mirror of
https://github.com/storytold/filmcraft.git
synced 2026-11-17 04:14:24 +00:00
* 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
filmcraft-speech
Speech-to-text for FilmCraft's Text panel ▸ Transcript (L2). User-facing behaviour, commands and model handling: docs/transcripts.md.
Transcribertrait (mono 16 kHz in, word-timedfilmcraft_project::Transcriptout) andFixedTranscriberfor tests and agents.models: Whisper model catalogue (pinned revisions, SHA-256, sizes, licences) and, with featuredownload, the verified downloader (ureq + rustls + RustCrypto). Weights are never bundled.whisper(featurewhisper): 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.