feat(tutorials/readspeech): add interactive Jupyter notebook tutorial

[shubhamNvidia]

Summary

• Add interactive Jupyter notebook walkthrough for DNS Challenge Read Speech audio curation pipeline
• Provides step-by-step execution with visualization of quality score distributions
• Demonstrates threshold tuning to control quality vs. data retention tradeoffs

What's Included

• tutorials/audio/readspeech/readspeech_tutorial.ipynb - Interactive notebook covering:
  1. Dataset download and manifest creation
  2. Pipeline stages: Mono → VAD → Band Filter → UTMOS → SIGMOS → Speaker Separation
  3. Quality filtering with configurable thresholds
  4. Visualization of intermediate outputs and score distributions

Test Plan

• Open notebook in Jupyter and run all cells
• Verify plots render correctly
• Confirm pipeline stages execute without errors

[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.

[shubhamNvidia]

/ok to test 972b1c1

[greptile-apps]

Greptile Summary

This PR adds an interactive Jupyter notebook tutorial (tutorials/audio/readspeech/readspeech_tutorial.ipynb) that walks through the DNS Challenge Read Speech audio curation pipeline, covering dataset download, all filter stages, quality-score visualization, and threshold tuning. The .secrets.baseline is regenerated to track the new notebook's embedded plot images in place of stale multimodal notebook entries.

Confidence Score: 5/5

Safe to merge; all findings are P2 style suggestions that do not affect correctness.

No P0 or P1 issues found. The two P2 comments cover a potentially misleading threshold-sensitivity chart scope and an unguarded Ray cleanup path — neither blocks functionality. The secrets baseline update is a clean regeneration.

No files require special attention.

Important Files Changed

Filename	Overview
tutorials/audio/readspeech/readspeech_tutorial.ipynb	New interactive tutorial notebook for DNS Challenge Read Speech pipeline; two minor P2 issues: threshold-sensitivity chart misleads with passing-only sample scope, and Ray cluster lifecycle has no cleanup on failure.
.github/workflows/config/.secrets.baseline	Baseline regenerated via detect-secrets scan: swaps stale multimodal notebook entries (base64 plot images) for the new audio tutorial notebook entries; timestamp updated to current run.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Cell 2: Config & Imports] --> B[Cell 4: CreateInitialManifestReadSpeechStage\nstandalone preview]
    A --> C[Cell 6: Full Pipeline]
    C --> D[CreateInitialManifestReadSpeechStage\nauto_download=True]
    D --> E[AudioDataFilterStage]
    E --> E1[Mono Conversion 48 kHz]
    E1 --> E2[VAD 2-60 s]
    E2 --> E3[Band Filter full_band]
    E3 --> E4[UTMOS >= 3.4]
    E4 --> E5[SIGMOS OVRL >= 3.5 / NOISE >= 4.0]
    E5 --> E6[Speaker Separation]
    E6 --> F[AudioToDocumentStage]
    F --> G[JsonlWriter -> RESULT_DIR]
    G --> H[Cell 9: load_jsonl_results]
    H --> I[Cell 11: Score Distributions]
    H --> J[Cell 13: Band Classification]
    H --> K[Cell 15: Speaker Distribution]
    H --> L[Cell 17: Threshold Sensitivity]
    G --> M[Cell 19: ray_client.stop]

Loading

_{Reviews (2): Last reviewed commit: "Merge branch 'main' into pr/notebook" | Re-trigger Greptile}

[greptile-apps]

Internal codename leaked into public tutorial

The comments reference "Xenna" — which appears to be an internal project codename — in two places. This will be confusing to external contributors and users who won't know what "Xenna" refers to. These should be replaced with a neutral description.

The second occurrence (# Use the default executor (Xenna), matching pipeline.py CLI defaults.) should similarly be reworded to something like # Use the default executor, matching pipeline.py CLI defaults.

[greptile-apps]

Pass rate denominator may be misleading

The pass rate is calculated as len(results) / MAX_SAMPLES, but MAX_SAMPLES is an upper bound on how many files to download — the pipeline may have actually processed fewer than MAX_SAMPLES inputs (e.g., if the dataset has fewer matching files, or some are skipped). Using MAX_SAMPLES as the denominator can silently understate the true pass rate.

Consider tracking the actual number of inputs from the manifest stage output and using that as the denominator, or at least adding a note that the denominator is the requested cap, not the actual input count.

[sarahyurick]

/ok to test e26ad56