Practical overview
When it fits
Use it when you need a code-based neural audio compression and reconstruction workflow, or a documented starting point for codec training and audio-generation research.
Who it is for
- Audio developers building compression or reconstruction into processing pipelines.
- Researchers working on neural audio codecs or generative audio workflows.
- Practitioners who want to train the repository's baseline model.
Documented uses
- Encode audio files into .dac compressed-code files.
- Decode .dac files into WAV audio.
- Use codec functions in a Python audio-processing workflow.
- Run baseline codec training with the supplied configuration.
Requirements & limits
The repository is public and specifies the MIT license. It documents pip installation and Docker-based workflows; no hosted service or account requirement is described.
- The repository reports 44.1 kHz encoding at 8 kbps and approximately 90× compression; treat these as project specifications rather than a universal quality result.
- Model weights are downloaded automatically on first CLI use, or can be cached for the 16, 24 or 44.1 kHz variant.
- The README warns that encoding long files as one signal may exceed GPU memory and points to compress/decompress functions for long files.
- Training documentation includes single- and multi-GPU commands and a Docker Compose development environment.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Encode audio and reconstruct it as WAV
Use the command-line tools to create .dac codes from audio and decode them into WAV files.
Before you start
- Install descript-audio-codec with pip or from the GitHub repository.
- Have input audio available; the codec downloads model weights automatically on first encode or decode.
Steps
Install with the command below or the command below.
pip install descript-audio-codecpip install git+https://github.com/descriptinc/descript-audio-codecRun the command below.
python3 -m dac encode /path/to/input --output /path/to/output/codesRun the command below.
python3 -m dac decode /path/to/output/codes --output /path/to/reconstructed_input
Expected outcome
The encode command creates .dac files; decode creates WAV files. Both preserve the input directory structure relative to the input root.
Guide limitations
- Encoding a long file as one signal in the programmatic interface may run out of GPU memory; the repository recommends compress/decompress functions for long files.
- The commands and documented model variants are repository instructions, not execution-tested here.
Prepare the baseline training environment
Build and launch the documented Docker Compose environment before running the baseline training command.
Before you start
- Install the repository's training dependencies with
pip install -e ".[dev]". - Docker and Docker Compose are needed for the documented container setup.
Steps
Build the development image with the command below.
docker compose buildLaunch the container with the command below.
docker compose run -p 8888:8888 -p 6006:6006 devRun the single-GPU baseline command:
export CUDA_VISIBLE_DEVICES=0followed by the command below.python scripts/train.py --args.load conf/ablations/baseline.yml --save_path runs/baseline/
Expected outcome
The README provides a baseline training command that saves to `runs/baseline/` from the launched training environment.
Guide limitations
- The guide documents setup and command syntax; it does not establish that training has been run successfully on a particular machine.
- The repository also documents a separate multi-GPU command, which requires GPU devices and
torchrun.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 14 October 2026
- Original website title
- descriptinc/descript-audio-codec · Official homepage title · retained in its source language