Objective Audio Quality Characterization – DRC Discrimination

This project provides a MATLAB implementation for analyzing and discriminating Dynamic Range Compression (DRC) profiles using audio descriptors.

The main objective is to identify features that are:

• ✅ Sensitive to compression profiles
• ✅ Robust to the original audio content

This is achieved using an inverse Fisher criterion, specifically designed to isolate processing effects from content variability.

---

Overview

Audio signals are first processed to extract descriptors.
Then, features are ranked according to their ability to discriminate between different DRC profiles independently of the original signal.

Finally, results can be visualized in a 3D feature space.

---

Project Structure

. ├── start_process_db.m ├── start_Fisher.m ├── start_display3d.m ├── originals/ └── generated/

---

Data Organization

Your dataset must follow this structure: /originals/ file1.wav file2.wav ...

/generated/ file1/ compressed_A.wav compressed_B.wav file2/ compressed_A.wav compressed_B.wav

🔹 Description

• originals/ contains reference (unprocessed) audio files
• generated/<filename>/ contains compressed versions of each corresponding file

⚠️ File names must match between originals/ and generated/.

---

⚙️ Scripts Description

▶️ start_process_db.m

Feature extraction

• Loads original and compressed audio files
• Computes audio descriptors for each signal
• Organizes and stores features for further processing

---

▶️ start_Fisher.m

Feature selection using inverse Fisher criterion

• Computes a score for each feature: F = variance_inter_profile / variance_inter_audio

• Ranks features in descending order

• Selects features that:

• Maximize differences between compression profiles

• Minimize dependence on audio content

---

▶️ start_display3d.m

3D visualization

• Projects selected features into a 3D space
• Allows visual inspection of:
• Separation between DRC profiles
• Robustness across different audio signals

---

Typical Workflow

🔹 Prerequisite: Dataset preparation

Before running the pipeline, you must create a dataset with the following structure:

/originals/ file1.wav file2.wav ...

/generated/ file1/ compressed_A.wav compressed_B.wav file2/ compressed_A.wav compressed_B.wav

• Each file in originals/ must have a corresponding folder in generated/
• Each folder must contain the same audio processed with different DRC profiles

---

1. Extract features

start_process_db

2. Select discriminative features (DRC-oriented)

start_Fisher

3. Visualize feature space

start_display3d