🎙️ Speaker Verification using Gaussian Mixture Models (GMM)

This repository implements a Speaker Verification System using Gaussian Mixture Models (GMMs) with MFCC, Delta, and Double Delta features. The system is trained and tested on a subset of a speaker recognition dataset containing 1-second WAV files per speaker.

---

📌 Project Overview

Speaker verification is an essential component in audio-based security and authentication systems. This project uses GMMs to learn speaker-specific acoustic patterns and verify the identity of a speaker based on short utterances.

🔍 Key Highlights

• 📂 Preprocessing includes noise addition, resampling, and pre-emphasis.
• 🎚️ Feature extraction based on MFCC, delta, and double-delta coefficients.
• 🎯 Model training using Gaussian Mixture Models with Expectation-Maximization.
• 📊 Equal Error Rate (EER) used for evaluation; best model achieves EER = 0.177.
• 👥 Includes speaker pair comparisons and real-world applicability demonstrations.

---

🛠️ Project Structure

.
├── Audio/                     # Contains 1-second audio files for each speaker
├── Noise/                     # Contains noise samples used for robustness
├── test_pairs.txt             # File containing speaker comparison test cases
├── ml-end.ipynb               # Main notebook for training, evaluation, and testing
├── README.md                  # You're reading it

---

🔄 Pipeline Summary

1. Preprocessing

• Silence and Noise Removal
• Resampling to 16kHz
• Pre-Emphasis Filtering

2. Feature Extraction

• MFCC (Mel-Frequency Cepstral Coefficients)
• MFCC Delta
• MFCC Double Delta

3. Model Training

• Trained a GMM per speaker using sklearn.mixture.GaussianMixture
• Parameters: n_components=4, max_iter=160

4. Evaluation

• Log-likelihood comparison for speaker prediction
• Equal Error Rate (EER) calculated for performance
• Also supports speaker pair comparison

---

📊 Results

Model Version	EER
MFCC + Delta + Double Delta	0.177
MFCC only	Higher

A lower EER indicates better performance (balance between false accept and false reject).

---

📚 References

Key papers referenced include:

• Reynolds & Rose (1995): Robust text-independent speaker identification using GMM
• Dehak et al. (2007): Modeling Prosodic Features with Joint Factor Analysis
• Jadhav et al. (2018): GMM + MFCC + EM-based speaker recognition

See full reference list in the report

---

📈 Future Work

• Train on larger datasets with more speakers
• Explore i-vector and x-vector embeddings
• Apply deep learning approaches (e.g., LSTM, CNN) for feature extraction
• Integrate with real-time APIs or voice assistants

---