Skip to content

docs(stt): record why transcript cuts are not snapped to quiet points - #1057

Merged
EtienneLescot merged 1 commit into
mainfrom
feat/1023-transcription-vocab-snap
Oct 7, 2026
Merged

EtienneLescot merged 1 commit into
mainfrom
feat/1023-transcription-vocab-snap

Conversation

@EtienneLescot

@EtienneLescot EtienneLescot commented Oct 7, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Measured the quiet-point cut snap suggested in #1023 on tools/stt-eval/word-timing (TTS corpus of 48 clips, clean and noisy, CTC aligner):

  • Snapping unconditionally to the quietest 10 ms lands inside words (a fricative or a stop closure is quieter than the boundary): clean one-word deletes go from 49% to 48/41/33% at ±10/20/30 ms.
  • Gated to near-silence (30 dB under the boundary), it moves 48 of 6,509 boundaries and changes no figure.
  • The only gain (+2 points), a cut in the middle of a short gap, is what cut-breath.ts already does.

The numbers now sit in transcription-and-captions.md, with the other ideas tried and left out. The vocabulary prompt (part 1) needs the whisper-stt helper to read a prompt field and set initial_prompt, so it is left for a follow-up with a helper rebuild.

Related issue

Part of #1023

Type of change

  • Documentation

Release impact

  • No release note needed

Desktop impact

  • Not platform-specific

Testing

  • Docs only. Measured with the installed v2.0.0 helper (Vulkan) and the app's cached CTC aligners; not on real speech (LibriSpeech is not on disk).

🤖 Generated with Claude Code

Summary by CodeRabbit

  • Documentation
    • Added measured results for word-boundary snapping at three time ranges, compared with the existing cut-breath behavior and an earlier RMS-based approach.
    • Documented that restricting snapping to near-silence affected 48 of 6,509 boundaries and did not change the measured results.

…#1023)

Measured on the word-timing harness with the CTC aligner: an energy-minimum
snap of the cut point lands inside words and costs up to 16 points of clean
one-word deletes; gated to near-silence it moves under 1% of boundaries and
changes nothing. The gain a short gap offers, cut-breath.ts already takes.
@coderabbitai

coderabbitai Bot commented Oct 7, 2026 •

Copy link
Copy Markdown
Contributor

Review in Change Stack →

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info
⚙️ Run configuration
  • Configuration used: defaults
  • Review profile: CHILL
  • Plan: Advanced
  • Run ID: d439e55c-a021-4b08-abc7-9de047d21076
📥 Commits

Reviewing files that changed from the base of the PR and between 6e27423 and 1a3920f.

📒 Files selected for processing (1)
  • technical-documentation/architecture/transcription-and-captions.md

Included review availability: This review used your included allowance. Your plan provides up to 8 included reviews per hour; 4 remain after this review.


📝 Walkthrough

Walkthrough

The architecture documentation adds measurements from evaluating cut snapping between words with gaps under 100 ms. It compares quietest-point snapping windows with existing cut-breath behavior and records results for near-silence and RMS snapping.

Changes

Cut-Snapping Evaluation

Layer / File(s) Summary
Document cut-snapping results
technical-documentation/architecture/transcription-and-captions.md
The documentation reports clean one-word delete rates of 48%, 41%, and 33% for quietest-point snapping within ±10, ±20, and ±30 ms, compared with 51% for existing cut-breath behavior. It also records that near-silence snapping changed no results and that the old RMS snap had similar results.

Priority: ⬇️ Low

Estimated code review effort: 1 (Trivial) | ~4 minutes

Change: Other

Suggested reviewers: vitaligusatinsky

Merge Risk: ⚪ Minimal · up to 1a392

The documentation contains an inaccurate pointer to the RMS-snap discussion, but this is a minor issue and does not materially affect merge readiness.

🚥 Pre-merge checks | ✅ 5
✅ Passed checks (5 passed)
Check name Status Explanation
Title check ✅ Passed The title clearly describes the main change: documenting why transcript cuts are not snapped to quiet points.
Description check ✅ Passed The description covers the change, related issue, documentation type, release and platform impact, and testing. It also explains the measurement limits and deferred follow-up. The screenshots section …
Docstring Coverage ✅ Passed No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 0…
Linked Issues check ✅ Passed Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check ✅ Passed Check skipped because no linked issues were found for this pull request.
✨ Finishing Touches
🧪 Generate unit tests (beta)
  • Commit to this branch
  • Create a new PR
  • Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

@EtienneLescot
EtienneLescot merged commit 9317871 into main Oct 7, 2026
19 checks passed
@EtienneLescot
EtienneLescot deleted the feat/1023-transcription-vocab-snap branch October 7, 2026 12:03
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant