descript-audio-codec

What it is and what it does

Descript Audio Codec is a neural audio codec that compresses audio into discrete codes at 8 kbps bitrate while preserving high fidelity. It uses an improved RVQGAN architecture and works universally across speech, music, and environmental audio. The package provides both command-line tools and a Python API for programmatic use via its runtime dependencies including torch, torchaudio, einops, numpy, argbind, descript-audiotools, and tqdm.

The codec is designed as a drop-in replacement for audio language modeling applications. Pre-trained model weights for 16 kHz, 24 kHz, and 44.1 kHz are automatically downloaded and cached on first use. The package is actively maintained and has no known vulnerabilities.

Use it for:

• Compress audio for storage or transmission while maintaining perceptual quality in generative audio models
• Use as a backend codec for audio language models or music generation systems
• Encode audio files into discrete codes for downstream machine learning tasks
• Reconstruct high-fidelity audio from compressed .dac files for playback or further processing
• Replace existing codecs in audio ML pipelines that require a universal codec across domains

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Compresses audio into discrete codes at 8 kbps bitrate and reconstructs it with high fidelity, supporting 16 kHz, 24 kHz, and 44.1 kHz sampling rates across speech, music, and environmental audio.

Yes, if you need audio compression for ML applications. The codec is actively maintained, has no known vulnerabilities, uses a permissive MIT license, and offers strong compression with high fidelity. Install friction is low for developers already using PyTorch. Not suitable if you need real-time or streaming inference without GPU, or if you require lossless compression.

Install

descript-audio-codec on PyPI

pip install descript-audio-codec

uv add descript-audio-codec

poetry add descript-audio-codec

Installing descript-audio-codec

Before you install

Low install friction with a pure Python wheel. Requires torch and torchaudio as runtime dependencies, which are substantial downloads but standard for audio ML work. Repository is actively maintained with recent commits.

License in practice

MIT license permits commercial and private use, modification, and distribution with minimal restrictions—suitable for most projects.

Quickstart

pip install descript-audio-codec

import descript_audio_codec
from descript_audiotools import AudioSignal

model = descript_audio_codec.DAC.load(model_path)
model.to('cuda')

signal = AudioSignal('input.wav')
signal.to(model.device)
x = model.preprocess(signal.audio_data, signal.sample_rate)
z, codes, latents, _, _ = model.encode(x)
y = model.decode(z)
y.write('output.wav')

Requires CUDA-capable GPU or CPU; torch and torchaudio must be installed. Model weights are automatically downloaded on first use.

Verify before relying

• Whether the package supports real-time or streaming encoding/decoding, or only batch processing
• Memory requirements for typical audio file sizes and whether chunking strategies are built-in for long files
• Compatibility with non-PyTorch inference frameworks or ONNX export

Package facts

Evidence: descript_audio_codec-1.0.0-py3-none-any.whl

Keywords: audio, compression, machine learning

Tags

More Libraries packages

Further reading