Hebrish
Updates

Changelog

What changed, and when

Model versions, app releases and site updates.

Model versions, app releases and site changes, newest first. Model numbers are from the model pages and benchmarks, as measured when each entry was written; PoC-4 and PoC-5 were never benchmarked. The first five versions, published as v1 to v5, are now called PoC-1 to PoC-5. Models are now retrained together as numbered releases: see Model releases for which models and versions each one includes.

Show:

Each entry is filed under the platforms it concerns. A model update that reaches the apps counts for Android and Desktop too. The apps' What's new lists are built from this page by platform: Android shows Android entries, the desktop app Desktop and Site entries.

Get these by email

New model versions, app releases, and calls for voice recordings. Confirm by email; unsubscribe from any message.

Ubuntu 1.034

Ubuntu 1.034: Models in a window of their own

  • Models has its own window, opened from the new Models button at the top of the main window, or from the tray. It was a tab inside Settings. Nothing about the models themselves changes.
  • Windows and the AppImage stay on 1.026. Download.

Android 1.032

Android 1.032: The app and its models update themselves

  • App updates in the background (F-Droid and APK installs): once Hebrish is allowed to install apps, it checks for a new version every 12 hours on Wi-Fi, verifies the download, and installs it when no take is open. On Android 12 and later this is usually silent; Android 9 to 11 ask with a notification. A switch on Install and updates turns it off, and "Check for updates now" checks at once.
  • Models update in the background: every 6 hours, on Wi-Fi by default, Hebrish downloads new versions of the models you have installed. A take in progress is never interrupted: the switch to the new model waits until it ends. Switches on the Models screen.
  • Not yet tested on a phone. Download.

Android 1.031

Android 1.031: No full stop at the end, a live draft, and an edit row

  • No full stop at the end: a switch on the Dictation screen, and a small "No ." chip in the voice keyboard, leaves off the final full stop of each take, for searches, lookups and titles. Question and exclamation marks stay, and so do full stops inside the take.
  • Live draft (voice keyboard, experimental): while you speak, the words since your last pause appear underlined in the field and are heard again every half second, so an earlier word can change once the next ones make it clear. At a pause they are typed for good. Off by default.
  • Edit row in the voice keyboard: move the cursor, delete a word or the last sentence.
  • A take survives the keyboard hiding: it pauses and carries on when the keyboard comes back. An idle voice keyboard returns you to your regular keyboard.
  • Not yet tested on a phone. Download.

Ubuntu 1.033

Ubuntu 1.033: No full stop at the end, and a live draft

  • No full stop at the end: a new switch, in Settings and in the tray under Output, leaves off the final full stop of each take. Meant for searches, inventory lookups and titles. Question and exclamation marks stay, and so do full stops inside the take.
  • Live draft in the overlay (experimental): while you speak, the floating overlay shows your words since the last pause, dimmed, and hears them again every half second, so an earlier word can change once the next ones make it clear. What is pasted doesn't change: each piece still goes in at a pause. Off by default; on the Voice Typing tab under Insert as you go.
  • Windows and the AppImage stay on 1.026. Download.

Ubuntu 1.032

Ubuntu 1.032: Stop transcribes, and the Notepad gets formatting

  • Stop now transcribes. On the Transcription tab, the square Stop button stops the recording and puts the transcript in the Notepad. Until now Stop only kept the audio and waited for a separate Transcribe button, so a recording could look like it had vanished.
  • Voice Typing goes into the Notepad while the Transcription tab is open: dictate with your Voice Typing key and the words land at the Notepad's cursor, not in another app. Until now they went only to the clipboard and History.
  • The Notepad is a page you can format: a white page with bold, italic, highlight, headings and bulleted or numbered lists. Save and Copy write it as Markdown.
  • Windows and the AppImage stay on 1.026. Download.

Ubuntu 1.031

Ubuntu 1.031: Settings in their own window, Transcription as a Notepad

  • The main window holds the modes: Voice Typing, Transcription and History. Settings, Models, Dictionary, Updates, Diagnostics and About open in a separate Settings window, from the buttons at the top or Settings… in the tray menu.
  • Transcription in two panes: an Audio file pane (open a file, drop one, or try a sample) next to the Notepad, where you record and edit the transcript.
  • Test mic on the Transcription tab shows your microphone's level before you record.
  • Models lists the models made for a computer: Sabich, Shakshuka, Falafel and stock Parakeet. The rest are on the models page; one you already have installed stays listed.
  • Includes 1.030 (Enter waits for the last words), which was not released on its own. Windows and the AppImage stay on 1.026. Download.

Ubuntu 1.030

Ubuntu 1.030: Enter waits for the last words

  • Send pressed early: Hold the send key until they're in now also covers an Enter pressed while you're still recording (with The send key stops recording on). The recording stops, the last words are typed, and only then does the Enter go through, so a Claude Code prompt or a chat message is sent whole. If the words take more than 4 seconds, the Enter goes anyway and the rest go on the clipboard as before.
  • Default: Hold is the default on Ubuntu for new installs; existing settings are kept (Settings, Action keys). Windows and the AppImage stay on 1.026.

Android 1.030

Android 1.030: turn the floating mic off from the app

  • Floating mic: the Floating mic screen now has a Turn off the floating mic button. Before this, the only way to make it go away was to switch it off in Android's Accessibility settings. Turning it back on is still done there.

Android 1.029

Android 1.029: no more crash when you drag the floating mic

  • Floating mic: dragging the mic or its bar to a new place could close the app as you let go. It no longer does. This was the cause of the crashes some of you saw, not memory. Nothing else changes in this release.

Site

Falafel: the third model in active development

  • Falafel, the voice-command model, has its own page and sits beside Sabich and Shakshuka on /models and the home page. Two sizes, Falafel (desktop) and Falafel 110M (phones and laptop CPUs). Its first release is training now; not benchmarked. More in the announcement.

Ubuntu 1.029

Ubuntu 1.029: late words stay on the clipboard

  • Late words: when Enter goes in before the last words, Ctrl+V now pastes just the words that were missed. With "whole take on the clipboard" on, the whole take used to replace them a moment later.
  • Overlay: a new setting, Only while the last words come in, shows a small spinner from Stop until the last words are in, so you can see when Enter would still be too early. It is the default for new installs; existing settings are kept. Windows and the AppImage stay on 1.026.

Android 1.028

Android 1.028: Notes, and Falafel coming soon

  • Notes: a notepad you dictate into with the in-app mic, kept on the phone only. A note you leave untitled gets a title from the title model; a title you type is never replaced. Not yet tested on a phone.
  • Falafel: the new model for short voice commands is listed under Models, in a full size and a small one. It shows Coming soon until its first version is out, then downloads like any other model.
  • Floating mic: its bar now opens across the screen unless you picked a direction.

Ubuntu 1.028

Ubuntu 1.028: Falafel coming soon

  • Falafel: the new model for short voice commands has its own tab under Models, in a full size and a small one. It shows Coming soon until its first version is out, then downloads like any other model, with no app update. Windows and the AppImage stay on 1.026.

Models

Sabich v4.0: trained from scratch on a standardised dataset

  • Sabich v4.0 is retrained from stock NVIDIA Parakeet TDT 0.6B v2, not continued from earlier weights like the rolling updates v2.5 to v3.4. It is the new baseline later rolling updates continue from. 20 epochs on one NVIDIA H100 on Nebius AI Cloud, 192 minutes.
  • Dataset standardisation first: one set of transcript conventions (times, units, numbers, one spelling per term, NIS) applied to every published take, and every topic turned into its own Hebrish dataset in one standard format.
  • Data: 29,099 clips, about 45.9 hours (1,106 minutes code-switching). The apps pick it up on their own; on Hugging Face as Hebrish/Sabich-Rolling. Not benchmarked. More in the release post.

Models

Meorav v2.2: the voice model on 5.5 hours of recordings

  • Meorav v2.2, the Hebrish text-to-speech model, replaces v2.1: 3,995 takes, about 5.5 hours (v2.1: 1,103 takes, about 88 minutes), cleaned with NVIDIA NemotronLabs RE-USE instead of DeepFilterNet 3, with the same training settings.
  • On Hugging Face as Hebrish/Meorav (access by request) and Hebrish/Meorav-GGUF (Q8_0 and Q4_K_M). Not benchmarked. Listen to samples or read the release post.

Android 1.027

Android 1.027: Send, one-tap mic switching, and no more clashing takes

  • Keyboard and floating mic: starting a take in one now drops a take open in the other, so the new one records. Not yet tested on a phone.
  • Send: during a take the mic is a green send arrow on the floating mic and the keyboard, because Stop read as stopping, not transcribing. Retake is orange and Pause blue.
  • Changing mic: one tap on the mic name moves to the next connected mic in your order.

Ubuntu 1.027

Ubuntu 1.027: same version as Android

  • No changes: same app as 1.026, with the version number kept in step with Android. Windows and the AppImage stay on 1.026.

Apps 1.026

Apps 1.026: a joint stable release, with a smoother look

  • One version everywhere: Android, Ubuntu and Windows are all 1.026. Windows and the AppImage move up from 1.022.
  • Opening spin: the Hebrish logo turns once round as the app opens. Not yet tested on a phone.
  • Downloads: the progress bar glides instead of jumping, and shows a percentage.
  • Hebrew icon: the floating mic's Hebrew letter is now ה, for the Hebrew name היבריש.
  • Same words on every platform: Type as you go on Android is now Insert as you go, and the desktop's About page has the same tagline and a Privacy link. Download.

Site

Stickers to download and print

  • Stickers, under About: the logo in English and Hebrew, Corn the Sloth, and "I type with Hebrish", as print-ready PDFs and transparent PNGs. Also in Hebrew at /he/stickers.
  • The Hebrew pages now write the name in Hebrew: היבריש.

Android 1.019

A voice keyboard you can use while you dictate

  • Bottom row: the keys no longer sit under the navigation bar, so Enter, space and delete work while the mic is open.
  • Smaller: tighter buttons and no logo, so the keyboard takes less of the screen than before.
  • Back to your keyboard: the keyboard key is now gold, so it is easy to find.
  • Delete a sentence: press and hold delete to remove the last sentence.

Android 1.018

USB microphones, and your mics in your order

  • Mic order: Controls now has a list of mic types: USB, Bluetooth, wired and phone. Drag them into the order you want, and each take uses the highest one connected.
  • Pick a mic for now: tap the mic name while dictating, in the floating mic or the keyboard, to see every connected mic as an icon. Your choice lasts for the session.
  • USB microphones: they now work from the floating mic and the keyboard, and USB comes first by default.

