Architecture¶

The package is intentionally small and module-oriented.

audio_prep/
├── cli.py          # argparse entry point
├── config.py       # conversion configuration
├── converter.py    # discovery and FFmpeg conversion
├── chunker.py      # VAD chunking
├── validator.py    # output validation and duration probing
├── manifest.py     # JSONL manifest records and writer
└── exceptions.py   # package exceptions

Design Choices¶

FFmpeg For Decode And Encode¶

Source decoding is delegated to FFmpeg because real corpora often contain VBR files, odd headers, unusual containers, corrupt metadata, and other cases that pure Python decoders handle less consistently.

Result Objects Instead Of Batch Exceptions¶

Conversion and chunking return result objects per file. This lets a large corpus finish even when individual files are broken.

Validation After Conversion¶

FFmpeg success is not treated as enough. Converted output is reopened and checked for sample rate, channel count, readability, and minimum duration.

Chunking Dependencies¶

The base package includes Torch, Silero VAD, and tqdm so both convert and chunk are ready after pip install audio-prep-pipeline. Chunker imports are still lazy, so conversion-only commands avoid VAD startup cost.