Model inventory
All 17 models, by area: speech to text, text to speech and accessory models. Which are maintained, how often they are updated, what each is for and where it runs.
Which one should I use? Every model here can still be downloaded, and access requests for the older and less maintained ones are still approved. The up-to-date models (marked Recommended) are strongly recommended: they are the ones still being trained, so they recognise the widest vocabulary.
Recommended: maintained and updated. Available as is: downloadable, updated rarely or never. Legacy: replaced by a newer model. Separate project: listed for reference, not a Hebrish model. Latest versions of the models in the apps are read from the app download server when this page is built (2026-11-19).
Speech to text
Speech recognition (ASR): dictation, streaming and voice commands. All speech to text.
| Model | Status | Use case | Updates | Platforms | In the Hebrish apps | Latest | Last release | Weights |
|---|---|---|---|---|---|---|---|---|
| Sabich The default in the Hebrish apps: English with Hebrew words, tech and general vocabulary. | Recommended Active | Dictation (speech to text) | Rolling: retrained from the latest weights as new recordings arrive | Android, Linux, Windows, macOS | Android + Desktop | Sabich v5.1 | 2026-11-19 | Hugging Face |
| Falafel Two sizes: Falafel (0.6B) for desktops, Falafel 110M for phones and laptops. | Recommended Active | Voice commands | Updated alongside Sabich | Android, Linux, Windows | Android + Desktop | Falafel v1.1 | 2026-11-16 | Hugging Face |
| Jachnun Code-switching only. Sabich covers the same words and more. | Available as is Occasional updates Use Sabich instead | Dictation (speech to text) | Occasional updates, no schedule | Browser, Linux, Windows, macOS | No | Jachnun v1.0 | 2026-10-19 | Hugging Face |
| Sabich-Canary-Qwen-2.5B Sabich data on Canary-Qwen 2.5B, for desktop GPUs and servers. | Available as is Frozen Use Sabich instead | Dictation (speech to text) | None planned | GPU server | No | Sabich-Canary-Qwen-2.5B v1.0 | 2026-10-21 | Hugging Face |
| Sabich-Parakeet-110M Small Sabich for low-end phones and wearables. | Available as is Frozen Use Sabich instead | Dictation (speech to text) | None planned | Android, Low-power / wearables | No | Sabich-Parakeet-110M v1.0 | 2026-10-21 | Hugging Face |
| Sabich-Qwen3-ASR-1.7B One run comparing a Qwen3-ASR base with the Parakeet base. | Available as is Proof of concept Use Sabich instead | Dictation (speech to text) | None: a one-off experiment | GPU server | No | Sabich-Qwen3-ASR-1.7B v1.0 | 2026-10-21 | Hugging Face |
| Sabich-asr4all Smallest Sabich variants, streaming, for very low-power devices. | Available as is Frozen Use Sabich instead | Dictation (speech to text) | None planned | Low-power / wearables | No | Sabich-asr4all-S and -M v1.0 | 2026-10-21 | Hugging Face |
| Shakshuka Streaming: words appear while you speak. The streaming choice in the desktop app (Android support not released yet). Sabich is the recommended model: more accurate, and Shakshuka is not recommended until its accuracy catches up. | Available as is Occasional updates | Streaming dictation | Retrained now and then while its accuracy is worked on | Linux, Windows | Desktop | Shakshuka v2.2 | 2026-11-17 | Hugging Face |
| Whisper Hebrish The first Hebrish model, on Whisper large-v3-turbo. | Legacy Superseded Use Sabich instead | Dictation (speech to text) | None | GPU server | No | Version 2 | 2026-10-17 | Hugging Face |
| Whisper Hebrish Small ACFT Early Whisper fine-tune for FUTO Keyboard. | Legacy Superseded Use Sabich instead | Dictation (speech to text) | None | Android | No | small-en-en30-acft | 2026-10-17 | Hugging Face |
| Fritter English without Hebrew; a separate project, not a Hebrish model. | Active | Dictation (speech to text) | Regular incremental updates | Linux, Windows, macOS, Android | No | Fritter v1.2 | 2026-11-05 | Hugging Face |
| Gefilte Hebrew-script dictation in Daniel's accent, a separate project; the desktop app (1.042) can use it for Hebrew. | Proof of concept | Dictation (speech to text) | Early development | Linux, Windows, Low-power / wearables | Desktop | Gefilte PoC-1 | 2026-11-17 | Hugging Face |
Text to speech
Speech synthesis (TTS). All text to speech.
| Model | Status | Use case | Updates | Platforms | In the Hebrish apps | Latest | Last release | Weights |
|---|---|---|---|---|---|---|---|---|
| Meorav Code-switching text-to-speech in a designed voice; v3.0 is a proof of concept. Not in the Hebrish apps. | Recommended Active | Text to speech | New versions as the TTS work moves on | GPU server | No | Meorav v3.0 | 2026-11-15 | Hugging Face |
| Meorav-Daniel Text-to-speech in Daniel's own voice. | Available as is Frozen | Text to speech | None planned | GPU server | No | Meorav-Daniel v1.0 | 2026-11-15 | Hugging Face |
| Meorav-Multi-Voice Five designed voices in one model. | Available as is Frozen Use Meorav instead | Text to speech | None planned | GPU server | No | Meorav-Multi-Voice v1.0 | 2026-11-16 | Hugging Face |
Accessory models
Text tools that support the speech models. All accessory models.
| Model | Status | Use case | Updates | Platforms | In the Hebrish apps | Latest | Last release | Weights |
|---|---|---|---|---|---|---|---|---|
| Hebrew-Gendered-Sentence-Creator Text tool: writes Hebrew sentences already gendered for the speaker, for Hebrew text-to-speech. Stage 1 of 2 (gender-aware nikud planned). | Available as is Work in progress | Accessory (text tools) | In development | GPU server | No | v3 (work in progress) | 2026-11-18 | Hugging Face |
| Hebrew-Latin Token Classifier Text tool: finds romanized Hebrew words in English text, for text-to-speech. | Available as is Frozen | Accessory (text tools) | None planned | Linux, Windows, macOS, GPU server | No | v1 | 2026-10-15 | Hugging Face |
Machine-readable: /data/model-inventory.json. Help choosing a dictation model: Which model should I use?