Zum Inhalt springen
ModelAufgelistetExtern
Herunterladen

Whisper Small Hu

openai/whisper-small fine-tuned for Hungarian speech recognition on Mozilla Common Voice 26.0.

Quellehuggingface.co/SubZtep/whisper-small-hu/tree/3a6023e2806992205081f08e4ede704c2ce33bc2
ArtText modelBasismodellopenai/whisper-smallVersionBBewertung

Modellquelle

Art
Text model
Dateigröße
922 MB
Basismodell
openai/whisper-small
Version
vmodel
Aktualisiert
26. Okt. 2026
Downloads
89
Quelle
Hugging Face

Quellenbeschreibung

whisper-small-hu

openai/whisper-small fine-tuned for Hungarian speech recognition on Mozilla Common Voice 26.0.

WERCER
openai/whisper-small (zero-shot)46.7911.70
whisper-small-hu15.623.33

Both models evaluated on the full Common Voice 26.0 Hungarian test split (13,127 clips) with the same pipeline: greedy decoding, forced <|hu|><|transcribe|> prompt, predictions and references lowercased and whitespace-normalized before scoring.

Usage

from transformers import pipeline

asr = pipeline("automatic-speech-recognition", model="SubZtep/whisper-small-hu")
print(asr("audio.mp3", generate_kwargs={"language": "hungarian", "task": "transcribe"})["text"])

Training

  • Data: Common Voice 26.0 Hungarian train split (62,284 clips), silence-trimmed, 16 kHz
  • Schedule: 3 epochs (~11,700 steps), effective batch size 16 (8 × 2 gradient accumulation)
  • Optimizer: AdamW, lr 1e-5, linear decay with 500 warmup steps, fp16, gradient checkpointing
  • Hardware: single NVIDIA T4 (Google Colab)
  • Selection: best checkpoint by WER on a fixed 1,500-sample subset of the dev split

Training code: https://github.com/SubZtep/whisper-hu

Training progress

WER on the 1,500-sample dev subset during training:

EpochStepTrain LossVal LossWER
0.2610000.33020.328430.54
0.5120000.24900.267425.36
0.7730000.21300.233723.12
1.0340000.13800.201718.60
1.2850000.10110.196018.16
1.5460000.09230.186617.29
1.8070000.09130.177916.28
2.0680000.03900.176515.71
2.3190000.03880.178115.32
2.57100000.03940.176215.38
2.83110000.03690.174515.31

Framework versions

  • Transformers 4.44.2
  • PyTorch 2.11.0+cu128
  • Datasets 4.0.0
  • Tokenizers 0.19.1

Limitations

  • Trained on read speech (Common Voice sentences) — expect higher error rates on spontaneous conversation, noisy audio, or domain-specific vocabulary
  • WER is measured on lowercased text; the model itself outputs casing and punctuation, but their accuracy is not reflected in the score
  • Hungarian is agglutinative, so WER punishes single-suffix mistakes as whole-word errors — the CER of 3.33 is a better sense of how close transcripts are

Quellen

1 Quelle
Hugging FacePrimärVersion 3a6023e2806992205081f08e4ede704c2ce33bc2
huggingface.co/SubZtep/whisper-small-hu/tree/3a6023e2806992205081f08e4ede704c2ce33bc2

Verifiziert 26. Okt.

Modellartefakte

1 Artefakt

checkpoint-100/model.safetensors

safetensors · 922 MB · SHA-256 79ff233b2ac5…db7d · Hugging Face

Download file

Quellenauszüge

3 Auszüge
Readme7. Aug.huggingface.co/SubZtep/whisper-small-hu

--- language: hu license: mit library_name: transformers pipeline_tag: automatic-speech-recognition base_model: openai/whisper-small tags: - whisper - hungarian - common-voice metrics: - wer - cer model-index: - name: whisper-small-hu results: - task: type: automatic-speech-recognition name: Automatic Speech Recognition dataset: name: Common Voice 26.0 (Hungarian) type: mozilla-foundation/common_voice_26_0 config: hu split: test metrics: - type: wer value: 15.62 name: WER - type: cer value: 3.33 name: CER --- # whisper-small-hu [openai/whisper-small](https://huggingface.co/openai/whisper-small) fine-tuned for Hungarian speech recognition on [Mozilla Common Voice 26.0](https://commonvoice.mozilla.org/). | | WER | CER | |---|---|---| | openai/whisper-small (zero-shot) | 46.79 | 11.70 | | **whisper-small-hu** | **15.62** | **3.33** | Both models evaluated on the full Common Voice 26.0 Hungarian test split (13,127 clips) with the same pipeline: greedy decoding, forced `<|hu|><|transcribe|>` prompt, predictions and references lowercased and whitespace-normalized before scoring. ## Usage ```python from transformers import pipeline asr = pipeline("automatic-speech-recognition", model="SubZtep/whisper-small-hu") print(asr("audio.mp3", generate_kwargs={"language": "hungarian", "task": "transcribe"})["text"]) ``` ## Training - **Data:** Common Voice 26.0 Hungarian `train` split (62,284 clips), silence-trimmed, 16 kHz - **Schedule:** 3 epochs (~11,700 steps), effective batch size 16 (8 × 2 gradient accumulation) - **Optimizer:** AdamW, lr 1e-5, linear decay with 500 warmup steps, fp16, gradient checkpointing - **Hardware:** single NVIDIA T4 (Google Colab) - **Selection:** best checkpoint by WER on a fixed 1,500-sample subset of the `dev` split Training code: https://github.com/SubZtep/whisper-hu ### Training progress WER on the 1,500-sample `dev` subset during training: | Epoch | Step | Train Loss | Val Loss | WER | |:-----:|:-----:|:----------:|:--------:|:-------:| | 0.26 | 1000 | 0.3302 | 0.3284 | 30.54 | | 0.51 | 2000 | 0.2490 | 0.2674 | 25.36 | | 0.77 | 3000 | 0.2130 | 0.2337 | 23.12 | | 1.03 | 4000 | 0.1380 | 0.2017 | 18.60 | | 1.28 | 5000 | 0.1011 | 0.1960 | 18.16 | | 1.54 | 6000 | 0.0923 | 0.1866 | 17.29 | | 1.80 | 7000 | 0.0913 | 0.1779 | 16.28 | | 2.06 | 8000 | 0.0390 | 0.1765 | 15.71 | | 2.31 | 9000 | 0.0388 | 0.1781 | 15.32 |...

Note26. Okt.huggingface.co/SubZtep/whisper-small-hu

Source context: 89 downloads · 0 likes · Pipeline automatic-speech-recognition · Library transformers · Repo SubZtep/whisper-small-hu

Note7. Aug.huggingface.co/SubZtep/whisper-small-hu

Source context: 528 downloads · 0 likes · Pipeline automatic-speech-recognition · Library transformers · Repo SubZtep/whisper-small-hu