Let people retry "No speech found" meetings, and treat mis-taps as cancels

[claude]

Requested by Justin · project thread

Why

Before: a meeting judged "No speech found" became a failed row with no way to retry it, even though its audio was kept. Real speech could land there too: one "no speech" verdict from the diarizer ended the whole meeting, and a quiet track never got a second look. Stopping a meeting a few seconds after starting it by accident showed an error and left a failed row to clean up. A sub-second dictation mis-tap showed "Recording ended too soon".

After: a "No speech found" meeting keeps a Try again button while its audio exists. Before giving up on a meeting (or a split-mode mic) with no words, the pipeline takes one more careful look for quiet speech. Normal meetings with words are processed exactly as before. A meeting stopped within a few seconds that holds only silence closes like a cancel: no error, no "Saved" flash, no failed row. Anything that might hold a word, or any sign capture broke, still ends as a visible, retryable failure. A quick dictation tap closes like a cancel too.

This fixes customer pain points #5 and #10. PostHog, CI builds excluded, last 30 days: 66 no-speech skips on 31 devices (about 6% of stops), 42 of them under 10 seconds, and 0 retries because retry wasn't offered. Plus 196 dictation too-short events on 40 devices. Write-up: /mnt/project-files/customers/no-speech-analysis.md.

Product Impact

• Affects: meetings / dictation
• Lane: meeting reliability / dictation reliability / activation (12 of the no-speech skips were first meetings, 11 of those 2-15 second test taps)
• Why this matters: a person who really had a conversation can recover it, and a mis-tap no longer looks like the app failed.

What changed

• Retry: PipelineErrorKind.offersRetryForSavedAudio makes noSpeechDetected retryable (recordingTooShort still isn't). The copy now says "If people were talking, open Home and choose Try again." Import and saved-meeting retranscription no-speech failures keep their own copy, since they leave no Home row to retry from.
• Diarizer "no speech" is not final: on system audio with no mic, and on the split-mode mic (both the two-track and mic-only paths), it now becomes an empty channel instead of ending the meeting.
• Last-chance pass (lastChanceSpeechSweep):
  • When should it run?
    • On the call side, only when the whole meeting came out empty.
    • On a split-mode mic, whenever it produced no words, even if the call side did. If nothing is recovered, the mic is marked unusable as before.
  • How does it work?
    • It silence-splits the track, pads short pieces, and uses click-tolerant gain (capped at 3x for call audio, 12x for the mic).
    • It only runs on tracks that rise and fall like speech (hasSpeechLikeModulation), so silence, hum and DC never reach STT.
  • What gets dropped? A channel that only yields fillers ("Mm.") or a single word.
  • How are recovered words labeled? They stay unnamed ("Speaker 1" / "You"), and diarizer contexts and new profile ids are cleared.
• Normal meetings are unchanged: the click-tolerant gain (normalizeForSpeech(ignoringSpikes:)) and short-segment padding only run inside the last-chance pass, until a corpus A/B.
• Accidental starts (meetings): TranscriptionTaskManager.startTranscription takes the controller's sessionLength (Record to Stop).
  • Files under 2 s count as a tap only when the session is known and under 4 s and capture health is clean. Otherwise it's a visible "Recording too short" failure; when the session ran longer, the copy says capture stopped early instead of "Nothing broke".
  • A no-speech result is discarded only when all of these hold:
    • files and session are both under 10 s;
    • health is clean (no gaps, switches, missing or unusable track, or degraded capture);
    • no track has speech-like signal.
  • Otherwise it's the normal retained, retryable failure.
  • A discard deletes scratch and the journal and publishes DisplayStatus.discardedAccidentalStart.
  • Imports and retries never take this path.
• App side: TerminalTranscriptionOutcome.discarded settles to ready.
  • There's no "Saved" flash.
  • The cancel sound plays only when no other meeting is recording.
  • Telemetry: meeting_transcript_skipped with failure_kind=accidental_start and friction cancelled, plus the runtime-diagnostics terminal event.
• Dictation: recordingTooShort with a press under 1.5 s closes like a cancel, with friction cancelled. A longer press that captured too little still shows an error, now "Only a moment of audio came through. Try again, and if it keeps happening, check your microphone." instead of telling the user to speak longer. dictationFallback no longer calls a failure on 1 s+ of audio "too short", so that audio is kept as a model failure.

