DLAI_deep_DYN

DLAI 2025 project: DYN Audio Representation for Efficient Learning and Inference.

This repository implements DYN Audio Representation, a block-floating-point inspired audio representation (8-bit sample + 6-bit block gain) for neural audio models.

Project Structure

The source code of the project is organized in the following way, in the src folder:

src
├── datasets
├── dyn_encoder
├── models
└── utils

• datasets/ – dataset wrapper class for LibriSpeech
• dyn_encoder/ – DYN block (and per sample) encoder and decoder
• models/ – baseline and DYN model definitions
• utils/ – utilities for training and evaluation

Training and testing scripts are located directly in the src/ folder.

Installation and Setup

Warning

You may see a warning when loading the dataset with torchaudio 2.8.0:

"UserWarning: In 2.9, this function's implementation will be changed to use torchaudio.load_with_torchcodec under the hood. Some parameters like normalize, format, buffer_size, and backend will be ignored. We recommend that you port your code to rely directly on TorchCodec's decoder instead."

This warning comes from torchaudio.datasets.LIBRISPEECH itself and cannot be disabled in the dataset wrapper.
The issue is tracked here: pytorch/audio#3902.

The scripts in this repository work fine with torchaudio 2.7.1 and 2.8.0 despite the warning, it will probably work with version 2.9.0 as well.

1. Prerequisites

• uv: This project uses uv for fast dependency management. Install it following the instructions at https://docs.astral.sh/uv/.

2. Clone the repository

git clone https://github.com/Lorenzoantonelli/DLAI_deep_DYN.git
cd DLAI_deep_DYN

3. Install dependencies

Use uv to install the project's dependencies:

uv sync --frozen

4. Download the checkpoints

Checkpoints are available in this folder: download. Download them into the Checkpoints/ folder.

5. Run the tests

Run all evaluation scripts and save results to a CSV (customize file_name in the script if needed):

Linux/macOS:

chmod +x run_all_experiments.sh
./run_all_experiments.sh

Windows:

.\run_all_experiments.bat

6. Train a model

Baseline models:

uv run src/train_baseline.py -h  # show all options

There are also two predefined training scripts for the baselines, that you can run with:

uv run src/train_baseline_16bit.py
uv run src/train_baseline_mu_law.py

To train a DYN model, run the following command with the model you want to train:

uv run src/train_dyn.py