Hirecentive-D3l/Whisper-Hindi2Hinglish-Swift-ONNX

Whisper-Hindi2Hinglish-Swift (ONNX, q8)

ONNX conversion of Oriserve/Whisper-Hindi2Hinglish-Swift for in-browser use with Transformers.js (dtype: 'q8', WASM/CPU friendly).

The base model is openai/whisper-base (73M params) fine-tuned by Oriserve on ~550 hours of noisy Indian-accented Hindi. It transcribes Hindi-English code-mixed speech into romanized Hinglish — English words stay in Latin script.

Conversion

• —Exported with optimum-cli export onnx --task automatic-speech-recognition-with-past
• —Quantized with onnxruntime dynamic quantization (QUInt8 weights, EnableSubgraph for the merged decoder), matching the layout Transformers.js expects for dtype: 'q8':
• —onnx/encoder_model_quantized.onnx (23 MB)
• —onnx/decoder_model_merged_quantized.onnx (79 MB)

Usage (Transformers.js)

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

const asr = await pipeline(
  'automatic-speech-recognition',
  'TechHirecentive/Whisper-Hindi2Hinglish-Swift-ONNX',
  { dtype: 'q8' },
);
const { text } = await asr(audioFloat32Array16kHz, {
  language: 'hi',
  task: 'transcribe',
});

License

Apache-2.0, same as the upstream fine-tune. All credit for the model weights to Oriserve; this repo only repackages them as quantized ONNX.