[bryangruneberg]

(Keeps Whisper as an optional extra)

Adds a second transcription backend next to local Whisper, selected with transcriber: whisper | assemblyai (default unchanged: whisper).

AssemblyAITranscriber has the same interface as WhisperTranscriber and adds:

• speaker diarisation: segments carry a speaker, the transcript renders as [mm:ss] Speaker A: ..., and the summariser is given "Speaker A: ..." lines (summary_input) so action items can name an owner;
• upload as 16 kHz mono FLAC via ffmpeg. The recorder's 48 kHz WAV is ~11 MB/min, and a 34-minute meeting (377 MB) had its upload cut off mid-TLS; FLAC is lossless for what the model hears and ~10x smaller;
• upload retries on a dropped connection;
• submit()/wait() split so a caller can keep the transcript id and resume polling instead of re-uploading; AssemblyAIError carries status_code / transient / remote_failed for callers deciding whether to retry;
• a missing API key is reported at transcription time, so the app still opens and records without one.

build_transcriber(config) replaces the two duplicated construction blocks in app.py (startup and settings-saved). Settings gains a transcription picker.

Packaging: openai-whisper moves from core dependencies to a whisper extra (torch is several GB and not always available for the newest Python). all still includes it, so setup.sh's pip install -e .[all] installs exactly what it did before; a new assemblyai extra is the cloud-only set.

CI installs requests and runs tests/test_assemblyai.py.

[jamespember]

This is really strong work @bryangruneberg. Reading it end-to-end there's a lot of care in the parts that matter: uploading, retries, the split between submit/wait, and the packaging trade-offs.

What I like:

• Speaker labels flowing all the way through. TranscriptSegment.speaker is optional (Whisper leaves it None and renders unchanged), speaker_text() produces the Speaker A: … form that goes to the summariser, and format_segments renders **[mm:ss] Speaker A:** … in the saved transcript. Clean data flow, no leaky enums.
• The FLAC compression choice. 16 kHz mono FLAC really is what AssemblyAI listens to internally, so this is lossless for recognition and ~10× smaller — the 377 MB → 37 MB range makes the difference between "upload finishes" and "TLS cuts out at 198 MB" over a flaky connection. Falling back to the original recording when ffmpeg isn't there is the right graceful-degrade.
• AssemblyAIError carrying status_code / transient / remote_failed. Lets callers make sensible retry decisions instead of matching on error strings.
• submit() / wait() split. Genuinely useful — a callback restart or later poll can resume rather than re-uploading a multi-hundred-MB meeting. Not used from app.py today, but the interface is right for when it is.
• Deferring ValueError for missing key to transcribe(), not __init__. So the app opens and records without a key configured, and the failure is loud when it matters.
• build_transcriber(config) — kills the duplicated construction block that was in both __init__ and handle_settings_closed. Long-overdue tidy.

Packaging change is a breaking change for existing installs, but handled well:

Moving openai-whisper from core deps to a whisper extra is the right call — torch is several GB and often lags on the newest Python — and setup.sh uses pip install -e .[all] which still pulls it in. Anyone who ran pip install -e . by hand without an extra will need to re-run with .[whisper] or .[all]. Worth a line in the README changelog when we next cut a release; not blocking.

Tests are excellent. Injected FakeSession for the transcriber, FakeResponse shaped like requests, no network. Coverage includes:

• upload → transcript request → poll cycle (test_transcribe_uploads_requests_speakers_and_polls)
• speaker text formatting
• API-side error surfacing
• HTTP error status included in message
• build_transcriber routing
• Whisper's speaker-less rendering unchanged
• FLAC upload path (with @pytest.mark.skipif when ffmpeg is missing — nice)
• upload retry on SSLError
• upload give-up after 3 attempts
• submit → wait split
• redaction of assemblyai_api_key in to_safe_dict()

Minor nits (not blocking):

• AssemblyAIError.transient and .remote_failed are set but not read anywhere yet. Fine — they're part of the API the split enables.
• _check uses response.text[:200] for the fallback when JSON parsing fails; if the server ever streams a huge HTML error page that's a nice cap.

Rebased locally on top of #18 which just merged — clean rebase, 147 tests pass, ruff clean. Merging.

[jamespember]

Merged as commit 9aa486e on main (squashed locally and pushed because my gh token doesn't carry the workflow scope, so the automated gh pr merge refused to touch .github/workflows/ci.yml). GitHub sees the branch as closed rather than merged for the same reason, but the change is live on main. Thanks again @bryangruneberg — this landed with full commit attribution via Co-authored-by:.

[bryangruneberg]

Woop Woop! Glad Assembly made it in 😁