Site

The home page opens with the comparison, and a shorter footer

  • The first thing on the home page is now regular speech recognition typing a sentence and getting the Hebrew wrong, then Hebrish typing it, then the headline benchmark numbers.
  • The Models section also lists Jachnun, the early test model.
  • The footer is much shorter, and folds into accordions on phones.

Site

Home page second pass, and a dataset index

  • The home page leads with what Hebrish is, and regular speech recognition and Hebrish are side by side at the top again.
  • Apps are split into Android, Windows, Ubuntu and Server, and Models has an intro and an Other models section.
  • Benchmarks are now written as sentences and ratios, and a new section, The mission, explains why code-switching is hard and what the long tail of Hebrew words looks like.
  • New dataset index: every dataset, mirrored from Hugging Face, with how sentences are generated and why recordings are split by theme and blended.

Site

A shorter home page

  • The home page now says what Hebrish is in plain words, who it is for, and goes straight to the downloads, including a link to the server.
  • New short sections on the models, the benchmarks in wrong words per 100, ways to support the mission, and what code-switching is.
  • The same-recording comparison moved to Benchmarks.
  • Licensing now reads as two pathways that depend on how you use Hebrish, with one example of each and the two routes to more speakers and accents.
  • The home page ends with three questions and licensing. The FAQ gains the API, using Hebrish in your app, and what a fine-tune is.

Site

Just Sabich, and quicker Android downloads

  • The home page now says Sabich, with no version number. The model page shows the version in the apps, read live, and the latest benchmarked release.
  • The Android card has one button each for F-Droid and the APK, and Google Play greyed out as coming soon. Download buttons no longer show an app version.

Android 1.017

A noise filter, and one Sabich

  • Noise filter: room noise no longer turns into stray words as often. A piece of audio that is only background is now skipped. The Voice Typing screen has a 0 to 100 dial: turn it up if words appear when you aren't speaking, down if quiet speech gets missed.
  • Sabich: it now downloads from a single location, and the older names for it are gone from the app. If you already have the latest Sabich, nothing downloads again.
  • Not yet tested on a device. Download.

Desktop 1.025

Benchmarks, to see how Hebrish runs on your computer

  • Benchmarks: a new window, opened from Diagnostics, measures how fast Hebrish transcribes and how much memory and CPU it uses. Only the figures are saved.
  • Three ways to run it: the reference file (the first Hebrish podcast episode, which Hebrish recognises, so your result compares directly with other computers), any audio file, or a recording.
  • Results: the latest run shows as Markdown you can copy, with past runs listed below it. The transcript stays folded away until you open it.
  • Not tried on Windows by hand. Recording and the full 67-minute reference run haven't been tried by hand yet. Download.

Desktop 1.024

Calibrate the noise filter for your mic and room

  • Noise filter: a 0 to 100 dial on the Voice Typing tab, set a little stricter than before, to keep room noise from turning into stray words.
  • Calibrate: stay quiet for 5 seconds, speak for 10, and Hebrish suggests a Safe and an Optimized setting. Nothing changes until you apply one, past runs are listed by microphone, and the recordings aren't kept.
  • Stray words: a fix for phrases such as "I'm sorry." appearing at the start of a take when nobody said them.
  • Sabich: it now downloads from a single location, and the older names for it are gone from the app. If you already have the latest Sabich, nothing downloads again.
  • The splash screen shows the logo again while the app loads.
  • Not tried on Windows by hand. Download.

Site

Home page reordered, and Support the project moves under About

  • The home page now offers the apps before asking for help, and a new section links to the use cases, manual, screenshots, models, blog and roadmap.
  • Support the project sits under a new About menu, with a breadcrumb, and has its own callout in How you can help.

Site

Daniel's story, a new support page, and breadcrumbs on model pages

  • About now opens with who makes Hebrish and why, with a real photo and some dictation facts.
  • Support is in the header, and now sets out the mission, why Hebrish is free, licensing in short and the ways to help.
  • Hugging Face and GitHub are in the header.
  • Every model page has a breadcrumb back to /models.
  • Fritter, Daniel's English model without code-switching, is listed on /models. It isn't a Hebrish model.

Site

The training blend, and category charts

  • /models now explains the training blend: the five ingredients, what each one is for, and why LibriSpeech is in the mix, with Corn at the blender.
  • The training explorer is now a simple line of total hours per version. Pick a category, such as tech vocabulary, to see its hours and its share of the mix over time on separate charts.

Desktop 1.022

Windows catches up

  • Windows and AppImage: a new stable build with everything since Windows 1.012.
  • Windows: Hebrish now reads its data files as UTF-8, so history imports and model details with non-English text load correctly.
  • Not tried on Windows by hand. Download.

Android 1.014

Insert last only adds what was missed

  • Insert last: if part of a take went in and the rest missed the field, Insert last now adds only the missing words, not the whole take again.
  • Licensing: Hebrish now tells you once, when you first open it, how it is licensed, with the Creative Commons icons, and the wording matches /licensing.
  • Corn the Sloth, the mascot, comes along with that notice.
  • Not yet tested on a phone. Download.

Desktop 1.021

Paste last only pastes what was missed, and Corn on the start-up screen

  • Paste last: after an early Enter kept your last words out of a message, the Paste last key and tray item now paste only those words, not the whole take again.
  • Licensing: Hebrish now tells you once, when you first open it, how it is licensed, with the Creative Commons icons. About has the same text, and the wording matches /licensing.
  • Corn the Sloth, the mascot, is on the start-up screen.
  • The tray menu shows the running version.
  • Beta: Ubuntu 26.04 .deb. Download.

Site

Licensing in the top menu, a clearer models page, and Corn

  • Licensing is now in the top menu and opens with what each part of Hebrish is licensed under, with the Creative Commons icons.
  • Corn the Sloth, the mascot, is on the About page.
  • /models now spotlights Sabich and Shakshuka, the two models in continued training, with what's different about them. The others are grouped as mini models, and proofs of concept and server models, with a trailing badge.

Models

See what each Sabich and Shakshuka version was trained on

  • /models now leads with the two models in active development, Sabich and Shakshuka, with a training explorer: hours of audio at each version, then clips, speakers, the blend and topics for any version you pick. Jachnun stays available but is not being actively maintained at the moment.

Model

Shakshuka v2.1

  • Shakshuka, the streaming model, is updated for the first time since Release v2. It continues from v2.0 and now knows the same recordings as Sabich: 1,672 minutes of them, up from 642. The apps offer the update on their own. v2.1 is licensed CC-BY-NC-4.0, with NVIDIA's notices kept.

Desktop 1.020

Switch models from the tray, one Sabich, and one Voice detection switch

  • Active model: switch between Sabich and Shakshuka from the tray menu, or with the new Select Sabich, Select Shakshuka and Toggle model action keys.
  • One Sabich: the model that was Sabich (Rolling) is now just Sabich, and it is the default on Ubuntu and Windows too. Sabich (Stable) and Jachnun are no longer offered. If you were using one of them, Hebrish moves to a model you already have, or to Sabich, which downloads once.
  • Voice detection everywhere: one switch now covers Voice Typing and Transcription Mode, in the tray and on both tabs, so skipping silence works when typing at the cursor too. It has no effect on Shakshuka with Insert as you go. Download.

Android 1.013

One Sabich

  • The model that was Sabich (Rolling) is now just Sabich, the default. Sabich (Stable) and Jachnun are no longer offered. If you were using one of them, Hebrish moves to a model you already have, or asks you to download Sabich. Download.

Site

A user manual

  • The user manual covers installing and using Hebrish on Android, Ubuntu and Windows: voice typing, transcription, history, the dictionary, models, settings, the local API and troubleshooting. It is searchable and has its own version number, starting at v1.0, which describes Android 1.012, Ubuntu 1.018 and Windows 1.012. Screenshots are from Ubuntu 1.018, with example text; Android screenshots will follow.

Desktop 1.018

Let your coding agent read your history, fix recurring errors and update the app

  • Agent control (opt-in): a new switch in Settings → Local API lets coding agents on this computer read and search your transcript history, add find-and-replace rules to the personal dictionary, update the app and its models, switch model and start or stop dictation. It is off by default, and GET /v1 lists every route.
  • Agent skills: update-hebrish, read-hebrish-history and fix-hebrish-mistranscriptions ship with the app. Install them for Claude Code, Codex and Gemini CLI from Settings → Local API, or with hebrish --install-agent-skills. The last one tests each rule against your real history before adding it.
  • Try a sample: the Transcribe tab has test recordings, so you can try a model without finding a recording of your own.
  • The send key can stop recording: an opt-in setting, for when you leave recording on.
  • Cleaner joins: question marks, exclamation marks and spaces at the point where live pieces meet are handled better.
  • Better audio import: files opened without ffmpeg are resampled to 16 kHz properly.
  • Fixed: the Dictionary tab could overwrite words written to the dictionary file while the app was open.

Data

How the training sentences are made

  • A method note, How the Hebrish training sentences are made (PDF, 5 pages, with charts): where the words and sentences come from, how sentences are generated from a target term and its context, which models and batch APIs wrote them, the rejection policy, spelling and transliteration, and how the test splits are kept apart. Linked from Technical, Datasets and Docs.
  • A PDF library on Technical: the method note and the white papers (release 1 in English and Hebrew, code-switching text-to-speech) re-set in one house style, and the whole library as one PDF. Their wording is unchanged, except that compute costs are no longer listed.
  • The dataset page has the current sentence sources (6,812 clips). The earlier table counted sentences whose word Daniel chose but whose text an LLM wrote as "written by hand"; the new one separates them.

Licensing

How Hebrish will be licensed to organisations

  • A new Licensing page, also in Hebrew: free for individuals, research, education and small non-profits and volunteer organisations; an institutional licence for health funds, hospitals and public bodies; a commercial licence for businesses. Licensed organisations can add custom vocabulary or a custom fork, a model fine-tuned from the latest main release for their own needs. Licences open to early adopters once the models are trained on many speakers.
  • Terms of use version 1.4 (sections 3 and 4) to match.

Feedback

Tell Daniel how Hebrish works for you

  • A short four-step form at the top of Feedback: where you use Hebrish, the model and version, your accent, an accuracy rating, the kinds of words it struggles with and a list of words it gets wrong. Contact details are optional, and you can stay anonymous.

