feat: add language auto-detection for Whisper multilingual models

[jhlee111]

Closes #302

Summary

Implement automatic language detection for Whisper multilingual models. Previously, when no language was specified, the model silently defaulted to English. Now it runs a detection forward pass matching the Python transformers library behavior.

How it works

1. detect_language(input_features, generation_config) runs a single forward pass:

   • Encodes the audio through the encoder
   • Feeds only <|startoftranscript|> to the decoder
   • Masks logits to language token IDs only (from generation_config.lang_to_id)
   • Returns the argmax — the most likely language token ID

2. generate() calls detect_language() automatically when:

   • is_multilingual is true
   • No language is specified
   • inputs (audio features) are available

The detected language is set on generation_config.language, so _retrieve_init_tokens() picks it up normally. The init_tokens structure stays the same (SOT + lang + task + notimestamps), just with the detected language instead of hardcoded English.

Usage

// Auto-detect (no language specified)
const output = await model.generate({
  inputs: features,
  return_timestamps: true,
});

// Or call detect_language() directly
const langTokenId = await model.detect_language(features, generation_config);

Changes

• packages/transformers/src/models/whisper/modeling_whisper.js — 1 file
  • Add detect_language() method (~40 lines)
  • Call it from generate() when language is not specified (~10 lines)
  • Remove // TODO: Implement language detection comment

No breaking changes — when language is explicitly provided, behavior is identical to before.

Checklist

• Build passes (pnpm build)
• TypeScript type check passes
• Prettier formatting passes (pnpm format:check)
• No breaking changes to existing functionality
• Tests — happy to add if desired