Whisper's alignment makes words contiguous and places a start that
follows a pause late, so the previous word's span held the pause and
the next word's beginning; vad::tighten_words only trimmed silence at
the edges, which were voiced. A run of 200 ms or more of silence with
voice on both sides inside a word's span now ends the word, and the
next word, when contiguous, starts after it. When the voice before the
pause is too short to be the word (under half of 50 ms per letter) the
word starts after the pause, and a contiguous previous word takes the
scrap only when no pause separates its voice from it. Several pauses in
one span cascade through the following words.
On a 60 s probe (whisper-small) the 4 pauses that sat inside words are
all between words afterwards and no word keeps silence at an edge; on
a 48-minute interview the pauses inside a word's span go from 336 of
530 to 0, words longer than 1.5 s from 103 to 28, and the word after a
pause starts within 50 ms of the voiced onset 488 times instead of 192.
AI assistance: Mystic Intelligence for MysticArt Pictures.
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Move the test modules in whisper/tokenizer.rs and export/pipeline.rs to the
end of their files. Group the output path and optional open file in proxies.rs
into a Dest struct so transcode_with_file takes 7 parameters. Behaviour is unchanged.
- filmcraft-tts (feature kokoro): Kokoro-82M on candle, implemented from
the StyleTTS 2 / iSTFTNet / ALBERT papers, config.json and the
checkpoint's tensor shapes; official .pth read with candle's restricted
pickle reader; voice packs from their zip. Neural Voice: tts-text
phonemize -> Kokoro per chunk -> pauses -> pitch via audio-dsp.
- catalog: 9 US English voices (model card grade C+ or better) + CMUdict,
every file pinned to a revision and SHA-256 (336 MB).
- speech: generic download_files (same verified downloader).
- engine: tts.render (background synthesis into Session::tts_cache),
tts.downloadVoices (background job, cancellable), voices list with
package state; feature neural-voices (default on in the desktop app).
- ui: natural/basic voice groups, consent box (size, sources, licences),
download progress + Cancel, synthesis never blocks the window;
ui.set {tts:{...}} for agents.
Measured: Whisper-base WER on 50 original sentences 2.1-3.6 % for all 9
voices (built-in robotic voice 96.6 %); 2.2-2.3x real time on CPU.
Gates: fmt, clippy, layers, assets, wasm pass; tests pass except
shortcuts_tests::assign_reassigns_conflicts_and_undoes (fails on main)
and the frame pool test (flaky under load, passes alone).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
aac, bitstream, cfb, color, format, frame, geom, isobmff, media, scopes, speech, the desktop
app and the web app have no production unwrap/expect/panic/unreachable (the web app's two
`expect`s — frame-cache eviction and chunked file reads — become `let … else`); each now
carries `#![deny(clippy::unwrap_used, …)]` outside tests so it stays that way.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rust/clippy 1.99 added `chunks_exact_to_as_chunks` and tightened
`needless_late_init` and `drain_collect`, failing `clippy -D warnings`
on main.
- `chunks_exact(N)` / `chunks_exact_mut(N)` with a constant N become
`as_chunks::<N>()` / `as_chunks_mut::<N>()` (fixed-size arrays, so
several byte reads drop their per-element indexing). relink's
fingerprint uses the `(chunks, remainder)` split; its hash output is
unchanged.
- Late-initialised `let`s in ac3 bit allocation and opus CELT
bands/rate become `let x = if/match …`.
- `dpb.drain(..).collect()` becomes `std::mem::take(&mut dpb)`.
No behaviour change; `cargo xtask ci` passes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- crates/speech (L2): Transcriber trait, FixedTranscriber, Whisper model catalogue (pinned,
SHA-256 verified) with optional downloader (feature `download`: ureq + rustls + RustCrypto),
pure-Rust Whisper inference on candle with word timestamps (feature `whisper`), log-mel, VAD
word tightening, MFCC speaker clustering. Both features off by default and for wasm.
- project: per-media-item transcripts (Project::transcripts); schema v9 with a no-op v8->v9
migration and a v8 fixture.
- edit::transcript: sequence transcript derived from clip transcripts, search, paragraphs,
word ranges, pause/filler detection, ripple delete of many ranges, caption layout.
- engine: transcript.* commands (generate, set, delete, inspect, search, select, extract, lift,
renameSpeaker, removeFillers, removePauses, createCaptions, models, downloadModel);
Session::transcriber for hosts/tests; features `whisper` / `speech-download` (also on the
desktop and CLI apps). Without whisper, transcript.generate fails with a clear error.
- ui: Text panel Transcript tab (Transcribe button, speaker paragraphs of clickable words,
Shift+click selection -> In/Out, extract/lift/remove fillers/create captions, search) with
text.transcript.* automation ids; headless UI test.
- docs/transcripts.md, crates/speech/README.md, architecture and project-files updates.
whisper-tiny measures 12.2% WER over 797 words on local samples.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