Desktop 1.017

Transform a transcript into an email, a to-do list or minutes, and your clipboard comes back after a paste

  • Transform… (experimental): on the Transcription tab, rewrite a transcript with one of 206 ready-made prompts: a business or casual email, a to-do list, meeting minutes, a post, data formats and more. Filter them by type and style, or search. Edit any prompt to make it your own; your version is kept and the original can be restored. Your transcript is never changed: copy the result, or save it as a version beside the transcript.
  • Transform needs the transform model, an optional download in Models → Utilities (Qwen3 4B Instruct, 2.5 GB). It runs on your computer's processor, offline, and a transform takes tens of seconds on a laptop. It has not yet been tried end to end on a user's computer.
  • Fixed (Linux): after Hebrish pasted a transcript, the transcript stayed on the clipboard, so your next Ctrl+V pasted it again instead of what you had copied. Your previous clipboard is now put back after the paste. If Output → Clipboard is on, the transcript is still copied on purpose.
  • This is a beta: the Ubuntu .deb. Download.

Android 1.012

Clean-up and titles on the phone, and transcripts saved to a folder

  • Clean up: Transcription Mode can tidy a finished transcript with a small model on the phone: S1-mini by Superwhisper (the default) or S1-ASR-Cleanup (experimental), an optional download in Models → Add-on models. Needs a phone from roughly the last six years (ARMv8.2).
  • Save to a folder: Save writes the transcript as a Markdown file to a folder you pick, in year / month / day folders.
  • Titles: with the title model downloaded (Models → Add-on models), saved transcripts get a short title.
  • The add-on models are easier to find in the Models screen. Includes the 1.010 and 1.011 memory fixes.
  • Not yet tested on a phone. arm64-v8a only. Android 9+. Download.

Contributions

Send your own voice samples through Hugging Face

  • A batch of your own recordings can now be sent as a pull request on Hebrish/Community-Voice-Contributions, one folder per Hugging Face username. It is the only Hebrish dataset that accepts pull requests; we review them and add the recordings to the datasets. How to do it: Get involved.

Sabich (Rolling) r1

Sabich (Rolling) r1: the new default model in the apps

  • Sabich (Rolling) is Sabich v2.4 trained further from its own weights, for 6 more epochs, on all the recordings collected so far, instead of retraining from stock NVIDIA Parakeet TDT 0.6B v2. Same tokenizer, output style and hardware needs as Sabich: sherpa-onnx int8, about 661 MB download, phone or computer, no GPU.
  • From today it is the default (recommended) model in the Hebrish apps, Android and desktop, while it is being tested. Sabich v2.4 stays available in the Models screen. If testing goes well, Sabich will be retired or replaced by it. Later updates of the default model will be rolling releases (r2, r3, ...) that continue from the previous weights.
  • Hugging Face, both gated by request like Sabich: Hebrish/Sabich-Rolling (NeMo) and Hebrish/Sabich-Rolling-sherpa-onnx-int8 (the app build). No evaluation numbers are published for it yet.
  • The original model is now shown as Sabich (Stable) next to Sabich (Rolling). Only the display name changes: the Hugging Face repository, URL and DOIs are the same, and nothing changes for people using it.

Android 1.011

Stability release: the speech model is freed while the phone sleeps

  • 1.010 freed the speech model (about 700 MB) after 5 minutes without use, but that timer stalls while the phone is asleep, so the model could stay in memory for hours. Now it is also freed as soon as the screen turns off (unless you are recording), and the 5 minutes keep counting during sleep.
  • Only this change on top of 1.010. Not yet tested on a phone. Download.

Android 1.010

Stability release: much less memory when you are not dictating

  • The speech model (about 700 MB) no longer loads when the phone starts. It loads when you focus a text field or start a take, is freed after 5 minutes without use, and is freed straight away when the phone runs short of memory. Up to 1.009 it stayed in memory all the time, which could make the phone close other apps in the background. The first take after a pause can take a few seconds longer to start.
  • Dragging the floating mic is guarded against a crash reported on an older version.
  • Only these changes on top of 1.009. Not yet tested on a phone. Download.

Desktop 1.016

Download models in sets, update them all at once, and the title model downloads again

  • Bundles: the Models tab now starts with three sets you can download in one click. Everything (Sabich, Shakshuka, and the clean-up, title and paragraph models), Voice Typing (Sabich and Shakshuka) and Transcription (Sabich with the clean-up, title and paragraph models). Each shows what is still missing and roughly how big it is; anything already downloaded is skipped.
  • Check for updates now checks every model and utility you have downloaded, not only the one in use. When there is more than one update, Update all installs them in one go.
  • Fixed: in 1.015 the title model could fail to download with "the catalogue offers no clean-up model" if the app had last checked for updates before the title model was added.
  • Ubuntu. The Windows build follows later. Download.

Desktop 1.015

Saved transcripts get a folder, a title and a proper header; transcripts arrive in one go

  • Save writes straight to a Hebrish folder (Documents, Hebrish, transcripts by default), one folder per year, month and day, as Markdown, and never overwrites. The folder is shown under the Transcription tab's text box, with Open and Change.
  • Titles (optional): download the title model (292 MB) under Models, Utilities, Post-processing, and each saved transcript gets a short title in its file name and heading. You set the longest title there (60 characters by default).
  • Each saved file starts with YAML front matter (title, date, model, audio length), the title as a heading, the date and time under it, and which model transcribed it.
  • Versions are kept side by side: if you edited the transcript or clean-up changed it, the raw transcription is saved beside it. Save cleaned copy runs the clean-up model you choose (S1-mini by Superwhisper or S1-ASR-Cleanup) and saves the result as its own file, leaving the original as it was. A small JSON file links the versions.
  • The Transcription tab shows the finished transcript at once instead of filling it in piece by piece. A Transcribing timer appears only if it takes more than 5 seconds.
  • Next recording: choose whether a new recording or file adds to the transcript in the box or starts a new one.
  • Settings, Your data shows where your history, transcripts, dictionary and settings are kept, and Back up saves them as one tar.gz or zip file.
  • Ubuntu. The Windows build follows later. Download.

Desktop 1.014

Press Enter after silence, a local transcription API, and no more lost pieces

  • Press Enter after silence (optional, off by default, Settings, Action keys): while Voice Typing with Insert as you go, once text has been typed and you have been quiet for 8 seconds (2 to 60), Hebrish presses your Send key once, so a message to Claude goes without reaching for Enter.
  • Local API (optional, off by default, Settings, Local API): other programs on this computer can send audio files to Hebrish and get the transcript back, from an OpenAI-compatible endpoint on 127.0.0.1, port 8723. There is no key, and nothing leaves the computer. For coding agents there is a transcribe-with-hebrish skill.
  • Transcription no longer loses pieces: the model sometimes returned nothing for a whole stretch of a recording or file (about 20 seconds on one voice note). Such a stretch is now transcribed again in two halves.
  • Choose the clean-up model: S1-mini by Superwhisper (the default) or S1-ASR-Cleanup (experimental), under Models, Utilities, Post-processing.
  • What's new has its own Updates page, from a button in the header, with every entry newest first.
  • Ubuntu. The Windows build follows later. Download.

Models

Sabich v2.3: 20% more recordings

  • Sabich v2.3 replaces v2.2 as the recommended model in the apps. It is a Sabich-only update to Release v2, trained the same way as v2.2 on 1,296 minutes of audio (522.8 minutes of English with Hebrew words in it, up from 471.4). The apps offer it through their normal model update and delete v2.2 once it is installed; no app update is needed.
  • On held-out recordings it was never trained on, against v2.2 on the same clips: code-switching word error rate 4.74% (v2.2: 5.12%), Hebrew words not seen in training written correctly 58.0% (56.3%), and English 3.48% (3.90%). On public test set v2, though, it scores slightly below v2.2: 4.96% code-switching word error rate (v2.2: 4.91%) and 67.0% of Hebrew words right (70.1%). The two test sets disagree, and both differences are small.

Desktop 1.013

Optional clean-up model, and a lighter, clearer app

  • Clean up (optional, off by default): download S1-ASR-Cleanup, a fine-tune of S1-mini by Superwhisper, under Models, Utilities, Post-processing (about 660 MB), and turn on Clean up there or in the Transcription tab. Each finished Transcription Mode transcript is tidied once and set out in paragraphs. If anything goes wrong, you keep the text as transcribed.
  • The app does much less work while you dictate, and wakes far less often while it sits in the tray.
  • Warnings at the top of Voice Typing can be dismissed.
  • Saving, exporting, copying and deleting now tell you when something goes wrong instead of failing silently, and greyed-out buttons say why they are greyed out.
  • Check for model update and Paste last in the tray menu answer with a notification when the window is closed.
  • Changing a hotkey no longer gets stuck when the keyboard can't be read, and gives up after 15 seconds if no key is pressed. Transcription Mode remembers a gain you set by dragging the slider, and Diagnostics keeps updating in long sessions.
  • Ubuntu and Windows. Not yet tested on Windows hardware. Download.

Android 1.009

No dead ends, and steadier under load

  • Buttons that open a phone settings screen no longer close the app on phones that don't have that screen; the app tells you instead.
  • Try it is never stuck: with no model it offers Open Models, if loading fails it offers Try again, and "Set it up for every app" opens Setup.
  • If the phone runs out of storage during a Transcription Mode recording, the recording stops, is kept, and the app says why, instead of closing.
  • Choosing or deleting a model no longer freezes the Models screen, and downloads redraw their progress less often.
  • Discard and Retake no longer let a last piece of text slip in after them. Copy and Share say when there is nothing to copy yet, and TalkBack skips decorative images.
  • Not yet tested on a phone. Download.

Android 1.008

Switch between the phone's mic and Bluetooth with one tap

  • The top of the Hebrish voice keyboard shows which mic it will use: Phone, or BT and your headset's name. Tap it once to switch between the phone's mic and a connected Bluetooth headset, before you start or in the middle of a take; what you have already said is kept. With no headset connected, it says so.
  • On the floating mic, the mic chip in the bar now switches on the first tap, even while the headset is still connecting, with a short tick when it does.
  • The keyboard and the floating mic remember your choice for the next take.
  • Not yet tested on a phone. Download.

Desktop 1.012

New recording controls in Transcription

  • The Transcription tab's controls are round icon buttons with one colour for each action: Record red, Pause amber, Stop slate, Retake blue, Discard grey, and Transcribe a green send arrow. Hover a button for its name.
  • The transcript box belongs to Transcription Mode alone: a Voice Typing take no longer pastes into it when it has focus. The take goes to the clipboard and History instead, and the status bar says so.
  • Ubuntu and Windows. Not yet tested on Windows hardware. Download.

