AudioQuant

AI-powered ultra-low bitrate audio codec CLI built on SNAC

Compress audio at 0.98–2.6 kbps — up to 392x smaller than WAV — with near-original quality using neural audio codecs.

WAV:  1 min speech = 2.88 MB
MP3:  1 min speech = 960 KB
SNAC: 1 min speech = 7.35 KB  ← 392:1 compression

Features

• Ultra-low bitrate: 0.98 kbps (speech) to 2.6 kbps (music)
• Near-original quality: MUSHRA 88.4/100 for speech at 0.98 kbps
• 3 pretrained models: Optimized for speech (24kHz), music (32kHz), high-fidelity (44kHz)
• Custom .snac format: Compact binary format with 44-byte header + token data
• Quality metrics: SNR, SI-SNR, spectral distance, PESQ (optional)
• Streaming mode: Chunked encoding for large files (1-hour+ podcasts)
• GPU acceleration: CUDA, MPS (Apple Silicon), and CPU support
• Rich terminal UI: Progress bars, colored tables, styled panels

Installation

Requires Python 3.10+

pip install audioquant

# With quality metrics (optional)
pip install audioquant[metrics]

From source

git clone https://github.com/wjddusrb03/audioquant
cd audioquant
pip install -e ".[dev]"

Quick Start

# Compress audio (WAV 2.88MB → SNAC 7.35KB)
audioquant compress podcast.wav

# Decompress back to WAV
audioquant decompress podcast.snac

# Get file info
audioquant info podcast.snac

# Compare all models
audioquant compare podcast.wav

# Benchmark across multiple files
audioquant benchmark file1.wav file2.mp3 file3.flac

Commands

compress

Compress an audio file to .snac format.

audioquant compress input.wav                          # Default: snac_24khz
audioquant compress input.wav -m snac_44khz            # High-quality music
audioquant compress input.wav -o output.snac           # Custom output path
audioquant compress input.wav -d cuda                  # Force GPU
audioquant compress input.wav --no-progress            # No progress bar

Output:

╭──────── Compression Complete ────────╮
│ Input:  podcast.wav  (2.88 MB)       │
│ Output: podcast.snac  (7.35 KB)      │
│ Model:  snac_24khz                   │
│ Ratio:  392:1                        │
│ Time:   0.85s  (7.1x realtime)       │
│ Device: CUDA (NVIDIA RTX 4090)       │
╰──────────────────────────────────────╯

decompress

Decompress a .snac file back to audio.

audioquant decompress podcast.snac                     # Default: WAV
audioquant decompress podcast.snac --format flac       # FLAC output
audioquant decompress podcast.snac -o restored.wav     # Custom path
audioquant decompress podcast.snac -d cuda             # Force GPU

info

Display metadata about any audio or .snac file.

audioquant info podcast.wav       # Audio file info
audioquant info podcast.snac      # SNAC file info

compare

Compare compression across all three SNAC models.

audioquant compare input.wav                    # Table output
audioquant compare input.wav --json             # JSON output
audioquant compare input.wav -m snac_24khz,snac_44khz  # Specific models
audioquant compare input.wav --no-metrics       # Skip quality metrics

Output:

         Benchmark Results
┌────────────┬──────┬────────┬───────┬───────┐
│ Model      │ Rate │ Size   │ Ratio │ SNR   │
├────────────┼──────┼────────┼───────┼───────┤
│ snac_24khz │ 24k  │ 7.3 KB │ 392:1 │ 28.5  │
│ snac_32khz │ 32k  │ 14.2KB │ 203:1 │ 31.2  │
│ snac_44khz │ 44k  │ 19.5KB │ 148:1 │ 33.8  │
└────────────┴──────┴────────┴───────┴───────┘

benchmark

Benchmark models across multiple audio files.

audioquant benchmark *.wav                          # All WAV files
audioquant benchmark file1.wav file2.mp3 --json     # JSON output
audioquant benchmark file1.wav -m snac_24khz,snac_44khz  # Specific models
audioquant benchmark file1.wav -d cuda              # Force GPU

stream

Encode large files using chunked streaming mode (lower memory usage).

audioquant stream long_podcast.wav -o podcast.snac
audioquant stream lecture.wav -o lecture.snac --chunk-duration 2.0
audioquant stream lecture.wav -o lecture.snac -m snac_44khz -d cuda

