Audio Dataset Preprocessor

Batch audio preprocessing for ML training data. Resamples, trims/pads to uniform duration, outputs WAV.

Same tool implemented in 5 ways to compare performance and ergonomics.

Benchmark

320 WAV files, resample to 16kHz, max 10s duration, 8 threads, Apple M1:

Implementation	Time	Speedup
Python (librosa)	14.19s	1x
Bash (ffmpeg CLI)	8.45s	1.7x
Zig (FFmpeg bindings)	1.29s	11x
Rust (FFmpeg bindings)	0.93s	15x
C (FFmpeg bindings)	0.85s	17x

Usage

# Zig
zig build run -- ./input ./output --sample-rate 16000 --min-duration 1.0 --max-duration 10.0

# Rust
cd rust_src && cargo build --release
./target/release/audio_preprocessor ./input ./output --min-duration 1.0 --max-duration 10.0

# C
cd c_src && make
./audio_preprocessor ./input ./output --min-duration 1.0 --max-duration 10.0

# Python
cd python_src && uv run python audio_preprocessor.py ./input ./output --min-duration 1.0 --max-duration 10.0

# Bash
./shell_test.sh ./input ./output --min-duration 1.0 --max-duration 10.0

Test Data

MusicNet Dataset on Kaggle