Dataset index
Every individual Hebrish dataset in one list, mirrored from the index on Hugging Face. There are two main kinds: sentences to read, and the audio of those sentences being read.
Sentence datasets and audio datasets
A sentence dataset is text to read. Most sentences are generated by an LLM from a target word and its context, or on demand when Daniel picks a word in the recording app. Each generated sentence passes an automatic gate (6 to 18 words, each target term exactly once, allowed characters only), for most of the data a reviewer model from another family, and the reader, who skips any sentence that misuses a word or sounds unnatural. The code-switching dataset page has the shares, and How the Hebrish training sentences are made (PDF) has the full method.
An audio dataset holds recorded takes of those sentences, one dataset per theme: code-switching, Yiddish, tech, acronyms, currency and so on. Every take is recorded once, into the master dataset Read-Aloud-Voice-Sentences, and the theme datasets are views of it. The speech models are trained on one speaker's voice; contributing recordings is how that grows.
Why recordings are split by theme
- Thematic segmentation. Each phenomenon stays in its own dataset, so it can be browsed, reused, measured and extended on its own.
- Dataset blending. A model trains on the master dataset through a blend: a rule for which rows it takes. Choosing how much of each theme goes in helps avoid overfitting to one kind of sentence, and problems with recognising word formats, such as how numbers, acronyms and web addresses are written.
The training blend explains the ingredients, and Datasets shows what each theme trains and which models use it.
105 datasets: 67 published, 1 private and 37 planned. Mirrored from Hebrish/Dataset-Index, built 2026-11-06.
Sentence datasets (10)
Text to read, no audio.
| Dataset | Access | What it is |
|---|---|---|
| AI-Interactions-Sentence-Bank Hugging Face | Direct download | Sentences dictated to AI agents (text only). Audio: AI-Interactions-Sentences. |
| AI-Tech-Vocabulary-Sentence-Bank Hugging Face | Accept terms | AI and tech vocabulary sentence bank (text only). Audio: Tech-Sentences. |
| hebrew-latin-code-switching Hugging Face | Accept terms | Sentences with romanized Hebrew words labelled, for the token classifier. |
| Homophone-Context-Sentence-Bank Hugging Face | Accept terms | Homophone-in-context sentence bank (text only). Audio: Homophone-Context-Sentences. |
| Linux-Admin-Sentence-Bank Hugging Face | Accept terms | Linux administration sentence bank (text only). Audio: Linux-Admin-Sentences. |
| Master-Sentence-Bank | private | Every sentence written or read for the speech models, one row per unique sentence, categorised, with take counts. Text only. |
| October-7-Memorial-Sentence-Bank Hugging Face | Accept terms | In memory of those who were murdered and who fell on October 7, 2023: three sentences about each person's life, and places and terms connected with that day. Audio: October-7-Memorial-Sentences. |
| Personal-Vocabulary-Sentence-Bank Hugging Face | Accept terms | Personal vocabulary sentence bank (text only). Audio: Personal-Vocabulary-Sentences. |
| Smart-Home-Sentence-Bank Hugging Face | Direct download | Smart-home sentence bank (text only). Audio: Smart-Home-Sentences. |
| term-context-sentences Hugging Face | Direct download | Training data for the term-context sentence generator (Granite 4.2 3B). |
Audio datasets (35)
Read takes, one dataset per theme.
