Version 3 of my English-Hebrew code-switching fine-tune of NVIDIA's Parakeet TDT 0.6B v2 is out: Parakeet-TDT-0.6B-v2-English-Hebrew-Codeswitching.
Hugging Face Model danielrosehill/Parakeet-TDT-0.6B-v2-English-Hebrew-Codeswitching A fine-tune of nvidia/parakeet-tdt-0.6b-v2 for English speech with Hebrew words dropped in, the way English speakers in Israel talk: "pick up bamba from the makolet", "I have an appointment at Misrad Hapnim tomorrow".… ↓ 564Automatic speech recognitionCC-BY-4.0Updated Oct 2026 View model on Hugging FaceThe model transcribes English speech that contains Hebrew words, the way English speakers in Israel actually talk ("I have to go to Bituach Leumi tomorrow", "Can you grab some bamba from the makolet?"), and writes each Hebrew word in one consistent Latin spelling. The background, method and earlier results are in the v2 post. This post covers what changed in v3.
Results
Scored on a held-out test split of 219 code-switched clips. The split is term-disjoint: 242 Hebrew terms never appear in training, so the "never seen" figure measures generalisation, not memorisation.
| Test set | Stock v2 | v3 |
|---|---|---|
| Code-switching WER | 10.9% | 5.3% |
| Hebrew terms seen in training, correct | 29% | 66% |
| Hebrew terms never seen in training, correct | 27% | 48% |
| Long-form dictation WER (~3-minute files) | 9.0% | 5.0% |
| Speaker's own plain English WER | 3.5% | 2.9% |
| LibriSpeech test-clean WER | 1.2% | 1.0% |
Code-switching errors are halved compared with the stock model, and English accuracy is no worse on either the speaker's own English or LibriSpeech.
Every version on the same test
The test split grew from 121 to 219 clips for this release, so the v2 numbers published earlier are not comparable with these. Instead, the stock model and all three versions were re-scored on the new split:
| Measure | Stock v2 | v1 | v2 | v3 |
|---|---|---|---|---|
| Code-switching training audio | – | 77 min | 120 min | 196 min |
| Code-switching WER | 10.9% | 6.0% | 5.8% | 5.3% |
| Terms seen in training, correct | 29% | 57% | 60% | 66% |
| Terms never seen, correct | 27% | 47% | 47% | 48% |
| Long-form WER | 9.0% | 5.9% | 5.6% | 5.0% |
What 76 more minutes bought
v3 added 76 minutes of code-switched recordings (120 to 196 minutes), almost all made on the phone that runs the model. The gains are where the new data is:
Words with recordings improved: 60% to 66% correct.
Long-form dictation improved: 5.6% to 5.0% WER.
Words the model has never heard barely moved: 47% to 48%, one extra hit out of 176.
That last point is the main finding of three versions. The big jump on unseen words happened in v1, when the model learned the general mapping from my Hebrew pronunciation to romanized spelling. Since then, more audio improves the words it has heard, not new ones. Unseen-word recognition is stuck just under 50%.
So the practical lesson is about which words get recorded, not how many minutes. Before fine-tuning, the stock model never once got 706 of 1,139 Hebrew terms right across the recordings, and only 258 were always right:
The capture app now weights recording toward those never-right terms and toward terms that have only one or two takes.
Training
| Item | Detail |
|---|---|
| Code-switched speech | 196 min (1,921 clips) |
| Speaker's own English | 43 min |
| LibriSpeech replay | 28 min, with cased and punctuated LibriSpeech-PC transcripts |
| Mix by duration | 73% code-switching, 16% own English, 11% replay |
| Run | Full fine-tune from stock v2, 20 epochs, one L40S GPU, 29 minutes |
The English share follows the code-switching literature: fine-tuning on mixed-language data alone degrades monolingual accuracy, and around 10% replay of original-domain data limits the damage. The exact ratio is still untuned; the comparison of English shares moves to v4.
Downloads
| Use | File |
|---|---|
| Android dictation (Offline Voice Input) | GGUF Q8_0, 730 MB |
| Low-memory Android | GGUF Q4_K_M, 475 MB |
| GPU inference or further fine-tuning (NeMo) | .nemo |
The download links always point at the newest version; the model card's version history has pinned links to v1 and v2. Other formats are listed on the project page, and you can try the model against stock Parakeet in the demo Space.
Hugging Face Space danielrosehill/English-Hebrew-Codeswitching-Parakeet Versions, downloads, results and notes for the fine-tune StaticLiveCC-BY-4.0Updated Oct 2026 Open the Space Hugging Face Space danielrosehill/English-Hebrew-Codeswitching-Parakeet-Demo Stock Parakeet v2 vs an English-Hebrew code-switching tune GradioLiveCC-BY-4.0Updated Oct 2026 Open the SpaceWhat's next
v4 at 270 minutes of code-switched audio, with runs at 0%, 10%, 20% and 30% English to settle how much monolingual data is actually needed.
Unseen words: recording targeted at terms the stock model never gets right.
Numbers written as digits ("DeepSeek 4.1", not "four point one"): sentences are being recorded now.
This is still a single-speaker model, trained on my voice. Recordings from other English speakers in Israel are the most useful contribution anyone could make, and integration into dictation tools is very welcome. For either, get in touch via danielrosehill.com/contact.
The model and datasets are released under CC-BY-4.0. All recordings, including the audio, are public in Read-Aloud-Voice-Sentences, and progress is tracked on GitHub.