How I checked it

• scripts/dev/agent-preflight.sh
• python3 scripts/dev/check-build-source-lists.py, python3 scripts/dev/check-analytics-emitters.py
• bash build.sh --no-open, bash run-tests.sh, bash build-deps.sh --force, bash run-integration-smoke.sh, swift test: the session has no Swift toolchain, so CI runs these.
• Manual check: Mac test at /mnt/project-files/hardware-tests/no-speech-retry.md

New tests:

• spike-tolerant peak, and the default gain unchanged
• speech-like modulation
• the sweep recovering call-side speech
• the sweep recovering split-mode mic speech with and without call-side words
• a split mic with nothing recoverable marked unusable
• filler-only recovery dropped
• steady tracks still ending as no speech
• short silent sessions discarded
• short files from a long session kept and retryable
• no session length means audio is kept
• sub-2 s files count as a tap only for a known, short, healthy session (8 s session, unknown session and broken-track cases stay visible failures)
• the "capture stopped early" meeting copy
• isAccidentalStart edges
• unreadable tracks count as possible speech
• the dictation press-length rule and the long-audio fallback

Updated tests:

• the 2 second gate now expects .discardedAccidentalStart for a 1.4 s session
• the dictation too-short copy
• the split-speakers failed-queue test now uses a real STT error instead of silence

Risk Review

• Privacy / local-first behavior reviewed: accidental_start is a plain category and no new properties were added.
• Storage path or migration impact reviewed: legacy failed rows with "No speech detected" now decode as retryable. Discarded accidental starts delete only their own scratch audio and journal.
• Public-facing copy stays concrete
• Release/update impact: none
• Agent PR, draft until human review
• UI evidence: none. The UI change is that an error no longer shows up.
• No private data included

Notes

• Overlaps with other open PRs:
  • Mic-only meetings don't start the call-audio tap #1787 (mic-only meetings) touches TranscriptionTaskManager (healthInfo/micOnlyByChoice params) and the MeetingSessionController stop path. It also writes a silent system track, which the last-chance pass skips because silence has no modulation.
  • Show a "Mic only" note when a meeting records just the mic #1794 touches the overlay files and MeetingSessionController.
  • Keep system audio when Stop lands during system-audio setup #1777 touches only Audio.swift / AudioFileManager, so no overlap.
  • Sources/Meeting/CLAUDE.md: my retry-invariant line sits right next to the line Mic-only meetings don't start the call-audio tap #1787/Show a "Mic only" note when a meeting records just the mic #1794 edit. Expect a trivial doc merge conflict.
• Deep review (/mnt/project-files/reviews/next-release/1801.md): S1-S5 are addressed in ab1cc55. For S5, the gain and padding changes were split out of normal meetings rather than flagged. Re-check N1-N3 are addressed in 6a996c4 (N3: the copy test reads the messages from PipelineFailureDisplayCopy, so a wording change fails the test).
• Tradeoff: a quick dictation tap (under 1.5 s) that did hold a word used to show "speak for at least a second". Now it closes like a cancel.
• Telemetry continuity: sub-2 s meetings from a short, healthy session move from recording_too_short / give_up to accidental_start / cancelled. Dictation mis-taps keep their event name, but friction flips to cancelled. PostHog insights that count these need a note.
• Not fixed here: meeting_capture_stopped_under_controller always reports duration_bucket=lt_10s (MeetingSessionController stop telemetry), which skews PostHog duration numbers.

Agent handoff

COORD_DONE: BRIEF | this PR | retry for no-speech + last-chance speech pass + accidental starts as cancels | none | none | preflight, source-list and analytics checks (Swift checks via CI) | drive CI green, run the Mac test

🤖 Generated with Claude Code

https://claude.ai/code/session_01X89qzDDfTibdTyHR74c5C1