Firstly, install the conda env provided by the environment.yaml file. To run correctly the code, we have to consider 4 main files:

📁 audio-restore/
├── 📂 quality_estimator_code/
│   └── 📂 python/
│       └── estimate_quality.py
├── 📂 restore_code/
│   └── 📂 python/
│       └── start.py
├── music_gen.py
├── text_to_music_eval_prompts.csv

On the csv file, you can add all the prompts from which you want to generate music, then run music_gen.py.

In start.py you have to pay attention on 2 constants:

• INPUT_FILE: music_gen generated audio path
• OUTPUT_FIN: restored audio path

So write correctly these 2 paths.

Then you only have to execute the following command:

python -m quality_estimator_code.python.estimate_quality

It will try to restore the INPUT_FILE and then it will estimate a quality score.

🎧 Audio Restoration for Generative Models

This repository implements a modular enhancement pipeline designed to improve the perceptual quality of audio generated by text-to-music models such as MusicGen. While these models are capable of generating coherent musical ideas from text prompts, their outputs often suffer from compression artifacts, low bit depth, and a lack of high-frequency fidelity. Our pipeline addresses this gap by introducing a U-Net-based restoration module and a CNN-based perceptual quality estimator.

---

🧠 Abstract

Recent generative audio models like MusicGen produce impressive musical samples directly from text. However, these generations often suffer from low bit depth, quantization artifacts, noise, and a general lack of high-fidelity detail compared to studio-quality audio.

In this project, we propose a novel approach to enhance the perceptual quality of generated audio. We employ a U-Net architecture to perform post-processing on the audio, removing artifacts and restoring acoustic richness. Additionally, we introduce a CNN-based quality assessment module trained to predict perceptual audio quality. This module provides feedback that guides the U-Net during training. Our results show measurable improvements both in objective metrics and subjective listening tests.

---

🗂️ Repository Structure

📁 audio-restore/
├── 📂 restore_code/
│   ├── 📂 Jupyter/
│   │   ├── convert_wav_audio.ipynb        # Converts MP3 to mono WAV
│   │   ├── downsample_song.ipynb          # Simulates degradation (downsampling, noise)
│   │   └── audio_restore_train.ipynb      # Trains U-Net on degraded/clean waveform pairs
│   └── 📂 python/
│       └── start.py                       # Restore audio quality
├── 📂 quality_estimator_code/
│   ├── 📂 Jupyter/
│   │   ├── Discriminator.ipynb            # Trains CNN quality predictor on spectrograms
│   └── 📂 python/
│       └── estimate_quality.py            # Launches the full pipeline
├── music_gen.py                           # Wraps MusicGen inference from prompts
├── 📂 data/
│   └── 📂 train/                          # Example dataset (Free Music Archive subset)
├── 📂 models/                             # Saved weights for U-Net and CNN

---

🔩 Installation

Install dependencies with:

conda env create -f environment.yaml

---

💻 System Overview

The pipeline consists of the following components:

1. Music Generation

   • Uses a pre-trained model (e.g., MusicGen) to produce audio from text.

2. Audio Degradation

   • Downsamples and alters generated audio to simulate poor fidelity.

3. U-Net Restoration

   • Enhances low-quality waveform using an encoder-decoder network.

4. CNN Evaluation

   • Predicts perceptual quality of restored audio using spectrograms.

5. Loss Functions

   • Combined L1 loss and CNN-based perceptual loss guide training.

---

Experimental Setup

• Dataset: MusicGen outputs (10s, 32kHz)
• Degraded: Downsampled to 8kHz, then restored
• Train/Val split: 80/20
• Optimizer: Adam
• Epochs: 80
• Learning Rate: 5e-5

---

📊 Evaluation Metrics

• SNR (Signal-to-Noise Ratio): measures denoising
• LSD (Log Spectral Distance): frequency-domain fidelity
• CQS (CNN Quality Score): learned perceptual score
• MOS (Mean Opinion Score): subjective test on 10 participants

---

Results

Method	CQS	MOS
MusicGen (baseline)	0.991	0.871
U-Net Enhanced	0.998	0.893

Method	SNR	LSD
U-Net Enhanced	4.757	0.246

---

Authors

• Paul Alexandru Radu Loghin (raduloghin.1942544@di.uniroma1.it)
• Andrea Maggiore (maggiore.1947898@di.uniroma1.it)

-- 📚 This project was developed as part of an academic course at Sapienza Università di Roma.