[codex] docs: add long-form audio cutting guide (#2136)

* docs: add long-form audio cutting guide

Signed-off-by: Lawrence Lane <llane@nvidia.com>

* docs: clarify audio pretrain finalization

Signed-off-by: Lawrence Lane <llane@nvidia.com>

* docs: address reviewer feedback on long-form audio cutting guide

- state that source IDs must be unique after tar-key normalization
  (make_snippet_id replaces . / \ with _, so session.1 and session_1 collide)
- clarify that a missing segments key fails OverlapFilterStage validation,
  while an empty list is accepted and yields zero snippets
- warn that dry runs must use a fresh, nonexistent tar path so preview
  manifests are not reconciled against a stale archive
- document the missing_audio/corrupted_audio counters as sparse and
  recommend defensive .get() access; drop the misleading zero-valued key
  from the example metrics JSON

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Signed-off-by: Lawrence Lane <llane@nvidia.com>

---------

Signed-off-by: Lawrence Lane <llane@nvidia.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: lbliii

fern/versions/main.yml

@@ -474,6 +474,9 @@ navigation:
               - page: ALM Tutorial
                 path: ./main/pages/curate-audio/tutorials/alm.mdx
                 slug: alm
+              - page: Long-Form Audio Cutting for ALM Pretraining
+                path: ./main/pages/curate-audio/tutorials/audio-pretrain.mdx
+                slug: audio-pretrain
               - page: ReadSpeech Tutorial
                 path: ./main/pages/curate-audio/tutorials/readspeech.mdx
                 slug: readspeech

fern/versions/main/pages/curate-audio/index.mdx

@@ -140,4 +140,11 @@ quality-filtering
 basic-workflow
 </Card>
 
+<Card title="Long-Form Audio Cutting" href="/curate-audio/tutorials/audio-pretrain">
+Turn long-form diarized recordings into mono, resampled snippet archives and manifests for audio language model pretraining
+alm-pretraining
+long-form-audio
+webdataset
+</Card>
+
 </Cards>

fern/versions/main/pages/curate-audio/process-data/alm/data-builder.mdx

@@ -124,3 +124,4 @@ The `stats` dictionary contains the following counters:
 - **[ALM Overlap Filtering](/curate-audio/process-data/alm/overlap-filtering)**: Next stage in the ALM pipeline
 - **[ALM Pipeline Concepts](/about/concepts/audio/alm-pipeline)**: Architectural overview
 - **[ALM Tutorial](/curate-audio/tutorials/alm)**: End-to-end walkthrough with sample data
+- **[Long-Form Audio Cutting](/curate-audio/tutorials/audio-pretrain)**: Cut selected diarized time ranges into reusable snippet audio and tar manifests

fern/versions/main/pages/curate-audio/process-data/alm/overlap-filtering.mdx

@@ -101,3 +101,4 @@ Monitor the yield by comparing `filtered_dur` to `total_dur_window` in the outpu
 - **[ALM Data Builder](/curate-audio/process-data/alm/data-builder)**: Previous stage in the ALM pipeline
 - **[ALM Pipeline Concepts](/about/concepts/audio/alm-pipeline)**: Architectural overview
 - **[ALM Tutorial](/curate-audio/tutorials/alm)**: End-to-end walkthrough with sample data
+- **[Long-Form Audio Cutting](/curate-audio/tutorials/audio-pretrain)**: Filter diarization overlap before extracting physical ALM pretraining snippets

fern/versions/main/pages/curate-audio/tutorials/alm.mdx

@@ -297,6 +297,7 @@ The `ALMManifestReader` discovers all `.jsonl` and `.json` files in the director
 
 After completing this tutorial, explore:
 
+- **[Long-Form Audio Cutting](/curate-audio/tutorials/audio-pretrain)**: Extract, resample, and package bounded audio snippets when you need physical audio assets rather than metadata windows
 - **[ALM Data Builder](/curate-audio/process-data/alm/data-builder)**: Detailed reference for window construction
 - **[ALM Overlap Filtering](/curate-audio/process-data/alm/overlap-filtering)**: Detailed reference for overlap filtering
 - **[ALM Pipeline Concepts](/about/concepts/audio/alm-pipeline)**: Architectural overview

fern/versions/main/pages/curate-audio/tutorials/audio-pretrain.mdx

@@ -34,4 +34,11 @@ windowing
 speaker-diarization
 </Card>
 
+<Card title="Long-Form Audio Cutting" href="/curate-audio/tutorials/audio-pretrain">
+Extract bounded mono and resampled snippets from long-form diarized audio, then package them as a training manifest and WebDataset-compatible tar.
+alm-pretraining
+snippet-extraction
+webdataset
+</Card>
+
 </Cards>