diarization#85
Open
cjpais wants to merge 9 commits into
Open
Conversation
- transcribe_diarize_mode tri-state on run_params (DEFAULT/OFF/ON), validated like pnc/itn; non-DEFAULT against a non-supporting model WARNs and proceeds (TRANSCRIBE_FEATURE_DIARIZATION probe) - transcribe_segment gains trailing speaker_id (1-based; 0 = none) - new transcribe_speaker_segment result rows + init/n/get accessors, batch mirrors, TRANSCRIBE_ABI_SPEAKER_SEGMENT - session plumbing: SegmentEntry.speaker_id, SpeakerSegmentEntry vector through capture/snapshot/restore/clear - api_smoke + run_dispatch_unit coverage for the new contracts NOTE: run_params layout changed (diarize inserted after itn) — 0.x ABI break; abihash/bindings regen lands in a follow-up commit this branch. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
New arch/moss/diarize.{h,cpp}: strict-grammar parser for the model's
[start][Sxx]text[end] output. Default (diarize DEFAULT/ON) now returns
per-turn segments with t0/t1 from the bracket timestamps, 1-based
speaker_id, speaker-segment rows, and dediarized full_text (the same
normalization the WER gates scored). diarize=OFF keeps the previous raw
single-segment passthrough byte-for-byte.
Unrecognized bracket spans degrade to verbatim text (no byte loss);
unknown turn boundaries are patched from neighbors / audio length.
capabilities: max_timestamp_kind NONE -> SEGMENT, FEATURE_DIARIZATION
on. run_batch fast path threads run params into finalize_utterance.
tests: moss_diarize_parser_unit (13 grammar cases incl. malformed
spans, bare tags, CJK, rounding, dediarize equivalence).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Correction to the exploration-phase assumption: -plus diarization is a
prompt-selected task (IBM's verbatim "Speaker attribution: ..."
instruction), not always-on. So for granite the family DEFAULT resolves
OFF (plain transcription, byte-identical to today; gates untouched) and
diarize=ON opts into the SAA task. Upstream defines timestamps and
speaker attribution as separate tasks; diarize=ON + timestamps=word is
rejected with INVALID_ARG, which removes the word-index remapping scope
entirely.
New arch/granite/diarize.{h,cpp}: [Speaker N] splitter (1-based per the
model card; optional ':' + space consumed; unrecognized spans pass
through verbatim). Turns become segments with speaker_id and speaker
rows with t0=t1=0 (the SAA task carries no timing).
model.cpp: FEATURE_DIARIZATION set from the GGUF
stt.capability.speaker_diarization KV (-plus only, mirrors the
word_timestamps lowering); run() and the run_batch capture loop now
share one finalize_granite_result().
tests: granite_diarize_parser_unit (model-card example, no-marker
passthrough, colon-less markers, Speaker 0 sentinel guard, byte-loss).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- flags mirror --pnc/--no-pnc; unset keeps family DEFAULT - segments JSON gains "speaker_id" when attributed; segments now also emit for speaker-attributed turns without timing (granite SAA, kind NONE); new "speakers" array from the speaker-segment accessors in single-file and batch JSONL (p omitted when NaN) - human output prefixes attributed segments with "S<N>: " and prints timing-less speaker turns without the timestamp column Smoke-tested live: moss default 2-speaker (31 turns, S1->S2 boundary correct), moss --no-diarize raw passthrough, granite-plus --diarize 2-speaker split, granite default unchanged, diarize+word-timestamps INVALID_ARG. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- generate.py: include/transcribe.abihash, python _generated.py, typescript _generated.ts - cargo xtask bindgen: rust transcribe_sys.rs - swift pinnedHeaderHash bumped after review (new: diarize_mode enum + run_params field, segment speaker_id, speaker-segment struct + accessors, ABI enum entry) Intentional 0.x ABI break: run_params layout changed (diarize inserted after itn). Both drift gates pass (swift_abihash_check.py, generate.py --check). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…n change Backs the granite decoder out to the fixed 256-token generation budget (pre-d24eeed state). The rework (chunked KV growth, decode-to-ceiling, batch rerun path) ships separately once it has a granite_grow_kv test and a repetition guard; saved as ~/sandboxes/transcribe/tmp/granite-dynamic-kv-growth.patch. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
add diarization support