| Dataset | Access | What it is |
|---|---|---|
| Acronym-Sentences Hugging Face | Direct download | Natural sentences with acronyms read by one speaker, both spelled out (API, SSH, HTTPS) and said as words (NASA, JSON, YAML), each in its standard written form |
| AI-Interactions-Sentences Hugging Face | Direct download | Instructions dictated to AI coding agents (Claude Code, Codex and others) and the git and GitHub workflow around them, read aloud by one speaker: subagents, worktrees, compaction, plan mode, hooks, MCP servers, pull requests, rebases Audio: AI-Interactions-Sentence-Bank. |
| Brand-Name-Sentences Hugging Face | Accept terms | Natural sentences with company, product, AI model and open-source project names (Anthropic, OpenRouter, Amazon Bedrock, Qwen, Kubernetes, Nginx) read by one speaker, written in each name's official form |
| Community-Voice-Contributions Hugging Face | Direct download | Recordings contributed by anyone, for the speech models. |
| Currency-Sentences Hugging Face | Accept terms | One speaker saying amounts of money the natural way ("50 shekels", "a dollar fifty", "6 pounds 50") |
| Disfluent-Read-Sentences Hugging Face | Accept terms | Experimental |
| English-Hebrew-Codeswitching-Sentences Hugging Face | Accept terms | English sentences with everyday Hebrew words and Israeli names written in Latin letters, read aloud by one English speaker living in Israel |
| English-Hebrew-Mixed-Sentences Hugging Face | Needs approval | The first code-switching set: sentences with recordings (637 WAV files, checked 2026-11-05). Its sentences are also a source of the prompt bank. |
| Expanded-Acronym-Sentences Hugging Face | Accept terms | Chat and business shorthand said as letters by one speaker ("by E O D", "F Y I", "I M O") with the transcript writing it out in full ("by end of day", "For your information", "In my opinion") |
| General-English-Sentences Hugging Face | Direct download | Plain, everyday English sentences (emails, messages, general prose and homophone pairs) read aloud by one speaker, with no Hebrew or Yiddish words, Israeli names or tech vocabulary |
| Hard-ASR-Sentences Hugging Face | Accept terms | English sentences speech recognition finds hard, read by one speaker: Harvard sentences (IEEE 1969, public domain), tongue twisters, homographs, garden-path sentences, rare words, dense numbers and units, unusual proper nouns |
| Hebrish-TTS-Voice-Recordings Hugging Face | Accept terms | |
| Homophone-Context-Sentences Hugging Face | Direct download | English sentences whose context settles which spelling of a homophone is meant (their/there, principal/principle, and 58 more sets), read aloud by one speaker Audio: Homophone-Context-Sentence-Bank. |
| Irish-English-Sentences Hugging Face | Accept terms | Everyday sentences in Irish English read by one speaker born in Ireland: vocabulary (craic, grand, giving out, press, messages), grammar (I'm after doing it, so it is, ye), Irish-language words used in English (Taoiseach, Dáil, Gardaí, slái |
| Israeli-Name-Sentences Hugging Face | Accept terms | English messages and emails that mention an Israeli or Jewish person's name, a first name (Naftali, Yotam, Tzlil, Bat-El) or a surname (Buskila, Ifergan, Weinberg), written in its common Israeli-English spelling, read aloud by one speaker |
| Jewish-Religious-Terminology-Sentences Hugging Face | Needs approval | English sentences with Jewish religious vocabulary written in Latin letters: Shabbat and the festivals, prayer and the synagogue, Torah portions, Talmud and halacha, kashrut, and life-cycle events, read aloud by one speaker |
| Judaica-Sentences Hugging Face | Direct download | Real sentences from traditional Jewish works in English, read by one speaker, Hebrew terms in romanized form: Tanach (Metsudah), Mishnah (Mishnah Yomit, Joshua Kulp; Shraga Silverstein; Sefaria Community Translation; Eighteen Treatises from |
| Linux-Admin-Sentences Hugging Face | Direct download | Dictation-style English sentences about Linux desktop and server administration (KDE Plasma, GNOME, systemd, storage, networking, homelab), read aloud by one speaker Audio: Linux-Admin-Sentence-Bank. |
