🎙️ Linguistix: Speaker Recognition System

CSL2050 - Pattern Recognition and Machine Learning Project
Indian Institute of Technology, Jodhpur

👥 Team Members

• Shashank Parchure (B23CM1059)
• Atharva Honparkhe (B23EE1006)
• Vyankatesh Deshpande (B23CS1079)
• Abhinash Roy (B23CS1003)
• Namya Dhingra (B23CS1040)
• Damarasingu Akshaya Sree (B23EE1085)

📌 Abstract

Linguistix is a speaker recognition system that applies classical ML techniques to identify speakers from voice samples. It explores a wide range of models including KNN, SVM, Decision Trees, ANN, Naïve Bayes, GMMs, and K-Means. Dimensionality reduction (PCA, LDA) and ensemble learning methods (Bagging, Boosting, Stacking) are integrated for performance optimization.

🧠 Key Insight: Supervised dimensionality reduction (especially LDA) significantly boosts the accuracy and generalization of traditional ML models in speaker recognition.

---

📂 Dataset

• Source: Kaggle Dataset
• Samples: 2511 .wav files
• Speakers: 50 unique identities
• Feature Extraction: MFCCs → 4000-dimensional vectors per sample

---

🛠️ Tech Stack

• Language: Python 3.10
• Libraries: librosa, numpy, scikit-learn, pandas, PyTorch, matplotlib
• Environment: Jupyter Notebooks
• Cloud Deployment: Google Cloud VM (4GB RAM, 25GB Disk)

---

🧪 Models Implemented

🔹 Supervised Classifiers

• K-Nearest Neighbors (KNN)
• Support Vector Machines (SVM)
• Decision Trees
• Naïve Bayes
• Multi-Layer Perceptrons (MLP)
• Artificial Neural Networks (ANN)

🔹 Clustering Techniques

• K-Means
• Gaussian Mixture Models (GMM)

🔹 Ensembles

• Bagging
• AdaBoost & Improved AdaBoost (SAMME)
• Stacking (SVM + GMM + RF)

🔹 Dimensionality Reduction

• PCA
• LDA
• t-SNE & UMAP (for visualization)

---

📊 Key Results (Test Accuracy)

Model	Test Accuracy (%)
ANN with LDA	100.00
KNN with LDA	99.80
Bayesian with LDA	99.80
SVM with PCA	99.40
GMM (Semi-supervised)	99.67
CNN with LDA + PCA	99.80
Decision Tree + Bagging	82.70
K-Means with LDA + PCA	87.67

---

📌 Key Observations

• LDA > PCA: LDA consistently performed better across models due to class-separability.
• GMM + Supervision: Semi-supervised GMMs delivered the best clustering results.
• ANN + LDA: Delivered perfect classification across all splits.
• Ensemble Methods: Bagging and AdaBoost helped reduce overfitting in tree-based models.

---

🌐 Live Demo & Resources

• 🔗 GitHub Repository
• 🎥 Spotlight Video
• 💻 Web Demo (Hosted on Google Cloud)
• 📄 Project Page

---

🤝 Contributions

Shashank Parchure (B23CM1059)

• Implemented:
  • KNN with PCA
  • Bayesian Learning with Correlation-based Feature Selection
  • SVM with PCA
  • MLP with PCA
  • ANN with LDA
• Responsible for:
  • Report compilation
  • Spotlight video organization (content outline)
  • Deploying demo code on Google Cloud

Atharva Honparkhe (B23EE1006)

• Implemented:
  • Decision Tree with PCA, LDA, and Ensemble Methods
  • GMM
• Responsible for:
  • Demo code creation
  • Report compilation

Vyankatesh Deshpande (B23CS1079)

• Implemented:
  • KNN with LDA
  • Bayesian Learning with LDA
  • SVM with Correlation-based Feature Selection
  • ANN with PCA
• Responsible for:
  • MFCC feature extraction
  • Project page implementation
  • Report compilation
  • Google Cloud exploration

Abhinash Roy (B23CS1003)

• Implemented:
  • KMeans Clustering with LDA
  • Decision Tree on raw data
  • CNN with PCA and LDA
• Responsible for:
  • Spotlight video creation

Namya Dhingra (B23CS1040)

• Implemented:
  • Decision Tree with UMAP and t-SNE
  • KMeans Clustering with PCA
• Responsible for:
  • Spotlight video presentation organization
  • Content writing

Damarasingu Akshaya Sree (B23EE1085)

• Implemented:
  • KNN
  • Bayesian Learning with PCA
  • SVM with LDA
  • MLP with Correlation-based Feature Selection
  • ANN with Correlation-based Feature Selection
• Responsible for:
  • Spotlight video presentation organization
  • Content writing

---

📚 References

• NumPy Documentation
• Scikit-learn Documentation
• PyTorch Documentation
• Understanding Bootstrapping – Medium
• Ensemble Methods – Medium