SNAC Models

Model	Sample Rate	Bitrate	Quality (MUSHRA)	Best For	Params
snac_24khz	24,000 Hz	0.98 kbps	88.4	Speech, podcasts	19.8M
snac_32khz	32,000 Hz	1.9 kbps	—	Music, sound effects	54.5M
snac_44khz	44,100 Hz	2.6 kbps	—	High-fidelity music	54.5M

vs Other Codecs

Codec	Bitrate	Speech Quality	Music Quality	CLI Tool
SNAC (AudioQuant)	0.98 kbps	88.4	76.8	✅
EnCodec (Meta)	1.5 kbps	78.3	64.4	✅
DAC (Descript)	2.5 kbps	85.0	54.0	✅

How It Works

Encoding (Compression)

Audio File  →  Resample  →  AI Encoder  →  Multi-Scale Tokens  →  .snac File
(WAV/MP3)     (to model     (SNAC neural    (3-4 RVQ levels       (44-byte header
               rate)         network)         at different           + uint16 tokens)
                                              temporal rates)

SNAC uses Residual Vector Quantization (RVQ) at multiple temporal scales (rates vary by model):

• Level 0 (coarse, 10–14 Hz): Captures overall melody and tone
• Level 1 (medium, 21–29 Hz): Captures timbre and emotion
• Level 2 (fine, 42–57 Hz): Captures pronunciation and detail
• Level 3 (finest, 83–115 Hz): High-frequency detail (32kHz/44kHz models only)

.snac File Format

┌─────────────────────────────────────┐
│ Header (44 bytes)                   │
│   Magic: "SNAC"                     │
│   Version: 1                        │
│   Model: "snac_24khz"              │
│   Sample rate, duration, channels   │
├─────────────────────────────────────┤
│ Token Data                          │
│   Level 0: [count] [tokens...]      │
│   Level 1: [count] [tokens...]      │
│   Level 2: [count] [tokens...]      │
└─────────────────────────────────────┘

Supported Formats

Format	Read	Write
WAV	✅	✅
MP3	✅	✅
FLAC	✅	✅
OGG	✅	✅
OPUS	✅	✅
SNAC	✅	✅

Project Structure

audioquant/
├── src/audioquant/
│   ├── __init__.py       # Package version
│   ├── models.py         # Dataclasses & model registry
│   ├── codec.py          # SNAC wrapper (encode/decode)
│   ├── audio_io.py       # Audio file I/O (WAV/MP3/FLAC)
│   ├── format.py         # .snac binary file format
│   ├── metrics.py        # Quality metrics (SNR, PESQ, spectral)
│   ├── streaming.py      # Chunked encode/decode for large files
│   ├── benchmark.py      # Cross-model comparison
│   ├── display.py        # Rich terminal output
│   └── cli.py            # Click CLI commands
├── tests/                # Test suite
├── pyproject.toml
├── README.md
├── README_KO.md
└── LICENSE

Testing

# Run all fast tests (no SNAC model required)
pytest tests/ -v

# Run specific test file
pytest tests/test_format.py -v
pytest tests/test_metrics.py -v

Dependencies

Package	Purpose
snac	SNAC neural audio codec
torch	Tensor operations
torchaudio	Audio I/O and resampling
click	CLI framework
rich	Terminal formatting
numpy	Numerical operations
pesq (optional)	PESQ quality metric

Quality Metrics

Metric	Description	Range
SNR	Signal-to-Noise Ratio	Higher = better (dB)
SI-SNR	Scale-Invariant SNR	Higher = better (dB)
Spectral Distance	Multi-resolution STFT distance	Lower = better
PESQ	Perceptual speech quality	1.0–4.5
Compression Ratio	Original / compressed size	Higher = smaller

Issues & Contributions

Found a bug? Have a feature request? Please let us know!

• Bug reports: Create an issue
• Feature requests: Create an issue

All feedback is welcome. If something doesn't work as expected, please report it — it helps make AudioQuant better for everyone!

License

MIT

Acknowledgments

• SNAC — Multi-Scale Neural Audio Codec by Hubert Siuzdak
• TurboQuant — Inspiration for the practical CLI-on-top-of-research pattern