meeting-minutes: speaker diarization via siren's new endpoints

[Kabilan108]

Client half of the siren diarization rollout (M4/M5 + review hardening). Three commits touching only bin/meeting-minutes.

• Requests per-word timestamps on every chunk; preserves words in cached transcript parts.
• Diarizes the whole recording once (/v1/audio/diarize), aligns words to turns server-side (/v1/audio/align), and writes speakers + per-segment speaker into transcript.json additively.
• Prompt renders grouped [m:ss] SPEAKER_NN: turns (timestamps retained for the screenshot-candidate pass) plus an instruction to map labels to attendee names from content and turn-taking.
• Best-effort like the calendar pass: any diarize/align failure (including empty-turn responses) degrades to today's flat transcript; MEETING_MINUTES_DIARIZE=0 disables.
• Trailing sub-second ffmpeg segment slivers are dropped via decode-based duration (segment-muxer FLAC headers report cumulative end timestamps, not chunk length).

Verified end-to-end on a real recorded meeting against the siren branch: 211/211 segments speaker-tagged, prompt renders 102 timestamped speaker turns, flat fallback exercised. Requires the siren diarization-rollout branch server-side; degrades cleanly against older servers.

🤖 Generated with Claude Code

---

[greptile-apps]

Greptile Summary

This PR adds the client-side speaker diarization pipeline to meeting-minutes, calling two new Siren endpoints (/v1/audio/diarize, /v1/audio/align) and rendering the result as grouped [m:ss] SPEAKER_NN: turns in the Codex prompt. All diarization steps are best-effort: any failure degrades cleanly to the existing flat JSON transcript, and a MEETING_MINUTES_DIARIZE=0 env var disables the feature entirely.

• Per-word timestamps are now requested on every transcription chunk via timestamp_granularities[]=word; collect_transcript_words stitches them across chunks with index * CHUNK_SECONDS offsets and feeds them to the align endpoint.
• A new filter_degenerate_tail_chunks function decodes the last chunk with ffmpeg -f null to drop sub-second slivers created by the segment muxer before transcription; decode_duration (always decode-based) is correctly used here rather than probe_duration (header-based).
• probe_duration gains a second ffprobe strategy (stream=duration) and an ffmpeg decode fallback, though a structural gap means the fallback is unreachable if ffprobe exits non-zero rather than outputting N/A.

Confidence Score: 5/5

Safe to merge; diarization runs entirely in best-effort mode and every failure path degrades cleanly to the pre-existing flat transcript.

All three new pipeline stages (diarize, align, render) are wrapped in their own isolated try/except blocks that catch subprocess, OS, and application errors and set a failed status before continuing. The transcript.json is only overwritten after valid_alignment passes, and atomic_write_json ensures no partial writes reach the prompt. The one structural gap noted (probe_duration loop not catching a non-zero ffprobe exit before reaching the ffmpeg fallback) affects an uncommon code path and has no impact on the new diarization logic, which correctly uses decode_duration for the sub-second sliver check.

Files Needing Attention: bin/meeting-minutes — specifically probe_duration, which gained new fallback strategies but has an incomplete try/except structure in the loop.

Important Files Changed

Filename	Overview
bin/meeting-minutes	Adds end-to-end speaker diarization (diarize → align → speaker-tagged prompt rendering) with graceful fallback; one structural gap in the probe_duration loop prevents the ffmpeg decode fallback from being reached when ffprobe exits non-zero.

Sequence Diagram

sequenceDiagram
    participant J as process_job
    participant FC as filter_degenerate_tail_chunks
    participant DR as diarize_recording
    participant TC as transcribe_chunks
    participant AT as align_transcript
    participant S as Siren API

    J->>FC: chunks (sorted FLACs)
    FC->>FC: decode_duration(last chunk)
    alt "duration < 1s"
        FC-->>J: chunks minus trailing sliver
    else
        FC-->>J: chunks unchanged
    end

    J->>DR: diarize_recording(job)
    DR->>S: POST /v1/audio/diarize (whole-recording FLAC)
    S-->>DR: "{turns, speakers, duration}"
    DR-->>J: diarization payload

    J->>TC: transcribe_chunks(job, chunks)
    TC->>S: "POST /v1/audio/transcriptions (per chunk, +timestamp_granularities[]=word)"
    S-->>TC: verbose_json with word timestamps
    TC-->>J: transcript.json

    J->>AT: align_transcript(job, transcript_path, diarization)
    AT->>AT: collect_transcript_words(job, chunks)
    AT->>S: "POST /v1/audio/align ({words, turns})"
    S-->>AT: "{speakers, segments with speaker labels}"
    AT->>AT: overwrite transcript.json with speaker-tagged segments
    AT-->>J: transcript_path

    J->>J: "render_transcript_for_prompt -> [m:ss] SPEAKER_NN: ..."
    J->>J: generate_minutes (Codex prompt)

Loading

_{Reviews (2): Last reviewed commit: "Address Greptile review findings" | Re-trigger Greptile}

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 1 additional finding.