Desktop 1.011

Transcription Mode for long sessions

  • Built for sessions of 90 minutes and more (a recording can run up to 4 hours). If the microphone drops out (unplugged, the sound system restarts, the laptop wakes from sleep) it reconnects and the recording carries on; if the disk fills up, the recording stops cleanly and what was recorded is kept.
  • A Voice detection switch in the Transcription tab and in the tray menu, on by default in this mode and separate from Voice Typing's.
  • A small Recording light and a quiet timer next to the controls, and the tray icon shows when a transcription recording is running.
  • Save audio: keep each recording once it is transcribed, as the model hears it (16 kHz mono WAV, about 170 MB for 90 minutes), in a folder you choose.
  • With voice detection on, a long recording is scanned for speech once instead of twice before it is transcribed, so it starts sooner.
  • Ubuntu and Windows. Not yet tested on Windows hardware. Download.

Desktop 1.010

Better voice detection, and a tidier Models tab

  • Voice detection now uses Silero VAD, a small voice-activity model included with the app and on by default. It tells speech from fans, traffic and café noise, so silent takes are skipped and long pauses aren't transcribed. Turn it off under Models, Utilities, and detection goes by loudness as before.
  • The Models tab has two parts. Models has a tab for each family (Sabich, Jachnun, Shakshuka, Parakeet stock) with short rows; size, hardware and the technical notes are under Details. Utilities holds the small helper models: Silero VAD under Audio components, the paragraph model under Post-processing.
  • Ubuntu and Windows. Not yet tested on Windows hardware. Download.

Desktop 1.009

Your history at a glance, and a history that stays lean

  • Diagnostics has a History card: how much you have transcribed all told (rounded: 12K words, 1.2M words), hours of audio, days active and how much space the history takes. Show last 7 days draws a bar for each day's words.
  • The history tidies itself about once a day while you're not dictating: it compacts the search index and gives space from deleted transcripts back to the disk. Tidy now in Diagnostics runs it on demand.
  • Counts in the History tab are rounded the same way, with the exact figures in the tooltip.
  • Ubuntu and Windows. Not yet tested on Windows hardware. Download.

Desktop 1.008

A softer beep, now the default

  • The start and stop beeps are lower, rounder and quieter: easier on the ears, and on the people around you. The old beep is still there as Loud (Settings, Sounds, Beep); choosing one plays it.
  • Quiet hours can keep a sound: between the hours you set, play soft beeps only instead of nothing. With voice prompts on, quiet hours switch to soft beeps too.
  • Ubuntu and Windows. Not yet tested on Windows hardware. Download.

Models

Sabich v2.2: more recordings, speed perturbation, and a better score on every measure

  • Sabich v2.2 replaces v2.1 as the recommended model in the apps. It is a Sabich-only update to Release v2; Shakshuka, Jachnun and Meorav are unchanged. The apps offer it through their normal model update; no app update is needed.
  • More recordings: 1,079 minutes in the mix (899.7 of the speaker's takes, 471.4 of them code-switching, plus 100.1 of LibriSpeech), up from 848.4; 10,272 clips.
  • Speed perturbation: each clip is also trained at 0.9x and 1.1x speed, aimed at fast natural dictation. Averaging the last five epochs was tried too, but the single best epoch scored better on the validation split, so it was kept.
  • On the public test set v2 (scored 2026-10-27, same clips as v2.1) it is the best Sabich on all six measures: code-switching WER 4.91% (v2.1 5.21%), term recall 70.1% (65.6%), unseen-term recall 62.0% (56.2%), seen-term recall 83.3% (81.0%), English control WER 2.12% (2.42%) and tech WER 2.76% (3.02%). No test clip was in its training data.
  • Builds: the sherpa-onnx int8 build the apps run and the F16 GGUF are v2.2; the Sabich ONNX build (and the browser demo) stays at v2.1 and MLX at v1.0. CC-BY-NC-4.0, access by request on Hugging Face.

Android 1.007

Clearer controls in the voice keyboard

  • The Hebrish voice keyboard has one row of round, colour-coded buttons: Discard (red), Retake (amber), the mic in the middle, Pause (gold while paused) and Insert last. They are always in the same place, so Retake and Discard never need a scroll, and the keyboard no longer grows when you start talking.
  • The mic is green when it is ready and red with a stop square while you talk.
  • Not yet tested on a phone. Download.

Android 1.006

Enter waits for your last words, fewer repeated words, and Voice Typing

  • In the Hebrish voice keyboard, if you tap Enter after you stop and before the last words are in, the Enter now waits for them (up to 4 seconds), then sends. Your message no longer goes without its last words, and those words no longer land in the next one. If they take longer than that, Enter goes through and they are kept: Insert last puts them in. This is the keyboard's own Enter key, so nothing is taken over to do it. The floating mic can't do this, because there the Enter belongs to the other app.
  • Repeated words and false starts are cut, as on the desktop since 1.007: "I really wanted, wanted to go home" comes out as "I really wanted to go home". Words you double on purpose stay doubled ("really, really", "yalla yalla"), never more than twice, and numbers are left alone.
  • The Dictation screen is now Voice Typing.
  • Not yet tested on a phone. Download.

Desktop 1.007

Voice Typing, late words kept out of your next message, fewer repeated words

  • If you press Enter after you stop and before the last words are in, those words no longer land in whatever you type next, such as the next prompt in a chat box. They go on the clipboard, and the overlay shows Late words with a Paste button. Settings, Action keys, Send pressed early.
  • Optional: hold the Enter until the last words are in (up to 4 seconds), then send it. Off by default, because it briefly takes over the keyboard to do it.
  • Repeated words and false starts are cut: "I really wanted, wanted to go home" comes out as "I really wanted to go home". Words you double on purpose stay doubled ("really, really", "yalla yalla"), never more than twice, and numbers are left alone. The same rules go into the next Android release.
  • The Dictate tab is now Voice Typing, and the desktop window has the look of this site: the green, the paper header and a set of line icons, in light and dark themes.
  • Holding the Enter is not yet tested on a live desktop or on Windows hardware. Download.

Android 1.005 · Desktop 1.006

No more half-words where a long dictation is split

  • Longer dictations are decoded in pieces cut at your pauses. The first word after a pause sometimes lost its start ("ividual" for "individual", "chnically" for "technically"); it now comes out whole, on Android and on the desktop.
  • Found by reviewing real phone dictations from the app's history export.
  • This finishes the work on sentence boundaries in live dictation that began with Desktop 0.7.6 and Android 0.8.2: far fewer capitals and full stops in the middle of sentences (on test clips, 73 mid-sentence capitals became 30 and 27 false full stops became 10). How it works and what it measured: Live dictation without a streaming model.
  • Not yet tested on a phone or on Windows hardware. Download.

Android 1.004

Switch your remote's buttons off without losing them, and save sets of them

  • Settings, Sound, microphone and buttons has a new Use bound buttons switch. Turn it off and Hebrish takes no keys, so the same Bluetooth remote works in other apps; your bindings are kept for when you turn it back on.
  • Clear all removes every binding at once.
  • Saved sets: give the current bindings a name and save them, for example one set per remote, then switch between sets with Use. They are saved on your phone.
  • Not yet tested on a phone. Download.

Android 1.003 · Desktop 1.005

Models in plain words, a What's new list, and the floating mic types instead of pasting

  • Each model now says what it's built on, for example "Base model: Parakeet 0.6B v2 by NVIDIA"; the technical detail is under Technical details.
  • A What's new list is built into both apps (Android: More menu, side menu or About; desktop: the About tab), so it works offline.
  • No more <unk> in your text: the apps remove it, and Sabich v2.1, out today, no longer learns it.
  • Android 13+: the floating mic types at the cursor like a keyboard, so Android no longer says the app pasted from your clipboard.
  • Not yet tested on a phone or on Windows hardware. Download.

Models

Sabich v2.1: more recordings, and no more <unk>

  • Sabich v2.1 replaces v2.0 as the recommended model in the apps. It is a Sabich-only update to Release v2; Shakshuka, Jachnun and Meorav are unchanged. The apps offer it through their normal model update; no app update is needed.
  • More recordings: 848.4 minutes in the mix (722.9 of the speaker's takes, 449.2 of them code-switching, plus 80.3 of LibriSpeech), up from 713.9.
  • No more <unk> in the text: v2.0 had learned to write it for characters it cannot spell, and those are now mapped or removed from the training text (how it was found).
  • Words the speaker's own dictation got wrong get extra weight (repeated three times in training); paragraph passages were left out; code-switching and tech are kept in separate sentences.
  • Not benchmarked yet: a benchmark on a new, domain-balanced test set will follow. Builds: the sherpa-onnx int8 build the apps run and the F16 GGUF are v2.1; the Sabich ONNX and MLX builds stay at v1.0. CC-BY-NC-4.0, access by request on Hugging Face.

Desktop 1.003

Your microphones in your order, and no more lost words

  • Settings has a Microphones list: every mic Hebrish has seen on this computer, with a name you give it and a type (built-in, headset, wireless headset, shotgun, lav, dynamic and more). Put them in order and choose Preferred order as the microphone: each take uses the first one that is connected, and the system default when none is. A mic plugged in between takes is picked up. On Ubuntu a mic is recognised by its vendor and product id, so a USB mic stays the same mic on any port.
  • The last words after you stop now arrive. Spoken softly, the last few seconds of a take were sometimes taken for silence and dropped; they are now judged against the rest of the take.
  • Quiet takes are no longer rejected as "No speech". On a mic at a low level, such as a headset at its default gain, whole takes could be; a take now counts as speech when it rises clearly above its own background noise.
  • After you stop, the tray icon turns amber with three dots until the last words are in.
  • Inserting as you go uses slightly longer pieces: Fast about 4 seconds, Normal 6, Relaxed 10 (were 3, 5 and 8). If you were on a preset, you move to its new value.
  • Not yet tested on Windows hardware. Download.

Apps 1.002

Paragraphs in Transcription Mode, your choice of how

  • Transcription Mode, in both apps, has a Paragraphs setting: None (one block of text), Time-based (a new paragraph about every 30 seconds, as before) or Paragraph model. The setting is remembered. On desktop it is the menu in the Transcript header; on Android, Change under the progress bar.
  • Paragraph model is a small English classifier that reads the finished transcript and starts paragraphs where the topic moves on. It only adds line breaks and never changes your words. It runs on your device, once, when transcription finishes. Picking it the first time downloads it (34 MB). If it is missing or fails, that transcript gets time-based paragraphs and the app says so.
  • Switching the setting after a transcript re-flows it, as long as you haven't edited the text. Dictation is unchanged.
  • Not yet tested on a phone or on Windows hardware. Download.

