asounimelb / whisper-tiny-ckb-onnx

Whisper Tiny: Central Kurdish (ONNX)

openai/whisper-tiny (39M parameters), fine-tuned for Central Kurdish (Sorani, ckb ) speech recognition and exported to ONNX for use with Transformers.js (browser / Node.js) and ONNX Runtime.

Output is in the standard Kurdish Arabic script (e.g. کوردی ).

Whisper has no Central Kurdish language token, so this model was fine-tuned with the English token <|en|> as a placeholder. Always decode with language: "en" , task: "transcribe" . Other language settings will give poor results.

Other sizes

All three were fine-tuned the same way on the same data:

Files

The decoder is a merged decoder (with and without KV cache in one graph). Choose q8 for accuracy or q4 for a smaller download. In spot checks both produced near-identical transcripts.

Usage

Transformers.js (browser or Node.js)

npm install @huggingface/transformers

import { pipeline } from "@huggingface/transformers";

const transcriber = await pipeline(
  "automatic-speech-recognition",
  "asounimelb/whisper-tiny-ckb-onnx",
  { dtype: { encoder_model: "fp32", decoder_model_merged: "q8" } }  // or "q4"
);

// Accepts a URL, or a Float32Array of 16 kHz mono audio
const output = await transcriber("https://example.com/audio.wav", {
  language: "en",      // required: the model was trained on the <|en|> token
  task: "transcribe",
});
console.log(output.text);

For audio longer than 30 seconds, add chunk_length_s: 30, stride_length_s: 5 to the call options.

Training

Dataset

The dataset is single-speaker: a male native speaker (Aso Mahmudi) with a Mariwan accent, recorded in a home studio with a USB condenser microphone. The transcripts come from two kinds of source:

• the full text of the book Mesele-y Wijdan by Ahmad Mukhtar Jaff (1896–1935), about 49 minutes of audio;
• assorted texts from Kurdish websites (news, sport, general topics).

All transcripts were manually checked against the recordings.

Limitations

• Single speaker, small dataset. The training data is about 2 hours from one speaker. Expect lower accuracy on other speakers, accents and dialects (e.g. Sulaimani, Erbil, Kirkuk), and on spontaneous or conversational speech, noisy audio, and telephone-quality audio.
• Read speech only. The model learned punctuation and formatting from edited text, so it may insert punctuation that a speaker did not clearly signal.
• Word-spacing errors. Many remaining errors are spacing variants (e.g. بە کار vs بەکار ), a common ambiguity in Kurdish orthography.
• No formal benchmark. No evaluation on an independent test set has been published for this model. Evaluate it on your own data before relying on it.
• The usual Whisper caveats apply, including possible hallucinated text on silence or non-speech audio.

Citation

If you use this model, please cite the original Whisper paper:

@misc{radford2022whisper,
  title  = {Robust Speech Recognition via Large-Scale Weak Supervision},
  author = {Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman, Greg and McLeavey, Christine and Sutskever, Ilya},
  year   = {2022},
  url    = {https://arxiv.org/abs/2212.04356}
}