fix(diar): keep frames for words not yet tagged

[pskrunner14]

Without endpointing, words before the last ~10 min of long inputs all got one speaker (reported in #67). This PR adds a fix that keeps frames until their words are tagged.

AMI ES2002a (21 min) word speaker accuracy: 59.2% -> 99.7%

Fixes #67

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

• Configuration used: Repository: NVIDIA/NeMo-Speech.cpp/.coderabbit.yaml
• Review profile: CHILL
• Plan: Enterprise
• Run ID: 72b79303-de85-4a2a-9ff6-5e4a84167bf2

📥 Commits

Reviewing files that changed from the base of the PR and between 70468ff and dd0b1b9.

📒 Files selected for processing: 1

• src/asr/recognizer.cpp

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 11 remain after this review.

📝 Summary

Summary by CodeRabbit

• Bug Fixes
  • Improved speaker diarization reliability by retaining the audio frames needed to apply speaker labels to finalized words, including when a stream has not yet reached an endpoint.
  • Updated frame retention after finalization to preserve audio needed for speaker tagging while allowing later speech to continue processing.
  • Speaker labels are now associated with the frame containing the finalized word’s start, helping keep labels aligned with the recognized speech.

Diarization now preserves probability frames from stream start until a final result with words is available. It then advances the hold boundary to the frame containing the last word’s start time. Compaction respects the hold boundary, and resetting the stream rewinds a configured boundary to frame zero.

Changes: Diarization frame retention

Layer / File(s)	Summary
Hold boundary and compaction src/asr/diar/diar_pipeline.h, src/asr/diar/diar_pipeline.cpp	DiarStream adds a frame hold boundary, initialized as unset. Reset rewinds a configured boundary to frame zero. Compaction limits its cut to preserve held frames.
Recognizer hold lifecycle src/asr/recognizer.cpp	When diarization is enabled, recognition holds frames from frame zero. After a final result with words, it advances the boundary to the frame containing the last word’s start time.

Priority: ➖ Normal

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Bug fix · Severity of issue fixed: Medium

Merge Risk: ⚪ Minimal · up to dd0b1

For the documented diarization presets, the frame-retention change does not establish a speaker-attribution risk. No actionable merge-blocking issue was found, so the PR appears ready for normal validation.

Pre-merge checks: 4 · 1

Failed checks: 1 warning

Check name	Status	Explanation	Resolution
Docstring Coverage		Docstring coverage is 25.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 8 functions across 3 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

Passed checks: 4 passed

Check name	Status	Explanation
Title check		The title clearly and concisely describes the main diarization fix: retaining frames until words receive speaker tags.
Description check		The description directly explains the long-input speaker-tagging issue, the implemented fix, the reported accuracy improvement, and the linked issue.
Linked Issues check		The changes address issue #67. RecognitionStream holds diarization frames from frame 0 when diarization is enabled. After each final with words, it advances the hold point to the last final word sta…
Out of Scope Changes check		The changed files stay within issue #67. The DiarStream::hold_frames_from state, reset handling, compaction limit, and RecognitionStream integration directly support delayed word tagging. The comm…

Full details: Linked Issues check

Explanation

The changes address issue #67. RecognitionStream holds diarization frames from frame 0 when diarization is enabled. After each final with words, it advances the hold point to the last final word start, not to all pushed audio. DiarStream::maybe_compact limits compaction to frames before that hold point. A non-endpointed long input therefore retains frames until its words receive speaker tags. reset() also restores a configured hold to frame 0. Issue #67 has no explicit automated-test coding requirement.

Full details: Out of Scope Changes check

Explanation

The changed files stay within issue #67. The DiarStream::hold_frames_from state, reset handling, compaction limit, and RecognitionStream integration directly support delayed word tagging. The comment updates document this behavior. No unrelated behavior or feature change is demonstrated.

• Fix all pre-merge checks with AI

✨ Finishing Touches: 💡 1

📝 Generate docstrings: 💡

• Commit to this branch
• Create a new PR

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; the previous buffered-word speaker issue is fixed.

Summary

The PR keeps diarization frames until final words receive speaker tags.

• Diarization keeps word frames until their speaker tags are ready.

_{Reviews (2) · Last reviewed commit: "fix(diar): hold from the last tagged wor..." · Reviewed by Greptile}

[greptile-apps]

Want your agent to iterate on Greptile's feedback? Try greploops.

[anand-nv]

LGTM. Ok to merge once CI passes.