audio-prep-pipeline 0.1.0

Convert raw MP3 audio into pretraining-ready WAV/FLAC, with validation and dataset manifest output.

audio-prep

Convert source audio into pretraining-ready WAV/FLAC, with validation and a dataset manifest as the handoff artifact to the pretraining pipeline. Discovery supports common FFmpeg-readable audio/video containers by default and can scan all files with --extensions all.

Defaults: 16 kHz, mono — the standard input spec for Wav2Vec2 / XLS-R style self-supervised speech pretraining. Override via CLI flags if a different downstream model needs something else.

Repo layout

audio-prep-pipeline/
├── src/audio_prep/
│   ├── config.py        # ConversionConfig: target spec + behavior knobs
│   ├── converter.py      # find_audio_files, convert_file, convert_batch
│   ├── validator.py      # probe_duration, validate_output
│   ├── chunker.py         # (optional) ChunkConfig, chunk_file, chunk_batch -- Silero VAD speech chunking
│   ├── manifest.py       # build_manifest, write_manifest (JSONL output)
│   ├── exceptions.py     # ConversionError, ProbeError, ChunkingError
│   └── cli.py             # `audio-prep convert ...` / `audio-prep chunk ...` entry point
├── tests/
│   ├── conftest.py        # synthetic-audio fixtures (no binary files checked in)
│   ├── test_converter.py
│   ├── test_validator.py
│   ├── test_chunker.py    # fake-detector tests, no real VAD model needed
│   ├── test_manifest.py
│   └── test_cli.py
├── .github/workflows/ci.yml   # lint + typecheck + test matrix (3.10-3.12)
├── .pre-commit-config.yaml     # ruff + mypy + basic hygiene hooks, runs on every commit
├── pyproject.toml               # deps, ruff config, mypy config, pytest config
└── Makefile                      # `make check` runs everything CI runs, locally

Commands

audio-prep has two independent subcommands. Neither depends on the other running first -- both scan --input-dir for supported source files directly.

• audio-prep convert - conversion only: ffmpeg resample/remix/re-encode into the target WAV/FLAC spec, then validation, then an optional manifest. Does not chunk.
• audio-prep chunk - chunking only: Silero VAD speech chunking straight from source audio, with its own resample/format/manifest options. Does not convert or validate.

convert pipeline stages

1. discovery (find_audio_files) — recursively find supported source files under an input directory.
2. conversion (convert_file / convert_batch) — shell out to ffmpeg to resample/remix/re-encode into the target WAV/FLAC spec. Mirrors the input directory's subfolder structure on output. Runs in a process pool since each conversion is an independent subprocess call.
3. validation (validate_output) — re-opens each converted file with soundfile and checks it actually matches the requested sample rate, channel count, and minimum duration. This catches the case where ffmpeg exits 0 but silently produced something degenerate.
4. manifest (build_manifest / write_manifest) — JSONL file, one row per source file, recording status (ok / conversion_failed / validation_failed), output path, duration, sample rate, and any error.

Conversion failures don't abort the batch — a bad file in a 50,000-file corpus shows up as one conversion_failed row in the manifest, not a crashed job three hours in.

chunk pipeline stages

1. discovery (find_audio_files) — recursively find supported source files under an input directory.
2. chunking (chunk_file / chunk_batch) — runs Silero VAD over each file (decoding/resampling via ffmpeg) and splits it into speech-only chunks bounded by a [min, max] duration window, so silence-heavy source recordings don't waste pretraining compute. Needs the chunking extra (torch + silero-vad, see Setup below).
3. manifest (build_chunk_manifest / write_manifest), optional — JSONL file, one row per source file, recording status (ok / chunking_failed), chunk count, and chunk paths.

Chunking failures (e.g. no speech detected) work the same way: chunk_batch returns a ChunkResult per file instead of raising.

Setup

# ffmpeg is a system dependency, not a pip package
sudo apt-get install ffmpeg   # or: brew install ffmpeg

pip install audio-prep-pipeline

Install directly from GitHub:

pip install "audio-prep-pipeline @ git+https://github.com/nattkorat/audio-prep-pipeline.git"

For local development:

make install   # pip install -e ".[dev]" + pre-commit install

Chunking needs an extra install -- make install alone does not pull in torch/silero-vad/tqdm, since convert doesn't need them:

pip install -e ".[chunking]"

Or from GitHub:

pip install "audio-prep-pipeline[chunking] @ git+https://github.com/nattkorat/audio-prep-pipeline.git"

The first chunk run needs either the silero-vad pip package installed, or network access so torch.hub can download snakers4/silero-vad once (it's then cached under ~/.cache/torch/hub). If neither is available, pass --allow-energy-fallback to use a lower-quality offline detector instead of failing.

