Files
Dylan M. Taylor cefb980e6c speech: validate the Whisper attention head count (#4195)
* speech: validate the Whisper attention head count

The Whisper attention layer took its head count from config.json without
checking it: zero heads divided by zero in forward, and a huge count sized
every attention-logit tensor. A count must now be 1..=64 and divide
d_model.

* address review: separate head-count and divisibility errors
2026-11-03 23:19:06 -05:00
..

filmcraft-speech

Speech-to-text for FilmCraft's Text panel ▸ Transcript (L2). User-facing behaviour, commands and model handling: docs/transcripts.md.

  • Transcriber trait (mono 16 kHz in, word-timed filmcraft_project::Transcript out) and FixedTranscriber for tests and agents.
  • models: Whisper model catalogue (pinned revisions, SHA-256, sizes, licences) and, with feature download, the verified downloader (ureq + rustls + RustCrypto). Weights are never bundled.
  • whisper (feature whisper): Whisper inference on candle (CPU, pure Rust).
  • mel, vad, diarize: log-mel front end, word-bound tightening, MFCC-clustering speaker labels.

Both features are off by default and are never enabled for wasm (cargo xtask wasm checks the crate without them).

References

Implemented from the published description of the model: A. Radford et al., "Robust Speech Recognition via Large-Scale Weak Supervision" (OpenAI, 2022), the model cards and configuration files published with the weights (config.json, generation_config.json, tokenizer.json), and the candle tensor library's public API. Weights: OpenAI Whisper, MIT licence.