instush/ivrit-whisper-large-v3-turbo-timestamped-onnx

ivrit.ai whisper-large-v3-turbo — ONNX with word-level timestamps

Unofficial ONNX conversion of ivrit-ai/whisper-large-v3-turbo (a Hebrew fine-tune of openai/whisper-large-v3-turbo), for in-browser use with 🤗 Transformers.js.

What was modified relative to the source model (per Apache-2.0 §4(b)):

• Exported to ONNX with output_attentions=True, so the decoder graph exposes cross_attentions.* outputs. This is what enables word-level timestamps (return_timestamps: "word") in Transformers.js — the same technique as the onnx-community/*_timestamped exports. The official ivrit-ai/whisper-large-v3-turbo-onnx export does not include these outputs and therefore supports segment-level timestamps only.
• Quantized variants generated (fp16, q4); weights are otherwise unchanged.

All credit for the model itself goes to ivrit.ai (fine-tune, ~5,050 h of Hebrew training data) and OpenAI (base model).

Usage (Transformers.js)

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

const transcriber = await pipeline(
  "automatic-speech-recognition",
  "instush/ivrit-whisper-large-v3-turbo-timestamped-onnx",
  {
    device: "webgpu", // large model — WebGPU strongly recommended
    dtype: { encoder_model: "fp16", decoder_model_merged: "q4" },
  },
);

const out = await transcriber(audioFloat32Array16kHz, {
  language: "he",
  task: "transcribe",
  chunk_length_s: 30,
  stride_length_s: 5,
  return_timestamps: "word",
});

Notes:

• The model is Hebrew-only; per the source model card, set language: "he" explicitly (the fine-tune degraded language auto-detection).
• ~1.6 GB download at the recommended dtypes. Too large for the WASM backend on typical hardware — use WebGPU.

License

Apache License 2.0 — same as the source models. See the license text in this repo / the HF license tag. This is not an official ivrit.ai or OpenAI release.