Usage

audio-prep convert

audio-prep convert \
    --input-dir data/raw_mp3 \
    --output-dir data/wav16k \
    --format wav \
    --sample-rate 16000 \
    --workers 8 \
    --manifest data/manifest.jsonl

Flag	Default	Meaning
--input-dir	(required)	directory of source audio files
--output-dir	(required)	where converted output is written
--extensions	common FFmpeg audio/video extensions	comma-separated source extensions, or all to pass every regular file to FFmpeg
--format	wav	output format (wav or flac)
--sample-rate	16000	target sample rate
--channels	1	target channel count
--workers	4	parallel conversion workers
--min-duration-sec	0.5	validation fails files shorter than this
--overwrite	off	re-convert even if output already exists and passes validation
--normalize-loudness	off	apply EBU R128 loudness normalization (-23 LUFS)
--manifest	none	path to write a JSONL manifest

audio-prep chunk

Independent of convert -- scans --input-dir for supported source files and runs VAD chunking directly against them, decoding (and resampling, if --sample-rate doesn't match the source) via ffmpeg:

audio-prep chunk \
    --input-dir data/raw_mp3 \
    --output-dir data/chunks \
    --sample-rate 16000 \
    --format flac \
    --min-duration-sec 5 \
    --max-duration-sec 20 \
    --workers 4 \
    --manifest data/chunk_manifest.jsonl

Flag	Default	Meaning
--input-dir	(required)	directory of source audio files to scan and chunk
--output-dir	<input-dir>/chunks	where chunks are written
--extensions	common FFmpeg audio/video extensions	comma-separated source extensions, or all to pass every regular file to FFmpeg
--format	wav	output chunk format (wav or flac)
--sample-rate	16000	resample (via ffmpeg) to this rate before chunking if the source doesn't already match it
--min-duration-sec	5.0	drop chunks shorter than this
--max-duration-sec	20.0	split longer speech into windows this size
--workers	4	parallel chunking workers
--overwrite	off	re-chunk even if valid output already exists
--allow-energy-fallback	off	fall back to a low-quality energy detector if Silero can't load, instead of raising
--manifest	none	path to write a JSONL chunk manifest (source file, status, chunk count/paths)

chunk_file itself doesn't care about source extension -- it decodes whatever path it's given -- so the Python API can also chunk an existing WAV/FLAC corpus (e.g. convert_batch output) by passing source_files explicitly instead of relying on chunk_batch discovery:

from audio_prep import ConversionConfig, convert_batch, build_manifest, validate_output
from audio_prep import ChunkConfig, chunk_batch

config = ConversionConfig(output_format="wav", sample_rate=16_000, channels=1, num_workers=8)
results = convert_batch("data/raw_mp3", "data/wav16k", config)
validations = {r.output: validate_output(r.output, config) for r in results if r.success}
records = build_manifest(results, validations)

# source_files bypasses chunk_batch discovery, so this works
# directly against the already-converted WAV output above.
valid_outputs = [path for path, v in validations.items() if v.valid]
chunk_config = ChunkConfig(min_duration_sec=5, max_duration_sec=20, num_workers=4)
chunk_results = chunk_batch("data/wav16k", "data/wav16k/chunks", chunk_config, source_files=valid_outputs)

Development workflow

make format     # ruff format + autofix
make lint       # ruff check
make typecheck  # mypy --strict
make test       # pytest with coverage
make check      # all of the above -- run this before opening a PR

pre-commit (installed via make install) runs ruff + mypy + basic hygiene checks automatically on every commit. CI (.github/workflows/ci.yml) re-runs the same checks plus the full test matrix across Python 3.10–3.12 on every push and PR.

Extending this

Natural next additions, in roughly the order they'd come up:

• Streaming manifest writes for very large corpora, instead of holding all ConversionResults in memory before writing.

Note: This is the template, that you have to extend from. Main branch is protected so you have to create another branch to work on.