edersonff/whisper-large-v3_timestamped

whisper-large-v3_timestamped

ONNX export of openai/whisper-large-v3 with output_attentions=True for word-level timestamp support in transformers.js.

Usage

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

const transcriber = await pipeline("automatic-speech-recognition", "edersonff/whisper-large-v3_timestamped", {
  dtype: { encoder_model: "fp32", decoder_model_merged: "q8" },
  device: "webgpu",
});

const result = await transcriber(audio, {
  return_timestamps: "word",
  chunk_length_s: 29,
  stride_length_s: 5,
  language: "en",
});

Export details

• —Exported with attn_implementation="eager" and output_attentions=True
• —alignment_heads configured in generation_config.json
• —32 encoder attentions + 32 cross-attentions per decoder layer
• —Quantization variants: fp32 (original), q8 (quantized)
• —WebGPU recommended (large model, ~10GB fp32 / ~3GB q8)