Apps 1.001

Transcription Mode, the next take any time, and no more "um"

  • Transcription Mode, in both apps: record a long take, then transcribe it. It has its own microphone choice (you can switch mid-recording), a gain slider, a level meter with a clip light that you can hide, and Record/Pause, Retake, Stop, Transcribe and Discard, with the elapsed time shown small at the side. On desktop it replaces the File tab, now called Transcription; opening or dropping an audio file works as before. On Android it is on the home screen.
  • The recording is written to a file as you go, so a crash loses at most a couple of seconds, and it is deleted once it has been transcribed. A recording that wasn't transcribed is offered again.
  • Android: a new optional Setup step, Record in the background, lets Transcription Mode keep recording with the screen off or another app open, with Pause and Stop in a notification. It asks to allow notifications. Without it, keep the screen open while you record.
  • Start the next take while the last one is still transcribing. The desktop hotkey, Talk button, tray and overlay, and the Android floating mic and voice keyboard, used to ignore that press, which made it easy to lose track of whether you were recording. Now the mic opens at once and the texts arrive in order.
  • Hesitations are removed: um, uh, eh, er and hmm no longer appear in what you dictate, with every model.
  • Desktop: History stores just the new transcript when you transcribe a second file in the same session (it used to store the whole text box again).
  • Not yet tested on a phone or on Windows hardware. Download.

Apps: Android 1.000, Desktop 1.000

Hebrish 1.0: the floating mic gets out of the way, desktop output you can combine, and Paste last

  • Both apps are now version 1.000. From here, versions count 1.001, 1.002, and so on, on Android and desktop alike.
  • Desktop: Output on the Dictate tab (and in the tray menu) is now three buttons you can combine: At the cursor, Clipboard and This window. By default only At the cursor is on, so dictation goes where you are typing and no longer also fills the app's text box. Your earlier choice carries over.
  • Desktop: takes in the text box are separated by a blank line, so separate thoughts read as separate blocks, and the text is a little larger. Every take is still in History.
  • Desktop: Paste last transcription, an action key and a tray item that paste your last take again where the cursor is now, for when it went into the wrong window. Experimental and off by default: when a take finds no text field, the overlay offers a Paste button.
  • Android: while you talk, the floating mic's bar now opens along the edge of the screen by default, as a narrow column next to the mic, instead of across the text you are writing. It opens towards the middle of the screen. "Across the screen" is still an option.
  • The bar has a grip at its end: drag it to move the bar mid-take without touching the mic, which is Stop.
  • New settings for the bar: its size (50 to 100%), and a clear bar with no background, where each button sits on its own backdrop.
  • Size and opacity sliders have − and + buttons for one step at a time. Opacity goes down to 1%, size down to 20 dp.
  • New icons for the idle mic: the Hebrish H or the Hebrew letter ayin, on Hebrish green.
  • New installs start with a smaller mic on the left edge. If you have moved the mic before, it stays where you put it.
  • Android 12 and later say an app "pasted from your clipboard" every time the floating mic puts text in. Apps can't turn that off, but you can: Settings → Privacy → Show clipboard access. The app says so once, and the Floating mic screen has a button to the privacy settings.
  • Not yet tested on a phone or on Windows hardware. Download.

Android 0.10.0

A cleaner look, and the floating mic keeps your line breaks

  • The app's screens now use the Hebrish look from this site: the warm paper background, the green accent, the Hebrish logo (it was showing as a plain green square), headings in the Hebrish typeface and grouped settings on the home screen. Dark mode follows the same palette. The keyboard and the floating mic look as before.
  • The floating mic now pastes the transcript at your cursor instead of rewriting the whole text field. The field no longer jumps around while text goes in, and a line break or blank line before the cursor stays where it is. The transcript is left on the clipboard after a take.
  • Not yet tested on a phone; the new screens were checked on an Android 14 emulator. Download.

Apps: Android 0.9.0, Desktop 0.8.0

A personal dictionary: your words, written your way

  • Add the words you want written a certain way: names, brands, Hebrew words in the spelling you use. After each take, Hebrish replaces what the model wrote with your spelling. A word can have a "heard as" phrase (Hebrish, heard as "hebrew ish"); without one, it fixes the spelling and capitals of a word Hebrish already hears, also when the model splits it (git hub → GitHub).
  • Android: shared with the phone's own personal dictionary, so words you add in your keyboard work in Hebrish and words you add in Hebrish show up in your keyboards. Android allows this once the Hebrish voice keyboard is turned on in Settings; you don't have to use it.
  • Desktop: a new Dictionary tab, on Ubuntu and Windows.
  • It corrects the text after it is written, so it doesn't slow transcription or change what the model hears. Both apps follow the same rules.
  • Not yet tested on a phone or on Windows hardware. Download.

Android 0.8.2

Android: fewer capitals and full stops in the middle of sentences

  • Long takes are transcribed in pieces at your pauses, and each piece used to be treated as a sentence of its own, so a pause mid-sentence could leave "word. Capital". Now each piece is transcribed together with the end of the one before it, and a full stop at a pause is kept only if the sentence really ended there. Same fix as Desktop 0.7.6.
  • Takes shorter than about 8 seconds work as before. Long takes may take a moment longer after you tap stop.
  • Not yet tested on a device. Android 9+, arm64. Download.

Desktop 0.7.6

Desktop: fewer capitals and full stops in the middle of sentences

  • With insert as you go, stopping to think in the middle of a sentence no longer ends it. Each piece of text is now transcribed together with the end of the piece before it, so the model hears that the sentence carries on and writes it the way it would mid-sentence.
  • On a test with pauses inside sentences, capitals in the middle of a sentence fell from 73 to 30 and stray full stops from 27 to 10, with slightly fewer word errors. Transcribing takes about a fifth longer.
  • Ubuntu and Windows. Not yet tested on Windows hardware. Download.

Android 0.8.1

Android: transcription history, exported as CSV

  • Every take is now kept on your phone: the text, its length, the model, and how long it took to transcribe. Under Dictation, Export CSV saves it through the file picker, so you can put it straight into Google Drive. Nothing is uploaded by the app.
  • One row per take, with the time, text, word count, audio length, model, where you dictated (keyboard, floating mic or Try it) and the transcription timings (decode time, real-time factor, wait after stop) with the phone and app version. The first columns match the desktop history export.
  • Keep transcription history turns it off and Clear history deletes it. The history starts with this version.
  • Not yet tested on a device. Android 9+, arm64. Download.

Android 0.8.0

Android: minimise the floating mic with a double-tap

  • When the floating mic is in the way of what you are writing, double-tap it and it shrinks to a small dot in the same place. Tap the dot to bring the mic back. It never disappears, so it is there when you need it.
  • A single tap on the idle mic now starts the take after a short pause (about a third of a second), so it can tell a tap from a double-tap. Stop is still instant.
  • Not yet tested on a device. Android 9+, arm64. Download.

Android 0.7.6

Android: the floating mic keeps your line breaks

  • With the floating mic, a line break typed before dictating (to start a new paragraph, say) was deleted when the text arrived, and a space put in its place. In apps that don't report their trailing line breaks, Hebrish now pastes at the cursor, so nothing before it is touched.
  • Not yet tested on a device. Android 9+, arm64. Download.

Terms

Terms of use 1.2: every model listed

  • The licence table now lists every model line: Jachnun, Sabich and its variants on other bases, Shakshuka, Meorav and the Whisper fine-tunes, each with the licence it was released under. The base models and their licences are named in full, and the terms say what a commercial licence covers. PDF updated.

Desktop 0.7.5

Desktop: change how dictation is inserted from the tray

  • The tray menu has an Output submenu: paste at the cursor, clipboard only, or the app's own text box, and the paste shortcut (Ctrl+V, or Ctrl+Shift+V for terminals). It stays in step with Settings.
  • Ubuntu 26.04 and Windows. Not yet tested on Windows hardware. Download.

Desktop 0.7.4

Desktop: no more flicker when Hebrish pastes on KDE

  • On KDE Plasma 6 (Wayland), the window you were dictating into briefly lost focus every time Hebrish pasted, so its title bar flickered. Hebrish now sets the clipboard without taking focus. GNOME already worked this way.
  • Ubuntu 26.04 and Windows. Not yet tested on Windows hardware. Download.

Apps: desktop 0.7.3, Android 0.7.5

ALL CAPS, a tidier app, and history you can export

  • ALL CAPS: a setting (Dictation, on both apps) that types everything you dictate in capitals. Just for fun.
  • Desktop History keeps text only: the option to keep each take's audio is gone, and audio saved by earlier versions is deleted. Export your history to CSV (timestamp, transcription, model version) from History → Export….
  • Each desktop transcript records exactly which model produced it, so mistakes can be traced to a model version.
  • A polish pass on both apps: on Android, Try it no longer closes when the microphone is busy, Grant opens App info when Android won't ask again, and Setup recommends Sabich; on the desktop, Settings fits the window and the model descriptions match today's default (Sabich).
  • Not yet tested on a phone or on Windows hardware. Download.

Desktop 0.7.1

Desktop: dictation text arrives at a better rhythm, and Sabich is the default again

  • Sabich is the default for dictation again: it did better than Shakshuka for everyday dictation. If 0.7.0 moved you to Shakshuka, you are moved back once; Shakshuka stays downloaded and in Models.
  • Insert as you go follows your pauses: stop to think for a second and what you said so far goes in straight away. Normal is now every ~5 seconds (Fast ~3, Relaxed ~8).
  • No more stray full stops in the middle of a sentence when it is split at a short pause: the sentence carries on.
  • Shakshuka pastes a phrase at a time, a few words every 2-3 seconds and each sentence end at once, instead of a word or two at a time. Settings also offers a sentence or every word at a time.
  • Ubuntu 26.04 and Windows. Not yet tested on Windows hardware. Download.

Models

Release v2: Sabich v2.0, Shakshuka v2.0 and Meorav v2.0

  • The first model release: three models retrained together on more recordings (713.9 minutes in the speech-recognition mix, up from 623). Sabich v2.0 and Shakshuka v2.0 reach the apps through their normal model update; no app update is needed.
  • Meorav v2.0, the text-to-speech voice, trained on 1,103 read takes (88 minutes), is published on Hugging Face as Hebrish/Meorav (access by request). Not yet evaluated.
  • Later the same day, Meorav v2.1 replaced v2.0 in the same repository: the same training, on recordings cleaned with DeepFilterNet 3 first. Daniel picked it by listening; on 7 test sentences the low background noise is about 5 dB lower. Not otherwise evaluated.
  • Builds: each has a sherpa-onnx int8 build (what the apps run, also on CPUs and phones) and one GGUF, F16 (1.24 GB), for desktop GPUs. The v1.0 Q8_0 and Q4_K_M GGUF files of Sabich and Shakshuka are no longer offered. Sabich v2.0 is CC-BY-NC-4.0 (v1.0 stays CC-BY-4.0); Shakshuka keeps the NVIDIA Open Model License.
  • Not benchmarked: this release went out without a test-set evaluation. The other models (Jachnun, Sabich-Qwen3 and the small Sabich variants) stay at v1.0.