| Model-Checks Hugging Face | Accept terms | Takes checked by a speech recognition model while they were recorded, with the model's transcript, the speaker's verdict on it and, once judged, a second, machine verdict |
| Noisy-Condition-Sentences Hugging Face | Accept terms | Every take from the main dataset recorded in noise, whatever its category: outdoors, in a noisy room, with traffic, other people talking, background noise or a crying baby |
| Paragraph-Dictation-Passages Hugging Face | Direct download | Experimental |
| Personal-Vocabulary-Sentences Hugging Face | Direct download | English sentences built around names a speech model gets wrong for one speaker: his own name, his company and his project names (Daniel Rosehill, DSR Holdings, Carrot Cake AI, the podcast My Weird Prompts and its hosts Corn and Herman Poppl Audio: Personal-Vocabulary-Sentence-Bank. |
| Pronunciation-Variants Hugging Face | Accept terms | Takes in which a word was deliberately said one particular way, from the main dataset, stratified by `variant_category` |
| Read-Aloud-Voice-Sentences Hugging Face | Accept terms | Every take from the capture app; all speech models train on a pinned revision of it, selected by blend. Gated. |
| Repeated-Word-Sentences Hugging Face | Accept terms | Experimental |
| Self-Correction-Sentences Hugging Face | Direct download | One speaker correcting himself mid-sentence ("send it Tuesday, no, Wednesday") |
| Singing-Sentences Hugging Face | Direct download | Experimental |
| Smart-Home-Sentences Hugging Face | Direct download | Dictation-style English sentences about home automation (Home Assistant, Zigbee, Z-Wave, Matter, Thread, ESPHome, MQTT and device brands), read aloud by one speaker Audio: Smart-Home-Sentence-Bank. |
| Spoken-Aside-Sentences Hugging Face | Direct download | One speaker dictating a sentence and, partway through, saying something to someone in the room ("come in, Hannah") |
| Synthetic-Codeswitching-Voices-Pilot Hugging Face | Direct download | 100 synthetic clips (2026-10-28): English sentences with one Hebrew term. |
| Tech-Sentences Hugging Face | Direct download | Dictation-style English sentences dense in software, DevOps, cloud, self-hosting and AI vocabulary (product names, CLI tools, acronyms, model names), read aloud by one speaker Audio: AI-Tech-Vocabulary-Sentence-Bank. |
| Tech-Sentences-For-ASR-Training Hugging Face | Direct download | Early tech sentences with audio; a prompt source for the bank. |
| Voice-Command-Sentences Hugging Face | Direct download | One speaker reading desktop voice commands for a KDE Plasma voice-command agent: launching, switching to and closing apps (Konsole, Kate, Dolphin, Kdenlive, Audacity), browser and terminal tabs, windows and virtual desktops ("Move this to d |
| Weird-Voice-Sentences Hugging Face | Direct download | Experimental |
| Yiddish-Sentences Hugging Face | Accept terms | English sentences with Yiddish words written in Latin letters (shul, daven, cholent, mensch, freier, tachles), the way English speakers in Israel use them, read aloud by one speaker |
Benchmarks (2)
Held out from training.
| Dataset | Access | What it is |
|---|---|---|
| English-Hebrew-Codeswitching-ASR-Benchmark Hugging Face | Accept terms | Held-out test sets (v1, v2) the models are scored on; never trained on. |
| Hebrish-Validation Hugging Face | Direct download | Validation benchmark for the English-Hebrew code-switching speech models. |
Training mixes (2)
Exports of a blend, for reproducing a run.
| Dataset | Access | What it is |
|---|---|---|
| Parakeet-V2-Blend-All Hugging Face | Needs approval | Export of the `all` training mix, for reproducing Sabich runs. |
| Parakeet-V2-Blend-No-Hebrew Hugging Face | Needs approval | The exact data of training run r7-general, Daniel's English model without code-switching (not a Hebrish model). |
Error lists (2)
Words the models get wrong, which drive new sentences.
| Dataset | Access | What it is |
|---|---|---|
| ASR-Error-List Hugging Face | Accept terms | Words the models mishear in Daniel's own dictation, with the wrong forms; drives the capture app's priority words. Sabich v2.1 repeated its takes 3 times. |
| ASR-Error-Prediction Hugging Face | Direct download | Predicted mistranscription candidates for planning new sentences. |
Annotations (2)
Labels on the recordings.
