An audio classification tool that labels audio files into three categories: speech, music, and noise.

Overview

Janitor uses the Audio Spectrogram Transformer (AST) model pretrained on AudioSet to classify audio content. It consists of two components:

• Service: A GPU-accelerated inference server that runs the AST model with batched processing
• CLI: A client that extracts mel-filterbank features from audio files and sends them to the service for classification

Architecture

┌─────────────────┐      safetensors      ┌─────────────────┐
│   CLI Client    │  ─────────────────►   │  Inference      │
│                 │                       │  Service        │
│  - Audio decode │      HTTP POST        │                 │
│  - Resample     │  ◄─────────────────   │  - AST Model    │
│  - Fbank extract│      JSON label       │  - Batching     │
└─────────────────┘                       └─────────────────┘

Features

• Classify individual files or entire directories recursively
• Copy or move files to separate directories based on classification
• Concurrent file processing with configurable parallelism
• GPU acceleration with CUDA support
• Batched inference for improved throughput

Quick Start

1. Build and run the service

cd service

# Build the model (requires conda)
./build_model.sh

# Run on CPU
cargo build --release

# Or with CUDA support
TORCH_CUDA_VERSION=cu121 cargo build --release

# Start the service
cargo run --release

2. Run the CLI

cd cli
cargo build --release

# Classify a single file
cargo run --release -- audio.mp3

# Classify a directory
cargo run --release -- /path/to/audio/

# Move files to separate directories by label
cargo run --release -- move /path/to/audio/ \
  --speech-dir ./speech \
  --music-dir ./music \
  --noise-dir ./noise

Components

Requirements

• Rust 1.70+
• Conda (for building the model)
• CUDA 11.8 or 12.1 (optional, for GPU acceleration)

License

GPL-3.0