Models Site

Meorav, a text-to-speech line, and one page for every model

  • Meorav is the first Hebrish model that goes from text to speech: a Qwen3-TTS fine-tune that reads code-switched English aloud, in Irish-accented English with the Hebrew words pronounced correctly. Meorav is Hebrew for "mixed".
  • Meorav PoC-1 POC, the first training run (657 read takes, 50 minutes), was published open on Hugging Face as Hebrish/TTS-V1 under CC BY-NC 4.0 (since removed; current versions: Hebrish/Meorav). It is experimental: no benchmark yet, and the listening verdict is pending. The training code is public at Hebrish-ASR/Meorav-TTS.
  • Models now covers every model line, in three groups: speech recognition, small models and speech synthesis.

Desktop 0.7.0

Desktop: streaming dictation with Shakshuka, and a history of what you dictate

  • Shakshuka is now the default for dictation on Ubuntu and Windows. It streams: words are typed while you are still talking, about a second behind. If you already use another model, the app downloads Shakshuka in the background and switches once it is ready; switch back any time in Models. Sabich stays the pick for transcribing recordings.
  • History: every transcript is kept on your computer and searchable, with word counts for today, the last 7 days and all time.
  • Insert-as-you-go presets (Fast, Normal, Relaxed, Off, Custom), a Stop-and-send key, spoken prompts in Herman's, Corn's or Hilbert's voice, and beeps no longer end up in the transcript.
  • Model versions are shown simply, for example Sabich v2.0.
  • Not yet tested on Windows hardware. Download, or update through apt.

Android 0.7.4

Android: move the floating mic without closing it

  • The × next to the floating mic is gone. It sat where your finger lands to grab the mic, so trying to drag the mic closed it. Now you can drag the mic anywhere first, then tap it: the bar opens from where you left it. To hide the floating mic, turn it off in settings.
  • Model versions are shown simply, for example Sabich v2.0, without the commit number or training-run name.
  • Not yet tested on a phone. Download, or update from F-Droid.

Android 0.7.3

Android: the floating mic stays where you can reach it

  • With Show only while a text field is focused on, the floating mic used to disappear as soon as the field lost focus, for example when you closed the keyboard to read, or while you dragged it. Now it stays up once it has appeared, so you can drag it off what you are reading and then tap it.
  • A small × next to the idle mic closes it. It comes back the next time you tap into a text field.
  • Not yet tested on a phone. Download, or update from F-Droid.

Android 0.7.2

Android: a tidier floating mic bar

  • The buttons are now on the left of the bar (Pause, Retake, Discard) and the timer on the right, whichever side of the screen the mic is on.
  • Retake and Discard have their own coloured backgrounds, so they are easy to see and tap.
  • The microphone in use is a short chip: BT and the headset's name (shortened), or Internal for the phone's mic. A long headset name no longer stretches the bar.
  • Tap the chip to switch between Bluetooth and the phone's mic, also in the middle of a take. The choice is kept for the next takes.
  • Not yet tested on a phone. Download, or update from F-Droid.

Android 0.7.1

Android: steadier recording and transcription

  • If the microphone goes away mid-take (a Bluetooth headset switched off), Hebrish now stops and transcribes what it heard. Before, the take kept running with no audio and kept the phone busy until you pressed Stop.
  • If the microphone can't be opened because another app holds it, Resume now says so every time instead of appearing to record nothing.
  • Fixed a possible crash when the model changed while a take was being transcribed.
  • Not yet tested on a phone. Download, or update from F-Droid.

Desktop 0.6.3

Desktop app: the version at the bottom of the window, and screenshots in your software centre

  • The app's version now sits at the bottom right of the window (Hebrish 0.6.3); click it for this page.
  • On Ubuntu, Discover and GNOME Software show Hebrish with a description, screenshots and links once it is installed. The same screenshots are on the download page.
  • Every release is now tagged and checked automatically, so the version in the app always matches the release. If you have 0.6.2, Update now in the app installs it.

Desktop 0.6.2

Ubuntu: updates through apt, the app tells you when there's a new version, and agent skills to install it

  • Hebrish has a signed APT repository. Add it with one command and new versions arrive with your normal system updates (apt upgrade, Software Updater, Discover). The .deb adds the repository itself, so installing it once is enough.
  • The desktop app now checks for a new version at start and about once a day. On Ubuntu, Update now installs it after asking for your password and offers to restart; on Windows it links to the new zip. Turn it off or check by hand under Settings → App updates.
  • Want your AI agent to do it? Copy the prompt into Claude Code, Codex, Gemini CLI or another agent. It follows a public install skill from Hebrish-ASR/Hebrish-Agent-Skills, which also has a troubleshooting skill and works as a Claude Code plugin.

Apps 0.7.0 / 0.6.1

Pick a model by family and version, with a speed and accuracy guide

  • The Models screen (Android 0.7.0) and Models tab (desktop 0.6.1) now list each family (Sabich, Jachnun, Shakshuka, stock Parakeet) with its versions under it. Every version shows speed and accuracy dots (estimates, the same as Which model should I use?), its size, the hardware it needs and who it is for.
  • Sabich-Parakeet-110M (137 MB) can now be downloaded in the apps, for older phones and computers. Versions the apps can't run yet (Sabich-Qwen3-ASR-1.7B, Sabich-Canary-Qwen-2.5B, Sabich-asr4all, and Shakshuka's streaming) are listed with where they run instead.
  • The desktop app has Help me choose: your device, graphics card and its memory (detected where possible), and speed or accuracy, answered with the same rules as the website.
  • Not yet tested on a phone or on Windows. Download.

Models

Sabich-Canary-Qwen-2.5B, Sabich-asr4all-S and Sabich-asr4all-M published

  • Sabich-Canary-Qwen-2.5B v1.0, for desktop GPUs with about 6 GB of VRAM: the best plain English of the Sabich variants (2.4% WER on the speaker's own English, 0.9% on LibriSpeech), behind Sabich-Qwen3-ASR-1.7B on code-switching (6.9% against 5.4%). Long audio has to be cut into pieces of 40 seconds or less.
  • Sabich-asr4all-S and -M v1.0, 28M and 56M parameters, the smallest Hebrish models, streaming on the CPU, with an ONNX bundle. Code-switching WER 10.1% (-S) and 9.1% (-M), against 19.5% for the stock -M model. Numbers come out as words.
  • All three are on Hugging Face by access request and are not in the Hebrish apps yet. See Sabich's variants and Which model should I use?

Models Site

Which model should I use? New Sabich variants and Shakshuka v1.0

  • Which model should I use? asks about your device, GPU and VRAM and whether speed or accuracy matters more, and picks one model with a runner-up. Below it, every model sits on a speed-accuracy gradient (the positions are estimates).
  • Sabich now comes on more base models: Sabich-Qwen3-ASR-1.7B for desktop GPUs (code-switching WER 10.0% to 5.4% against stock Qwen3-ASR, on a held-out split; relabelled a proof of concept on 2026-10-25); Sabich-Parakeet-110M for low-end phones and wearables; and Sabich-asr4all-S and -M, the smallest, coming soon. See Sabich's variants.
  • Shakshuka v1.0, the streaming model, is published, trained on Sabich's data, with sherpa-onnx and GGUF builds. Not benchmarked yet.

Site

White paper reissued as model release 1, in English and Hebrew; the Hebrew page is easier to reach

  • The white paper is now Hebrish model release 1: Jachnun v1.0 and Sabich v1.0 (13 pages), also in Hebrew. It covers both first stable models together instead of treating Sabich as an addendum, adds a chart of the two on the same 422 held-out recordings, and has a section on Shakshuka, the streaming line, which is still in progress. No result changed; the old links lead to the new edition.
  • On phones, the switch to the Hebrew page (עברית) now sits next to the menu button instead of inside it, and hebrish.app/hebrew goes straight there.

Desktop 0.5.2

Desktop app: a clearer Models tab

  • The model in use is outlined in green, and every other model has a Switch to … button.
  • Sabich is listed first and marked as recommended. Jachnun and stock Parakeet are still one click away.
  • Ubuntu and Windows. The download links already point at it.

Models Site

Sabich is now the recommended model; meet the models on the home page

  • Sabich v1.0 is the recommended model in the Hebrish apps. The question was whether teaching it tech vocabulary would cost it accuracy on Hebrew words. On the same 422 held-out recordings (round R7, 2026-10-20) it didn't: Sabich v1.0 got 5.3% of words wrong and Jachnun v1.0 5.4%; it spelled 73 Hebrew words in 100 right, against 72; on tech dictation 1.54% against 1.74%. Plain English stayed level (LibriSpeech 1.3% against 1.2%). The gaps are small, so both models stay in the apps: pick the one that suits you, and switch in a tap. Details on the Sabich page.
  • The home page has a Meet the models section: a card for each line, Sabich, Jachnun and Shakshuka, with its mark and a link to its page, plus a short note on which one to pick. The same on /he.
  • Shakshuka, the streaming line, has its own mark: a pan of shakshuka.
  • Cartoons. Each model page opens with its mascot (a jachnun with headphones, a sabich at a laptop, a shakshuka with a microphone), and two gags show the problem Hebrish solves: a stock model hearing makolet as "Micollette" (home page, "How it works") and tofes as "TOFS" (use cases).

Desktop 0.5.1

Desktop app: dictate into the window, hear what it's doing, and see how fast it runs

  • Text appears as you talk: long takes are typed in every ~10 seconds, at a pause, so only the last few seconds are left when you stop. On by default.
  • The Dictate tab records into a text box you can edit, as well as typing at your cursor. It shows which microphone is in use, lets you switch, and has a level meter that tells you when you're too quiet, in the good zone, or clipping.
  • Voice detection: a take with no speech isn't transcribed, so no stray "Yeah." when you didn't say anything.
  • Sounds: softer beeps, or short spoken prompts ("Recording", "Paused", "Transcribing"), or none. Nothing plays when the text arrives, because you can see it.
  • Keys: Esc throws a take away; double-tap your dictation key to start again.
  • Models: a card for each model, Jachnun, Sabich and stock Parakeet, with its own Download button, so you can download one without switching to it. You choose the folder the models live in, and new model versions replace the old ones automatically, so each model is on your disk once.
  • A File tab for transcribing an audio file, a Diagnostics tab showing how fast the model runs on your computer (13 times faster than real time on a mid-range laptop processor), and history is now optional.
  • Ubuntu 26.04 and Windows. Not yet tested on Windows hardware. Download.

