Repository navigation
SONOLYTICA 0.2.2 — real-media trust fixes
SONOLYTICA 0.2.2
Trust fixes from real-media dogfood: no ghost music, no fallback noise,
clearer evaluation commands, language-aware ASR choice.
No schema change, no new model, no benchmark rewrite.
Fixed
- Ghost music on speech. The DSP now gates musical evidence before the
document is built. BPM, key, chords, beats and sections arenull/empty when
there is no evidence; loudness remains valid. Clean speech (the 11-second
JFK fixture) reportsmusic not detected; a click track and a sustained
triad still pass. Query JSON exposespresent: false/true. - Meaningful tempo confidence. Tempo confidence is now a normalized
cross-correlation, comparable across sparse clicks and dense speech. The old
raw ratio could reportscore 0.000while still publishing a BPM. - Fallback warning noise. A missing optional fallback (pyannote when
Sortformer is present) no longer emitsprovider_unavailableon every run.
If no provider covers a requested capability, one warning still fires;
sonolytica doctorstill lists every unavailable model. - Evaluation CLI.
sonolytica-eval report DOCUMENTnow works positionally;
--documentremains a deprecated alias.weraccepts
--reference-text/--hypothesis-textfor raw strings and-for stdin,
not only file paths. - Language-aware ASR choice. English uses
whisper-base.enby default;
non-English requests preferwhisper-smallwhen installed; an explicit
--model(includingauto) wins.sonolytica doctorprints the default.
Default model
whisper-base.en remains the fast English default. whisper-small is the
quality option and the automatic choice for explicit non-English languages.
This avoids silently changing the 26× realtime benchmark; switching the English
default requires a separate measured decision.
Upgrade
Re-run sonolytica analyze on documents produced by 0.2.1 to apply the music
gate. Existing documents are still readable: query treats low-confidence music
as present: false and hides BPM/key/chords.
Unchanged evidence
The hard/long benchmark, DER 6.40 %, 14 models, schema 1.6.0, MCP tools and
Windows/macOS/Linux packaging are unchanged. Tests: 184 model-free / 201
all-features. The honest ASR number remains
6.05 % WER / 2.25 % CER on LibriSpeech test-clean; 0.00 % on an 11-second
JFK clip is a sanity check, not a general accuracy claim.
Known limits
- Beatless music with weak or noisy triads may be under-reported by the
conservative gate. It will not hallucinate music on speech. whisper-smallis not the English default; use--model whisper-small.- Real overlapping speech still needs an overlap-aware diarizer.