ishaan-g/whisper-hinglish-swift_timestamped

Whisper Hindi2Hinglish Swift (ONNX, word timestamps)

ONNX weights of Oriserve/Whisper-Hindi2Hinglish-Swift (a fine-tune of openai/whisper-base that writes Hindi and Hinglish speech in Roman script) for Transformers.js, exported with cross-attention outputs so return_timestamps: "word" works. Used by CaptionsEasy to caption videos in the browser.

The model itself is unchanged: all credit to Oriserve. Licensed Apache-2.0, like the original.

Files

file	use
onnx/encoder_model_fp16.onnx + onnx/decoder_model_merged_q4f16.onnx	WebGPU
onnx/encoder_model_quantized.onnx + onnx/decoder_model_merged_quantized.onnx	WASM / CPU (8-bit)

The output projection shares its weights with the token embedding (as in onnx-community/whisper-base_timestamped), so downloads are the same size as stock Whisper base.

Usage

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

const asr = await pipeline("automatic-speech-recognition", "<this repo>", {
  device: "webgpu",
  dtype: { encoder_model: "fp16", decoder_model_merged: "q4f16" },
});
// Oriserve trained with the English language token; output is Roman-script Hinglish
const out = await asr(audio16kMono, { language: "en", task: "transcribe", return_timestamps: "word" });

Export: transformers.js scripts/convert.py --output_attentions (alignment heads of whisper-base), then fp16 / q8 / q4f16 quantization.