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)