Repository navigation
Releases
Releases · anatolykoptev/ox-say
v0.1.41
v0.1.40
0.1.40 (2026-11-19)
Added
- app: a setting to hide OxSay overlays from screen sharing and recordings (#659) (8b7a3c0)
- daemon: serve the notes model over an OpenAI-compatible endpoint (#537) (#643) (de46dfd)
- glossary: a recurring misrecognition in dictation counts as a sighting (#645) (d046158)
- mcp: meeting_live, the live transcript of the call being recorded (#657) (291dea0)
Documentation
v0.1.39
0.1.39 (2026-11-18)
Added
- app: a setting to hide the dictation pill (#649) (eae19e1)
- stt: Cyrillic and mixed-script spoken forms in the code packs (#651) (8dd830d)
- stt: default to the parakeet-ultra q8_0 weights (#655) (7af8e8c)
Speech recognition accuracy
The new default weights, parakeet-ultra (Moondream's post-trained parakeet-tdt-0.6b-v3, CC-BY-4.0, same size and speed), make fewer errors than v3 on every set we measured. Word error rate, same engine, lower is better:
| Set | v3 q8_0 (before) | ultra q8_0 (default now) | ultra f16 (optional) |
|---|---|---|---|
| Russian-accented English (proxy) | 18.85% | 16.53% | 15.67% |
| Native English | 5.03% | 4.89% | unchanged |
| Russian | 1.46% | 0.97% | unchanged |
- English words that came out transliterated into Cyrillic: 8 → 3 clips.
- f16 vs q8_0 on accented English: −0.86 points, paired bootstrap 95% interval −1.57 to −0.26. f16 is 1.26 GB instead of 669 MB and about twice as slow per clip; it is opt-in (see the f16 weights).
- How it was measured: synthetic speech, 125 English sentences read by two macOS voices plus 20 Russian sentences, on the same engine build. Not yet measured on human recordings.
- The converted weights and their provenance: Koptev/parakeet-ultra-ggml.
v0.1.38
v0.1.37
v0.1.36
v0.1.35
v0.1.34
0.1.34 (2026-11-16)
Added
- clean: clean up the voice in a file, from the Расшифровка page (#616) (3202e62)
- glossary: activation per meant term, Russian case matching, threshold by length (#588) (815862b)
- meetings: live captions while a meeting records (#597) (b6cc33c)
- meetings: per-app call tap for verified apps only, with diagnostics and an outside-playback notice (#608) (c82d5f1)
- meetings: sync the recording to disk every 2 s and hold the Mac awake while recording (#607) (cc7bebb)
- notes: report the GPU offload the notes child states in /status (#590) (5b47546)
- voices: report when the ref_text check could not run (#589) (5a732e6), closes #440
- window: live captions for the meeting being recorded (#623) (0457f87)
Fixed
- clean: verify.py's contract checks survive python -O (#591) (a5bf7d2), closes #447
- dictation: refuse the dictation start while a voice take records (#602) (f3e23c9)
- meetings: keep the first-time notes folder out of iCloud and meeting audio out of backups (#605) (960c85e)
- shots: compile the harness again, build it in CI, run render.sh without GNU timeout (#587) (0c1f115)
- shots: fail loudly when the display is asleep or a capture fails (#601) (24f57e5)
- speak: a changed voice stales the page; a failed clip is told and cuts nothing (#606) (610da44)
- speak: a notification target survives a daemon that does not answer (#592) (73ceeaf), closes #513
- stt: harden the term-boost wire format; log silent cuts; mutation evidence for #576 (#596) (def5090)
- updater: show why an install was refused; skip releases above this Mac's macOS (#610) (a88a6f3)
Changed
Documentation
v0.1.33
0.1.33 (2026-11-15)
Added
- glossary page and owner edit/purge (#575) (926ba1f)
- stt: wire spoken-form decoder term boosting into the daemon (Stage B, supersedes #574) (#576) (5f10527)
Fixed
- daemon: refusal quotes the real encoder window, not the test seam (#565) (11d88dd), closes #143
- tech-debt batch A (#561) (fbd63d0)
- tech-debt batch B (#563) (5ba319b)
- tech-debt batch D (#567) (0c84b5e)
- tech-debt batch E (#562) (bfe8c16)
- tech-debt batch F (#566) (d3d57b5)
- tech-debt batch G (#578) (46582d2)
- tech-debt batch H (#580) (3a4b0b9)
- tech-debt batch J (#581) (3a42788)
- tech-debt batch K (#569) (9c12c1f)
- tech-debt batch L (#577) (9669a27)