fix(audio-postprocessing): accept diarizer segment shapes; make TimestampMapperStage agent-ready

[shubhamNvidia]

Contracts: describe() + *_key params on TimestampMapperStage (maps split-clip timestamps back to the original file).

Bug fix — diarize → timestamp-map pipelines crashed.
Two producers emit two different segment shapes: VAD and speaker-separation emit [start, end] pairs, while the diarizers emit {start, end, speaker} dicts. The stage only read pairs, so a real diarizer's output raised KeyError: 0 and killed the pipeline. It now accepts both shapes, and echoes back the shape it was given — rewriting a diarizer's dict as a bare pair would have thrown away the speaker label, which is the whole point of running diarization. A single malformed segment is now skipped instead of taking down the batch.

Depends on the agent-ready foundation, #2332 — that must merge first (this branch imports nemo_curator/stages/audio/_agent/, so CI here stays red until it lands).

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[greptile-apps]

Greptile Summary

The PR makes TimestampMapperStage agent-ready and adds configurable field names while extending diarization handling to preserve both pair and speaker-labeled dictionary segment shapes.

• Adds stage contract metadata, configurable input/output keys, and agent readiness.
• Maps diarization-only timing through concatenation metadata and prioritizes precise start_ms/end_ms ranges when available.
• Preserves speaker labels, skips unreadable segment shapes, and adds regression coverage for the supported pipelines.

Confidence Score: 4/5

The malformed-bound handling should be fixed before merging because a single non-finite diarization timestamp can still terminate processing despite the new skip-on-malformed contract.

The new parser accepts every float-convertible bound, allowing non-finite values to reach integer timestamp conversion and reversed intervals to produce negative duration data.

Files Needing Attention: nemo_curator/stages/audio/postprocessing/timestamp_mapper.py

Important Files Changed

Filename	Overview
nemo_curator/stages/audio/postprocessing/timestamp_mapper.py	Adds agent contracts, configurable keys, and dual-shape diarization support, but malformed numeric bounds can still crash processing or emit invalid durations.
tests/stages/audio/postprocessing/test_timestamp_mapper.py	Adds focused regressions for dictionary and pair segments, speaker-label preservation, overlap handling, sanitization, and precise multispeaker mapping.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[AudioTask] --> B{segment mappings?}
  B -->|Yes| C{start/end present?}
  C -->|Yes| D[Translate precise range]
  C -->|No| E[Translate diarization segments]
  B -->|No| F[Derive timing from start/end, diarization, or duration]
  D --> G[Allowlisted output]
  E --> G
  F --> G

Loading

_{Reviews (1): Last reviewed commit: "feat(audio-postprocessing): make Timesta..." | Re-trigger Greptile}

[greptile-apps]

Malformed numeric bounds remain readable

When a diarization segment contains NaN, infinity, or an end timestamp earlier than its start, _segment_bounds accepts it as readable, causing integer timestamp conversion to raise or negative duration fields to be emitted instead of skipping the malformed segment.