Site

Models in the main menu

  • Models is now in the top menu. Filter by vendor (NVIDIA, OpenAI, Meta) and by base model, together.
  • Each model page lists its parameter count, headline evaluation with links to the benchmark and result files, download links for every format and quant, its Hugging Face collection, and its DOIs.
  • New DOIs from Hugging Face: Jachnun, Sabich, the Jachnun GGUF, ONNX and sherpa-onnx builds, the Sabich sherpa-onnx build, and the org copy of the Hebrew-Latin Token Classifier. Earlier DOIs still resolve.
  • The home page results now show Sabich v1.0 on the newest held-out round (422 recordings): 5.3% of words wrong against stock Parakeet's 11.8%.

Models

Sabich in ONNX; Mac (MLX) builds of Jachnun v1.0 and Sabich v1.0

  • Sabich v1.0 is now on Hugging Face in ONNX, fp32 and int8, at Hebrish/Sabich-ONNX (access request), for onnx-asr, ONNX Runtime and Voxtype. Same files and layout as Jachnun's ONNX build.
  • Format check on five held-out clips against the original model: fp32 matched it on 5 of 5; int8 on 3 of 5 (one capital letter and one comma differed)
  • The app build gains bpe.vocab, which sherpa-onnx uses for hotword boosting. The model files themselves are unchanged, so the apps have nothing new to download.
  • For Macs with Apple silicon: Jachnun-MLX is now v1.0 (it was PoC-3), and Sabich-MLX is new. Both run with parakeet-mlx --fp32. Checked on ten clips against the original model: Sabich matched on 9 of 10 (one comma), Jachnun on 8 of 10 in fp32. In bf16, parakeet-mlx's default, Jachnun turned one clip into unknown tokens, hence --fp32.
  • Every format and DOI: Jachnun, Sabich.

Models

The streaming model is now called Shakshuka; Sabich has its own home

  • The Nemotron streaming fine-tune now has a name, like Jachnun and Sabich: Shakshuka. Its two checkpoints so far are proofs of concept, Shakshuka PoC-1 and PoC-2. The first retrain will be Shakshuka v1.0.
  • Sabich v1.0's NeMo model is now at Hebrish/Sabich on Hugging Face (access request), next to Hebrish/Jachnun.
  • The demo Space (retired 2026-10-25) transcribes with stock Parakeet, Jachnun v1.0 and Sabich v1.0 side by side.

Desktop 0.3.0

Ubuntu and Windows: two hotkeys, a floating overlay, and a more finished look

  • Two hotkeys: a toggle key (tap to start, tap to stop; holding it is push-to-talk too, unless you turn that off) and a separate push-to-talk key. Either can be any key or a foot pedal. Your old hotkey carries over.
  • An optional floating overlay, like the Android floating mic: the timer, a level meter, Pause, Retake and Discard over any app, then "Transcribing…" and "Inserted". Drag it anywhere; it never takes the keyboard focus, so the text still goes where your cursor is.
  • A start-up screen while the model loads, and a redesigned About tab.
  • Better on Wayland: the window runs natively on GNOME and KDE, and the clipboard works on GNOME's Wayland session too. Download.

Site

Who makes Hebrish, said plainly

  • The home page now has a "Created by" card: Hebrish is a one-person, not-for-profit project by Daniel Rosehill, with no company or organisation behind it. The About page says the same, and the site no longer talks as "we".

Site

A page for every model, with its DOIs

  • Models is now an index you can filter by base model: NVIDIA Parakeet and Nemotron, or OpenAI Whisper. Each model has its own page listing every format and quant, every version, and the DOI of each Hugging Face repository, with a citation to copy: Jachnun, Sabich, Nemotron, Whisper Small ACFT, Whisper Hebrish.
  • The Hebrew-Latin Token Classifier now has its own page under Small models, and a home in the Hebrish organisation: it finds Hebrew words written in Latin letters, the first step in getting text-to-speech to say them as Hebrew.
  • How the token classifier fits a text-to-speech stack: a six-stage pipeline (detect, transliterate, add niqqud, tag, speak), with code and real outputs in a new repository and a white paper (PDF).
  • Access requests: ask on Hugging Face first; if you have not heard back after a few days, the contact form now has a "Request access to a model" option.

Android 0.6.0

A setup checklist, a side menu, simpler models and a tidier voice keyboard

  • Home opens with a setup checklist: every step in order, each ticking itself off, with a button that goes straight to the right settings page. The same steps with pictures are on Set up Hebrish on Android.
  • A side menu with every screen, Explore the lexicon, Contribute and a new About page. The app and model versions are shown at the bottom.
  • Models as cards, each with its mark: Jachnun (English + Hebrew vocabulary) first and recommended, Sabich (Hebrew + tech vocabulary), stock Parakeet last. A new install uses Jachnun; an existing one keeps its model. The model location is always shown, and you can move the models back to the default.
  • The voice keyboard has the mic at the top, the Hebrish mark and a settings button above it, and comma and full stop keys.
  • A retake is now easy to see: the mic flashes amber and spins back once. Optional beeps at start, stop and retake; choose your microphone (Bluetooth first); control it from a Bluetooth button or Tasker. Not yet tested on a phone. Download.

Site

An Android setup checklist

  • New page: Set up Hebrish on Android. Eight steps, in the same order as the app's Setup checklist: install, microphone, model, voice keyboard, floating mic, battery, "Pause app activity if unused", and all files access.
  • It covers the parts Android hides: unlocking the floating mic on Android 13 and newer with Allow restricted settings, and keeping the microphone permission from being taken away after months unused.
  • Each step has its own link, such as hebrish.app/setup#step-mic, so the app can open the right one.

Model Sabich v1.0

Sabich v1.0: a second model in the apps, code-switching plus tech

  • A second Parakeet fine-tune, trained on everything: the code-switched speech plus tech vocabulary, brand names, numbers and currency, Irish-English and the other task recordings. 560.5 minutes of the speaker's takes (373.1 of them code-switching), 20 epochs, 52 minutes on one L40S. Named after the Israeli pita with a bit of everything in it.
  • The Hebrish apps now offer it in the model picker next to Jachnun v1.0 and stock Parakeet. No app update needed: it shows up on the next update check.
  • Public benchmark: 73.8% term recall, 68.4% on unseen terms, 3.3% WER, 2.0% English control WER, level with Jachnun v1.0 (73.3%, 70.1%, 3.3%, 2.2%) within noise. Benchmarks · details.
  • Jachnun v1.0 stays the code-switching model. Still one speaker's voice: tell us how either does on yours.
  • App build on Hugging Face (open) · main repository (access request; .nemo and GGUF).

Keyboard 4.2-hebrish.1

Hebrish Keyboard: a whole keyboard with Hebrish built in

  • A second Android app: Hebrish Keyboard, HeliBoard 4.2 with Hebrish dictation on the mic key. Tap to talk, tap again to insert; no switching to a separate voice keyboard.
  • It installs alongside HeliBoard and is in the Hebrish F-Droid repository. Pick it or the Hebrish app, not both: each downloads its own copy of the model.
  • New HeliBoard releases are merged and re-released automatically every week, so it keeps up with HeliBoard.
  • Very early: released before it was tried on a phone. Problems go to its issue tracker. Download.

Android 0.5.0

Move the floating mic while you talk, and choose how it looks

  • You can drag the floating mic by any part of it, including the bar with the timer and buttons while you talk. The buttons still work when you tap them.
  • Press and hold, wait for the buzz, then drag: holding still before you move no longer stops the drag.
  • Two new settings: how see-through the bar's background is while you talk, and the mic's colour (six choices; it still turns red while recording). Not yet tested on a phone. Download.

Android 0.4.0

Faster dictation, longer takes, and a floating mic that works in more apps

  • Hebrish now transcribes while you talk, so the wait after you tap Stop is short. A take can run 20 minutes; it used to stop at 90 seconds.
  • The floating mic now finds the text field in apps such as ChatGPT, where it used to say there was no field.
  • The floating mic's setup guide has a new step that lets Hebrish run without battery limits, so phones like OnePlus don't stop it in the background.
  • If no field was selected when your text was ready, tap one and then Insert. You can now place the mic anywhere, and set its size and how see-through it is.
  • The app has a screen for each area: models (download one or all, and switch between them), voice keyboard, floating mic, dictation, install and updates, and help. Not yet tested on a phone. Download.

Android 0.3.0

A floating mic for any app

  • Hebrish for Android can now show a mic button that floats over every app. Tap it in any text field, talk, and tap Stop: the text goes in at the cursor, and your keyboard stays as it is.
  • While you talk, the mic has Pause, Retake and Discard next to it, and a timer.
  • You choose where it sits: drag it to either edge, or set the side, height and size in the app. It can also show only while you are typing.
  • It is optional, and turned on in the app's setup screen as an accessibility service. This first version has not been tested on a phone yet. Download.

Model

The model is called Jachnun

  • Jachnun is now the name of the Hebrish model as a whole, the Parakeet fine-tune, rather than of one version. The current release is Jachnun v1.0, the same weights as before; later retrains will be Jachnun v1.1, v1.2 and so on.
  • Jachnun has its own repository in the Hebrish organisation on Hugging Face: Hebrish/Jachnun. The proof-of-concept checkpoints, PoC-1 to PoC-5, stay where they were published, at danielrosehill/Parakeet-TDT-0.6B-v2-English-Hebrew-Codeswitching (DOI 10.57967/hf/10995).
  • The proof-of-concept checkpoints no longer have food names. The Nemotron streaming fine-tune does not have a name yet.

Site

A simpler home page, F-Droid first, and jump links

  • The home page is shorter: the side by side, the results, the vocabulary (now an A–Z cloud of words), how it works, and downloads.
  • On Android we now recommend F-Droid, so updates arrive on their own. The direct APK is still there.
  • Download links now always give the newest version: hebrish.app/get/android, /get/ubuntu, /get/windows.
  • Long pages have an "On this page" bar to jump between sections.

Android

Hebrish on F-Droid

  • Hebrish now has its own F-Droid repository, https://fdroid.hebrish.app/repo. Add it once and new Android versions arrive as app updates: how to add it.
  • It serves the same APKs as the download button, signed with the same key, so an existing install updates in place.

Apps 0.2.2

Hebrish 0.2.2, and downloads from our own CDN

  • The apps now download the model from cdn.hebrish.app, Hebrish's own content delivery network, instead of Hugging Face. An existing v1.0 install is kept, not downloaded again.
  • The app downloads themselves are served from the CDN too: download.
  • The in-browser demo loads its model from the CDN.

Apps 0.2.1

Hebrish 0.2.1: model download fixed, terms of use

  • Fixes downloading the code-switching model. Its main repository now needs an access request, so the apps download v1.0 from the open app build instead. Installed 0.2.0 apps pick this up on their next update check.
  • Both apps now show the terms of use: free for personal and non-commercial use; commercial use, including telephony, needs a licence.
  • The desktop About tab shows v1.0's benchmark results. Download.

