Pack short meeting speech segments into one speech-to-text pass

[r3dbars]

Why

Before: after Stop, a meeting made one Parakeet call per speech segment. Parakeet pads every call to a fixed 15 s window, so a typical 4 s segment cost the same work as 15 s of audio. Nothing kept the Mac awake while the job ran, so App Nap or idle sleep could stretch long jobs.

After: short segments are laid end to end (0.5 s silence between them) into one 15 s pass, and the text is split back per segment by token start times. A meeting needs roughly a third of the calls. The Mac stays awake (no App Nap, no idle sleep) while a transcription job runs.

Product Impact

• Affects: meetings
• Lane: meeting reliability
• Why this matters: real-user P50 for 30 min+ meetings is ~66 s after Stop, with a long tail. The speech-to-text stage is most of that.

What changed

• SpeechSegmentPacking (TranscriptedCore): layout of packed audio, word-level split by first-token start against mid-gap boundaries, and an in-order batcher. Segments longer than the window still go alone.
• SpeechToTextEngine gains optional packedSegmentWindowSamples / transcribePackedSegments (default nil, so other engines are unchanged).
• TranscriptionPipeline: system-audio loop and default mic loop go through the batcher. A packed call that fails or returns the wrong count falls back to one call per segment; cancellation still rethrows.
• ParakeetEngine.transcribePackedSamplesWithTokenTimes + STTRouter wiring: Parakeet TDT v2/v3 only, auto language only. Whisper and explicit-language paths are untouched.
• MeetingSTTAdapter: packing path, custom dictionary per segment, timings still recorded. TRANSCRIPTED_MEETING_STT_PACKING=0 turns it off.
• TranscriptionJobActivity.keepingMacAwake wraps live, import and retry pipeline runs in a .userInitiated activity.
• CLI import-audio: same packing switch plus a timing line ("N speech-to-text calls taking Y s (packing on/off)") for the before/after run.
• Tests: SpeechSegmentPackingTests (layout, split, batcher, zero window, fallback).

How I checked it

• scripts/dev/agent-preflight.sh
• bash scripts/dev/linux-checks.sh (runs without Swift) — 48 passed
• Selected checks from .agents/test-matrix.yml for the files changed (CI)
• bash build.sh --no-open (CI)
• bash run-tests.sh (CI)
• Performance budget passed
• bash run-integration-smoke.sh (CI)
• swift test (CI)
• swift test --package-path Tools/TranscriptedCLI (CI)
• Manual check: before/after on a Mac

Checks I could not run, and why:

• No Swift toolchain in the cloud session, so build and tests rely on CI.

Mac or hardware test still needed? Yes:

1. TRANSCRIPTED_MEETING_STT_PACKING=0 transcripted import-audio <30-60 min recording> --output-dir <tmp> then the same without the variable. The timing line should show about a third of the calls and a shorter speech-to-text time with packing on. On main there is no packing, so the call count matches the "off" run.
2. Diff the two transcripts: words should match apart from segment-edge differences.

Risk Review

• Privacy / local-first behavior reviewed (on-device only, no new off-device data)
• New analytics properties or Sentry tags avoid the sanitizer's drop fragments (none added; one local-only event meeting_packed_segments_transcribed)
• Checked the text-pin tests for every file I edited (check-source-pins.py --changed-only PASS)
• Storage path or migration impact reviewed (none)
• Public-facing copy stays concrete and matches current product scope (no copy)
• Release/update impact reviewed (none)
• Agent PRs got an independent deep review of the full diff
• UI changes include sanitized .agent-review/visuals/ evidence (no UI)
• No private transcripts, audio, tokens, personal paths, or customer data are included

Notes

Follows #1861 (stage timings on meeting_transcript_saved), which will show the effect in real meetings via stt_calls, stt_ms and sleep_ms. Main risk: a word near a segment edge landing in the neighbouring segment.

Agent handoff

COORD_DONE: BRIEF | this PR | packing + keep-awake + CLI timing | none | decisions: Mac before/after run needs Justin's OK | linux-checks, source pins | CI green + deep review

🤖 Generated with Claude Code

https://claude.ai/code/session_01LcH2PBaqN6yYkxeoaYrj45

---

Generated by Claude Code

Mac results (2026-09-28, Apple M5 Max): passed

Built with this branch combined with main plus #1791 and #1788. Checks passed: build-deps --force, build.sh, run-tests.sh, swift test (incl. SpeechSegmentPackingTests), run-integration-smoke.sh, swift test --package-path Tools/TranscriptedCLI, and the CLI manifest and packaging tests.

transcripted-cli import-audio (--no-speaker-identification), packing off vs on:

Recording	STT calls	STT time	Transcribe total
43 min, 5 speakers	251 → 156	20.00 → 15.63 s	24.55 → 19.72 s
27-min lecture with human captions	544 → 174	33.06 → 14.73 s	38.17 → 19.83 s

Accuracy:

• 43 min: same 106 utterances and 5 speakers. 1 of 9,727 matched words moved speaker. 53 words (0.54%) differ, in 29 spots of 6 words or fewer.
• Lecture vs its human captions (jiwer, same normalizer on both sides): WER 24.45% → 22.79%. Fewer substitutions (409 → 368), deletions (840 → 783) and insertions (461 → 443).

Packing is faster and no less accurate here.