File size: 5,280 Bytes
ea06575
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
861f5fa
ea06575
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
---
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, **专讜讝讛讬诇**. 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 专讜讝讛讬诇 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% |
| 专讜讝讛讬诇 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 |
|---|---|---|
| 讗驻砖专 诇专砖讜诐 讗转 讝讛 注诇 砖诐 专讜讝讛讬诇? | 讗驻砖专 诇专砖讜诐 讗转 讝讛 注诇 砖诐 专讜讝讬诇? | 讗驻砖专 诇专砖讜诐 讗转 讝讛 注诇 砖诐 专讜讝讛讬诇? |
| 讬砖 诇讬 转讜专 讗爪诇 讛专讜驻讗讛, 注诇 砖诐 讚谞讬讗诇 专讜讝讛讬诇. | 讬砖 诇讬 转讜讗专 讗爪诇 讛专讜驻讗讛 注诇 砖诐 讚谞讬讗诇 专讜讝讛专. | 讬砖 诇讬 转讜讗专 讗爪诇 讛专讜驻讗讛 注诇 砖诐 讚谞讬讗诇 专讜讝讛讬诇. |
| 讛拽讜谞讟讬讬谞专 谞讜驻诇 讻诇 驻注诐 砖讛讜讗 注讜诇讛. | 讛拽讜谞讟讬讬谞专 讗讜 驻诇诐 讻诇 驻注诐 砖讛讜讗 注讜诇讛. | 讛拽讜谞讟讬讬谞专 谞讜驻诇 讻诇 驻注诐 砖讛讜讗 注讜诇讛. |
| 诪住讚 讛谞转讜谞讬诐 讗讬讟讬 讛讬讜诐, 讗讜诇讬 讞住专 讗讬谞讚拽住. | 讞讬驻讛. 诪住注 讝讛 谞转讜谞讬诐 讗讬讟讬 讛讬讜诐, 讗讜诇讬 讛住专 讗讬谞讚拽住. | 诪住注 讝讛 谞转讜谞讬诐 讗讬转讬 讛讬讜诐, 讗讜诇讬 讛住专 讗讬谞讚拽住. |
| 讛诐 讗诪专讜 诇讬 砖诇讗 讬讻讜诇讬诐 诇注砖讜转 讗转 讝讛. | 讛诐 讗诪专讜 诇讬 砖诇讗 讬讜讻诇讬诐 诇注砖讜转 讗转 讝讛. | 讛诐 讗诪专讜 诇讬 砖诇讗 讬讻讜诇讬诐 诇注砖讜转 讗转 讝讛. |
| 讗转讛 诇讗 讬讻讜诇 诇讛转讞讘讗 诪诪谞讬. | 讗转讛 诇讗 讬讻讜诇 诇讛转讞讘讗 诪诪谞讬. | 讗转讛 诇讗 讬讻讜诇 诇讛转讞讘讗 诪诪谞讜. |
| 讗转讛 爪专讬讱 诇砖诪讜专 注诇 注爪诪讱. | (same) | (same) |
| 讗谞讬 专讜讗讛 讻诪讛 讝讛 诪住讜讻谉. | (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.