IEEE Signal Processing Cup 2024 - Team Wavemasters

🚀 Project Overview

This project tackles the challenging task of text-independent far-field speaker recognition under noise and reverberation conditions, specifically for mobile robots. Our goal is to develop a robust speaker recognition system using advanced signal processing techniques and machine learning algorithms.

📝 Project Description

In this project, we focus on creating a system capable of accurately identifying speakers from speech signals captured by mobile robots in challenging acoustic environments. Our approach is composed of the following key components:

• 🎧 Noise Removal and Amplification:
  Leveraging Wavelet Transform for effective signal denoising.

• 📊 Feature Extraction:
  Utilizing Mel-Frequency Cepstral Coefficients (MFCCs) to extract robust audio features.

• 🛠️ Data Augmentation:
  Applying techniques like Gaussian noise addition, time stretching, pitch scaling, and random gain modulation to expand and diversify the training dataset.

• 🤖 Speaker Recognition:
  Implementing the VGGish pre-trained model for high-level semantic embedding extraction, tailored for downstream classification tasks.

🔍 Methodology

1. Wavelet-based Signal Denoising

• Decomposition:
  The audio signal is decomposed into approximation and detail coefficients using Discrete Wavelet Transform (DWT).

• Thresholding:
  Noise is mitigated through soft and hard thresholding techniques applied to the detail coefficients.

• Reconstruction:
  The signal is reconstructed using the modified coefficients, omitting high-frequency details for a cleaner output.

2. Feature Extraction

• Mel-Frequency Cepstral Coefficients (MFCCs):
  We extract MFCCs using the librosa library, capturing essential characteristics of the audio signals in the frequency domain.

3. Data Augmentation

• Diverse Techniques:
  We employ various augmentation techniques such as Gaussian noise addition, time stretching, pitch scaling, and random gain modulation to generate additional training samples, enhancing the model’s generalization capabilities.

4. Model Implementation

• VGGish Pre-trained Model:
  We use VGGish, a model pre-trained on large-scale audio data, to extract 128-dimensional embeddings from denoised audio signals. These embeddings serve as the input for downstream classification tasks.