Short chunk overlaps often contain only one participant, so returning
speakers receive new transcript IDs. MAI can also split one voice into
several IDs within a single request, which overlap stitching cannot fix.
Add an opt-in audio comparison pass: --speakers N selects Gemini 3.8 Flash
with a known count, and --speaker-model allows another compatible model.
Keep request-local identities separate, select up to three clean excerpts
per ID, and compare all excerpts in one OpenRouter request. Gemini uses
Google AI Studio because the Vertex backend limits prompts to one audio
file. No local model or extra Python dependency is required.
Only reconcile identities whose sample assignments agree. Preserve
uncertain identities separately, reject incomplete or malformed replies,
and retain the original words and timestamps. Bound sample count and
request size, clip sampling to the real audio duration, and extend worker
cleanup and buffered output to the comparison step. Document the sampled
method's limitations and additional provider request.
Test Plan:
- python3 -B -m unittest -v: all 82 tests passed, including real MP3
encoding with fake transcription/comparison responses and cancellation.
- bash -n transcribe: passed.
- git diff --cached --check: passed.
- Live API and voice accuracy unverified: this workspace could not access
the configured 1Password credential. Mocked responses prove pipeline
behavior, not model accuracy on the supplied recording.
Render transcript-wide speaker numbers instead of exposing upload chunks in
reader-facing headers. Reconcile adjacent requests using at least two
unambiguous word-and-time matches in their shared audio, including when
provider speaker IDs change between requests. Reject conflicting mappings
and never treat matching numeric provider IDs as identity evidence.
Assign fresh IDs where continuity cannot be established and report that
uncertainty on stderr. Number only speakers retained after overlap trimming.
Document that overlap matching cannot identify a returning speaker who was
absent from the overlap, and add the reader-facing boundary to AGENTS.md.
Test Plan:
- python3 -B -m unittest -v: all 46 tests passed, including swapped IDs,
ambiguous evidence, continuity across three requests, and plain headers.
- git diff --cached --check: passed.
- Tests use synthetic speaker labels; live diarization accuracy was not
verified because provider credentials are unavailable in this environment.
Sending entire recordings in one request can exceed provider limits.
Decode audio once to a shared PCM sample clock and split uploads at a
configurable maximum of 1 to 30 minutes, defaulting to 30 minutes.
Prefer nearby pauses and include six seconds of overlap to protect words
at audio cuts while keeping each encoded MP3 within the duration cap.
Request word timestamps for chunked text and SRT output. Match overlapping
words and times to retain each passage once, warn when a join falls back
to timestamps, and generate SRT cues on the original recording timeline.
Reuse subtitle timestamp validation without reordering regressing words.
Only emit output after every request and conversion succeeds.
Document the Python dependency, temporary storage, boundary limitations,
and lack of automatic retry or resume. Keep single-request text formatting.
Test Plan:
- python3 -B -m unittest -v: all 29 tests passed, including mocked provider
failures, native MP3 payload checks, boundary stitching, and SRT timing.
- bash -n transcribe and shellcheck transcribe: passed.
- git diff --cached --check: passed.
- Locally decoded and encoded the supplied 74-minute WebM into three MP3
chunks; ffprobe confirmed each duration was below 30 minutes.
- Live transcription was not verified: the configured 1Password credential
was unavailable.
ddidderr