Audio: add local duration-aware ASR batching

[mohammadaaftabv]

Summary

• rebases the audio batching work onto current main
• ports PR Aaftabv/qwen1967 local noio control #2093's local duration/cost batching and model-safe input segmentation
• stitches segmented FastConformer results back into original parent-row order
• adds an adapter-call cap and exact NeMo model.transcribe batch mapping
• replaces the inert policy-only implementation from Audio: add local duration-aware bucketing #2225 on the correctly named branch

Validation

• 256 passed, 7 deselected across audio inference, ASR adapters, and segmentation
• Ruff format/check and git diff --check pass
• local RTX 3080 Ti parity passed against nithinraok/Curator:nkoluguri/integration-test
• policy-on calls: [1, 2]; policy-off and reference calls: [3]
• raw transcripts, normalized transcripts, identities, and row order match exactly across all arms

This PR intentionally remains a draft.

[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

Adds local duration-aware ASR batching, fixed-duration model-input segmentation, ordered result scattering, and parent transcript stitching, together with configuration documentation and tests.

• Introduces a validated duration bucket and aggregate-audio packing policy.
• Adds independent adapter-call item caps and restores results to original item order.
• Splits long waveforms into bounded model inputs and reassembles their parent results.
• Expands tutorial, Hydra configuration, adapter-mapping, segmentation, and batching coverage.

Confidence Score: 4/5

The fixed context-free segmentation should be corrected before merging because it can corrupt transcripts around chunk boundaries; the segmented diagnostics schema should also be stabilized.

Long inputs are cut at arbitrary sample positions and independently decoded without overlap before plain text concatenation, while multi-segment adapter diagnostics are emitted under a different schema than single-segment results.

Files Needing Attention: nemo_curator/stages/audio/inference/asr/stage.py, nemo_curator/stages/audio/model_input_segmentation.py

Important Files Changed

Filename	Overview
nemo_curator/stages/audio/inference/asr/stage.py	Integrates segmentation, local batching, result scattering, and stitching; fixed context-free cuts can damage boundary transcription and stitched diagnostics acquire a duration-dependent schema.
nemo_curator/stages/audio/inference/batch_policy.py	Implements validated, stable duration bucketing and greedy aggregate-duration packing while preserving original indices.
nemo_curator/stages/audio/model_input_segmentation.py	Produces complete bounded sample intervals, but its hard boundaries provide no overlap or speech-aware context for ASR.
nemo_curator/models/asr/base.py	Extends the documented adapter input contract with actual chunk duration.
tests/stages/audio/inference/test_asr_stage.py	Thoroughly covers batch membership, caps, scattering, and literal stitching behavior, but does not test speech spanning a real model boundary or stable extras schema.
tests/stages/audio/inference/test_batch_policy.py	Covers policy validation, boundaries, packing, ordering, and exact-once index behavior.
tests/stages/audio/test_model_input_segmentation.py	Covers sample interval completeness and duration-boundary edge cases.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Parent audio rows] --> B[Load and resample]
  B --> C[Fixed-duration segments]
  C --> D[Duration buckets]
  D --> E[Audio and item caps]
  E --> F[Adapter calls]
  F --> G[Scatter to segment order]
  G --> H[Stitch parent transcripts]
  H --> I[Original parent-row order]

Loading

_{Reviews (1): Last reviewed commit: "docs(audio): document local duration buc..." | Re-trigger Greptile}

[greptile-apps]

Hard cuts lose speech context

When speech crosses max_inference_duration_s, the waveform is cut at a fixed sample offset and each side is decoded independently without overlap before their text is joined, causing words around the boundary to be omitted, duplicated, or mistranscribed.

Knowledge Base Used: Audio curation stages

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[greptile-apps]

Segmenting changes diagnostics schema

For a segmented parent with adapter extras, this replaces the documented adapter-specific dictionary with {"chunks": [...]}, so consumers must handle a second duration-dependent schema instead of reading established keys such as detected_language at their normal location.

Knowledge Base Used: Model integrations and inference

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!