Site

Terms of use, and access requests

  • Terms of use: free for personal, research, education and non-profit use; commercial use, including telephony and call centres, needs a commercial licence. Versions up to v1.0 stay under CC-BY-4.0; from v1.1, new versions will be CC-BY-NC-4.0.
  • The model repositories on Hugging Face now need an access request, or write to access-requests@hebrish.app. A fillable attestation form covers non-commercial use. The apps and the in-browser demo keep working without one.

Site

Mobile menu, and what's inside the dataset

  • On phones and tablets the navigation is now behind a Menu button in the sticky header, with large tap targets.
  • The dataset page shows what it covers: 13 groups of words, from religious life to Israeli streets, with examples and live counts, plus Yiddish.
  • A section on Ashkenazi and Israeli pronunciation: recordings are now made both ways, and balancing the dataset across them is planned.

Benchmark

v1.0 benchmarked: the best scores so far

  • Public benchmark (84 clips, frozen): v1.0 writes 73.3% of Hebrew terms in the exact spelling, up from 67.4% for PoC-3, with a word error rate of 3.3% (PoC-3: 4.0%). On terms never seen in training it scores 70.1%. Benchmarks.
  • Held-out test split (387 clips, dataset revision 550bfcaa): code-switching WER 5.1%, against 11.3% for stock Parakeet and 6.0% for PoC-3 on the same clips. The speaker's own English improved (1.7% against 3.1% for stock); LibriSpeech stayed level (1.2%).
  • The Nemotron streaming fine-tune, scored offline: 62.2% term recall and 4.7% WER on the public benchmark, behind v1.0 on every test.
  • Nemotron while streaming, at 560 ms chunks: 5.65% code-switching WER, against 5.81% offline. Streaming costs it nothing measurable.
  • A white paper (PDF) on how v1.0 was trained and scored, and reusable figures.

Site

Projects board, the three models side by side, Hebrish on iPhone

  • A projects board: what is under consideration, in progress and done, with the funding each project needs.
  • Models and developer notes: the Parakeet, Nemotron and Whisper Small fine-tunes listed separately, what each suits (Nemotron streams, so text appears while you speak; Whisper Small runs in FUTO Keyboard), and suggested runtimes for each. The nav's "Model" link is now "Models".
  • A post on what an iPhone version would take: the tasks, why it ships through TestFlight, and its cost, about $199 plus testing.

Model v1.0

Jachnun v1.0: the first release we stand behind

  • The sixth checkpoint, and the first one past proof of concept. The five before it showed word error rate moving the right way; they are now called PoC-1 to PoC-5 (published as v1 to v5). Later retrains on more data will be v1.1, v1.2 and so on.
  • 45 more minutes of code-switched speech (294 → 339 min, 3,224 clips) and 13 more minutes of the speaker's own English (83 → 95 min). Same recipe as PoC-5: full fine-tune from stock Parakeet, 30% English, 20 epochs, 41 minutes on one L40S.
  • Benchmarked the same day (see the entry above): code-switching WER 5.1% against 11.3% for stock Parakeet, and 73.3% term recall on the public benchmark. Validation WER during training was 5.0% (raw, 218 clips).
  • All formats except MLX, which is still PoC-3: .nemo, GGUF Q8_0 / Q4_K_M / F16, ONNX fp32 and int8, sherpa-onnx int8. Model on Hugging Face.
  • Still one speaker's voice. Try it and tell us how it does on yours: download or try it in the browser, then write to us or info@hebrish.app.
  • Retraining pauses for a while. Next: v1.1 at about 450 minutes of code-switched speech.

Experimental

Nemotron streaming variant, second checkpoint

  • Retrained on the same data and recipe as Jachnun v1.0: 46 minutes on an L40S, best epoch 18 of 20.
  • Validation WER 6.3% (raw, offline decoding) against 5.0% for Jachnun v1.0 on the same validation set. Benchmarked offline the same day: 62.2% term recall and 4.7% WER on the public benchmark. Streaming at 560 ms chunks: 5.65% code-switching WER on the held-out split, against 5.81% offline. No app uses it yet.
  • Model on Hugging Face (NVIDIA Open Model License).

Site

Clip archive

  • A clip archive: every recording in the dataset, playable, with its transcript, target terms, split and date. Each clip has a unique ID (the dataset row id) to cite it by, and a link that opens the archive at that clip.
  • Clips are being tagged with how the Hebrew is pronounced (Israeli, Ashkenazi or mixed), shown in the archive with a filter.

Site

Try it in your browser, version names

  • Try it in your browser: record or upload a clip and the model transcribes it on your own device. Nothing is uploaded. It downloads about 660 MB once.
  • Each model version now has a name, an Israeli food in alphabetical order: Amba, Bamba, Chamin, Ful, Halva. All of them are proofs of concept (since renamed PoC-1 to PoC-5).
  • An "Experimental" note on every page, and platform icons on the download buttons.

Site

Send a word list

  • A word-list form for Hebrew words dictation gets wrong or the lexicon is missing. Paste a list in any shape; Tidy the list makes it comma-separated and alphabetical, and you review it before it is sent.

Site

Mission, blog, brand page, lexicon table

  • Our mission, and what not-for-profit means here: free apps, open model and data, no ads, audio stays on your device. The about page now has a photo and a link to Daniel's full bio.
  • A blog, starting with Introducing Hebrish and what word error rate means. RSS feed.
  • "What does this mean?" buttons next to the numbers, and results shown as mistakes per 100 words.
  • The lexicon as a sortable, filterable table, and as a CSV download (lexicon.csv) refreshed with the daily counts.
  • Brand page: voice tips, copyable templates, logos and colours. Use cases are in the top menu.

Apps

0.2.0: Hebrish all the way through

  • The last traces of the old name, Parlance, are gone: the Android app is now app.hebrish.dictation, and the Ubuntu package and command are hebrish. Downloads: Hebrish-v0.2.0.apk, hebrish_0.2.0-1_amd64.deb, Hebrish-v0.2.0-windows-x64.zip.
  • Android 0.2.0 installs as a new app. Install it and turn the keyboard on again; your 0.1.x install becomes Parlance, a separate general-purpose dictation app.
  • On the desktop the default hotkey is now F14, so Hebrish and Parlance can run side by side.
  • Release notes: GitHub releases.

Apps

0.1.2: the app is now called Hebrish

  • On Android and the desktop, the app now shows as Hebrish in the launcher, the keyboard list, the window and the tray, so searching your apps for "Hebrish" finds it.
  • Settings and downloaded models carry over; it installs over earlier versions. The Android and Windows downloads are now named Hebrish too (Hebrish-v0.1.2.apk, Hebrish-v0.1.2-windows-x64.zip); the Ubuntu package and the repository keep the old name, Parlance.

Apps

0.1.1: Hebrish icon, feedback by email

  • Android and desktop now use the Hebrish mark as the app icon.
  • Report a bug, Request a feature and Send feedback open an email to feedback@hebrish.app, with the app and system version already filled in.
  • Release notes: GitHub releases.

Site

Share buttons, newsletter

  • Email updates: sign up on the home page, here, or on the support page.
  • Every page can be shared to X with the text filled in, or sent on WhatsApp. Copy link, and your phone's share sheet where the browser has one.

Site

Charts, lexicon explorer, roadmap

  • Interactive results charts on the home page and model card, with targets for the next checkpoints (goals, not results).
  • Lexicon explorer: 2,989 words and names in 13 domains, with how often each is recorded.
  • Roadmap with live dataset counts, refreshed daily from Hugging Face.
  • New pages: use cases, integration guide, support (donations coming soon).

Site

hebrish.app launched

  • The site, with download, model card, dataset, FAQ, about and contact pages.

Apps

Parlance 0.1.0 for Android and desktop

  • First release of Parlance, which replaces Voice To Hebrish and Parakeet-Type-Ubuntu.
  • Choose between stock Parakeet and the English-Hebrew code-switching model; each is a separate ~660 MB download, kept until you delete it.
  • Android: voice keyboard (works with HeliBoard) and a Try it notepad. arm64, Android 9+.
  • Ubuntu 26.04 .deb and a portable Windows zip (untested on real Windows hardware): hotkey dictation, a Transcribe tab, model updates with SHA-256 checks.
  • Release notes: GitHub releases.

Experimental

Streaming variant on Nemotron

  • The same data and recipe as PoC-5 (v5), on NVIDIA Nemotron Speech Streaming EN 0.6B, which streams in 80 to 1,120 ms chunks.
  • Validation WER 6.4% (offline decoding) against 4.9% for the Parakeet PoC-5 run on the same validation set. Benchmarked offline the same day: 62.2% term recall and 4.7% WER on the public benchmark. Streaming at 560 ms chunks: 5.65% code-switching WER on the held-out split, against 5.81% offline. No app uses it yet.
  • Model on Hugging Face (NVIDIA Open Model License).

Model PoC-5 POC

PoC-5, published as v5: 2,829 sentences

  • 25 more minutes of code-switched speech (269 → 294 min, 2,829 clips) and 10 more minutes of the speaker's own English (73 → 83 min).
  • A smaller step than the usual 1.5×, released by choice. Not benchmarked yet.
  • model files at this version

Model PoC-4 POC

PoC-4, published as v4: 2,622 sentences

  • 73 more minutes of code-switched speech (196 → 269 min, 2,622 clips).
  • The training mix reached its 70 / 20 / 10 target (code-switching / own English / LibriSpeech) for the first time. Not benchmarked yet.
  • model files at this version

Model PoC-3 POC

PoC-3, published as v3: 1,921 sentences

  • 76 more minutes of code-switched speech (120 → 196 min, 1,921 clips).
  • Code-switching WER down from 5.8% to 5.3%, long-form WER from 5.6% to 5.0%. Hebrew terms seen in training written correctly: 66%.
  • model files at this version

Model PoC-2 POC

PoC-2, published as v2: 1,248 sentences

  • 120.1 minutes of code-switched speech (1,248 clips).
  • Code-switching WER 5.8%, against 10.9% for stock Parakeet.
  • model files at this version

Model PoC-1 POC

PoC-1, published as v1: 873 sentences

  • The first fine-tune: 77.4 minutes of code-switched speech (873 clips), plus the speaker's own English and LibriSpeech replay.
  • Code-switching WER 6.0%, against 10.9% for stock Parakeet.
  • model files at this version

Each sentence is one read-aloud recording. Next: v1.1 at about 450 minutes of code-switched audio; retraining is paused until then. Progress: project tracker.