whisper-tiny-uz-onnx

ONNX (int8) build of mirodil/whisper-tiny-uzbek for in-browser Uzbek speech recognition with transformers.js.

The decoder is exported unmerged, int8-quantized per branch, then re-merged — so the whole model runs as ~60 MB instead of the ~330 MB of the small-sized Uzbek checkpoints.

File	Size
onnx/encoder_model_quantized.onnx	~10 MB
onnx/decoder_model_merged_quantized.onnx	~50 MB

Output is Latin-script Uzbek.

Usage (transformers.js)

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

const asr = await pipeline(
  "automatic-speech-recognition",
  "davron112/whisper-tiny-uz-onnx",
  { dtype: "q8" }
);

const out = await asr(audio, { language: "uzbek", task: "transcribe" });
console.log(out.text);

Attribution

Fine-tuned weights © the authors of mirodil/whisper-tiny-uzbek. This repo only re-packages them as quantized ONNX for the browser runtime.