--- license: cc-by-4.0 language: - he pipeline_tag: automatic-speech-recognition base_model: Hebrish/Zipformer2-66M-Streaming-He tags: - speech - asr - hebrew - streaming - zipformer - transducer - icefall - k2 - sherpa-onnx - proof-of-concept --- # Gefilte (PoC-1) **Gefilte** is Daniel Rosehill's Hebrew speech-recognition fine-tune: a Hebrew model adapted to his own non-native (Anglo oleh) accent and the vocabulary he actually needs. The name is a joke: Daniel's other models are named after food, and an Ashkenazi immigrant's Hebrew deserved gefilte fish. > **Proof of concept.** PoC-1 was trained on **4.9 minutes** of one speaker reading Hebrew from nikud. Its only > purpose is to show that a few minutes of recordings can teach the base a word it has never heard. The test word is > Daniel's surname as he writes it in Hebrew, **רוזהיל**. It is not a general improvement on the base, and it is not > recommended for anyone else's voice. It is not part of the Hebrish code-switching models. ## Base [`Hebrish/Zipformer2-66M-Streaming-He`](https://huggingface.co/Hebrish/Zipformer2-66M-Streaming-He): the Hebrish project's unchanged fork of [`ivrit-ai/zipformer2-66m-streaming-he`](https://huggingface.co/ivrit-ai/zipformer2-66m-streaming-he), **developed by Yakov Kolani** and trained on [ivrit.ai](https://www.ivrit.ai) Hebrew audio (CC BY 4.0). It is a 66M Zipformer2 streaming transducer (icefall / k2). Gefilte keeps its architecture, tokenizer (`bpe.model`, 1,493 tokens) and streaming configuration (chunk 64 frames = 1.28 s, left context 256 frames), and changes only the weights. ## Data Takes from the `takes` tab of [`danielrosehill/Gefilte-Hebrew-Sentences`](https://huggingface.co/datasets/danielrosehill/Gefilte-Hebrew-Sentences), recorded in Daniel's capture app. He reads each Hebrew sentence from a pointed (nikud) version, with [DictaBERT menaked](https://huggingface.co/dicta-il/dictabert-large-char-menaked) supplying the nikud. The prompts are graded Tatoeba sentences plus a priority set: his surname and spoken dev and AI Hebrew. Targets are the unpointed Hebrew `text`, with the punctuation the base vocabulary can write (`, . ? ! : ; " ( ) -`). | | Takes | Minutes | |---|---|---| | Trained on | 84 | 4.9 | | Held out (never trained on) | 8 | | The held-out takes are the dataset's test rows, the two newest takes containing רוזהיל and the two newest other priority takes. ## Training The icefall `egs/librispeech/ASR/zipformer` recipe (commit `d7c83d52`): pruned RNN-T loss, ScaledAdam with Eden, base lr 0.002, 12 epochs, batch 8. It ran on one NVIDIA L40S (Nebius) on 2026-11-17. Code: `training/gefilte/` in Daniel's training repository. ## Results (held-out takes, greedy streaming decode) | | Base | Gefilte PoC-1 | |---|---|---| | WER | 21.2% | 13.5% | | CER | 7.7% | 3.0% | | רוזהיל recognised | 0 of 2 | 2 of 2 | There are only 8 held-out takes, from one speaker, so treat these numbers as a sanity check, not a benchmark. Every held-out take (reference, then base, then Gefilte): | Reference | Base | Gefilte PoC-1 | |---|---|---| | אפשר לרשום את זה על שם רוזהיל? | אפשר לרשום את זה על שם רוזיל? | אפשר לרשום את זה על שם רוזהיל? | | יש לי תור אצל הרופאה, על שם דניאל רוזהיל. | יש לי תואר אצל הרופאה על שם דניאל רוזהר. | יש לי תואר אצל הרופאה על שם דניאל רוזהיל. | | הקונטיינר נופל כל פעם שהוא עולה. | הקונטיינר או פלם כל פעם שהוא עולה. | הקונטיינר נופל כל פעם שהוא עולה. | | מסד הנתונים איטי היום, אולי חסר אינדקס. | חיפה. מסע זה נתונים איטי היום, אולי הסר אינדקס. | מסע זה נתונים איתי היום, אולי הסר אינדקס. | | הם אמרו לי שלא יכולים לעשות את זה. | הם אמרו לי שלא יוכלים לעשות את זה. | הם אמרו לי שלא יכולים לעשות את זה. | | אתה לא יכול להתחבא ממני. | אתה לא יכול להתחבא ממני. | אתה לא יכול להתחבא ממנו. | | אתה צריך לשמור על עצמך. | (same) | (same) | | אני רואה כמה זה מסוכן. | (same) | (same) | WER and CER count words only: punctuation is removed from the reference and the output alike. On the 84 training takes Gefilte scores 1.1% WER, so it has memorised them; only the held-out rows above say anything. ## Files - `model.pt`: icefall state dict under `model`, config under `config`, tokenizer under `tokenizer`, the same layout as the base. - `bpe.model`: SentencePiece tokenizer, unchanged from the base. - `sherpa-onnx/`: `encoder.int8.onnx`, `decoder.int8.onnx`, `joiner.int8.onnx`, `tokens.txt` for [sherpa-onnx](https://github.com/k2-fsa/sherpa-onnx) (`model_type="zipformer2"`). They are used exactly like the base's files; see the base card for code. ## Licence and credit CC BY 4.0, as the base. Credit: Yakov Kolani and [ivrit.ai](https://www.ivrit.ai) for the base model and its data; [Dicta](https://dicta.org.il) for the nikud model used to prepare the prompts.