| Dataset | Access | What it is |
|---|---|---|
| Room-Tone-Annotations Hugging Face | Direct download | Room-tone recordings and labels for noise conditions. |
| Take-Labels Hugging Face | Direct download | Hand labels for takes in the master dataset. |
Reference data (6)
Sources the other datasets draw on.
| Dataset | Access | What it is |
|---|---|---|
| Background-Noise Hugging Face | Accept terms | Background noise recordings. |
| Example-Audio-Sentences Hugging Face | Accept terms | Format reference for audio datasets: the project's standard columns, with a schema file and three real takes with the place block blanked. Audio: Example-Sentence-Bank. |
| Example-Sentence-Bank Hugging Face | Accept terms | Format reference for sentence datasets: the project's standard columns, with a schema file and a handful of invented neutral sentences. Audio: Example-Audio-Sentences. |
| Hebrew-Term-Pronunciations-IPA Hugging Face | Accept terms | How Hebrew words are pronounced inside English speech, in IPA, measured from real recordings. |
| Israeli-Street-Names Hugging Face | Accept terms | Israel's national street list, source for the street-name prompts. |
| Speakers Hugging Face | Direct download | Accreditation: one row per contributor whose voice is in the project. |
Other (2)
| Dataset | Access | What it is |
|---|---|---|
| English-Hebrew-Mixed-Sentences Hugging Face | Direct download | Card only, no data files (checked 2026-11-05). See Hebrish/English-Hebrew-Mixed-Sentences. |
| Hebrish-Ideas Hugging Face | Direct download | Ideas recorded in the capture app. |
Related (7)
Outside the Hebrish speech models.
| Dataset | Access | What it is |
|---|---|---|
| Annotated-Voice-Notes Hugging Face | Direct download | Real voice notes with annotations. |
| ASR-Cleanup-Before-After Hugging Face | Direct download | ASR cleanup before and after pairs. Separate project. |
| ASR-Paragraph-Breaks Hugging Face | Direct download | Hand-annotated paragraph breaks for flat speech-to-text output. Separate project. |
| Proofed-ASR-Cleanup-Training-Data Hugging Face | Direct download | Audio paired with clean, proofed text, for ASR cleanup training. Separate project. |
| Voice-Note-Audio Hugging Face | Direct download | Real-world voice recordings with transcripts and annotations. |
| Voice-Notes Hugging Face | Direct download | Real voice notes. |
| Wake-Words Hugging Face | Direct download | Jewish and Israeli first-name wake words, for local wake word models. |
Planned (37)
Registered themes whose dataset appears once takes are recorded: AI-Integration-Sentences, AI-Model-Sentences, API-Ops-Sentences, Colloquial-Reduction-Sentences, Data-Tool-Sentences, Date-Time-Sentences, Dev-Stack-Sentences, Dropped-Article-Sentences, Entity-Dense-Sentences, False-Start-Sentences, Fast-Speech-Sentences, Fragment-Sentences, GitHub-Trending-Sentences, Hardware-Sentences, Idiom-Sentences, Implied-Quote-Sentences, Israeli-Business-Sentences, Israeli-Place-Sentences, Israeli-Street-Sentences, Israeli-TV-Sentences, Latin-Phrase-Sentences, NATO-Alphabet-Sentences, Name-Cipher-Sentences, Numerics-Sentences, October-7-Memorial-Sentences, Open-Source-Sentences, Phone-Number-Sentences, Procurement-Sentences, Run-On-Sentences, Shingle-Sentences, Slow-Speech-Sentences, Spelled-Out-Sentences, Spoken-Punctuation-Sentences, Tanach-Verse-Sentences, Thinking-Pause-Sentences, Web-Address-Sentences, Whispered-Sentences.