A comprehensive speaker verification system using statistical methods and identity vectors, developed as part of my Bachelor's degree project. This system implements custom algorithms for speaker verification on a Romanian speakers corpus, resulting in a published IEEE paper.
📄 IEEE Publication: Speaker Verification Experiments using Identity Vectors, on a Romanian Speakers Corpus
This project has been peer-reviewed and published in IEEE Xplore, demonstrating its academic and technical merit.
This project implements a speaker verification system using statistical methods and identity vectors, specifically designed for Romanian speakers. Unlike black-box deep learning approaches, this system builds all core algorithms from scratch, providing full control over the mathematical foundations and feature extraction processes.
🔬 Academic Context: Developed as a Bachelor's degree thesis project focusing on statistical signal processing, identity vector computation, and speaker verification algorithms - all implemented from first principles.
- Custom Statistical Algorithms: All core functions implemented from scratch
- Identity Vector Computation: Statistical approach to speaker modeling
- Romanian Corpus Optimization: Specifically tuned for Romanian language characteristics
- Mathematical Foundation: Complete control over underlying mathematics
- Verification System: Speaker verification (1:1 matching) implementation
- Forensic Applications: Suitable for legal and security applications
- Performance Analytics: Comprehensive statistical evaluation metrics
- Python 3.x - Primary programming language
- NumPy - Numerical computations and matrix operations
- SciPy - Statistical functions and signal processing
- Matplotlib - Visualization and analysis plots
- Wave/Audio Libraries - Basic audio file handling
- Statistical Feature Extraction - Hand-coded feature computation
- Identity Vector Algorithms - Custom statistical modeling
- Distance Metrics - Implemented similarity measures
- Verification Algorithms - Custom speaker verification logic
- Performance Evaluation - Statistical significance testing
- Statistical Modeling - Gaussian distributions, covariance matrices
- Linear Algebra - Matrix operations, eigenvalue decomposition
- Signal Processing - Spectral analysis, windowing functions
- Probability Theory - Likelihood estimation
Background Speakers
│
▼
[Speaker Speech Detection]
│
▼
[Feature Extraction]
│
├───► [Train GMM-UBM Model]
│ │
▼ ▼
[Statistics Calculation]
│
├───► [Train Total Variability Space (T-Matrix)]
│ │
▼ ▼
[i-vector Extraction]
│
├───► [Compute Projection Matrix]
│ │
│ │
▼ ▼
[Train GPLDA Speaker Model]
Outputs:
- UBM Model
- T-Matrix
- Projection Matrix
- GPLDA Model
Known Speaker
│
▼
[Speaker Speech Detection]
│
▼
[Feature Extraction]
│
▼
[Statistics Calculation] ◄────────── Uses UBM Model
│
▼
[i-vector Extraction] ◄────────── Uses T-Matrix
│
▼
[Session Compensation] ◄──────── Uses Projection Matrix
Known Speaker
│
▼
[Speaker Speech Detection]
│
▼
[Feature Extraction]
│
▼
[Statistics Calculation] ◄────────── Uses UBM Model
│
▼
[i-vector Extraction] ◄────────── Uses T-Matrix
│
▼
[Session Compensation] ◄──────── Uses Projection Matrix
│
▼
[Score Calculation] ◄─────────── Uses GPLDA Model
│
▼
[Decision] ◄──── Threshold
- Audio Preprocessing Module: Handles audio loading, noise reduction, and normalization
- Feature Extraction: Extracts audio features (MFCC, spectral and temporal features)
- Statistical Modeling: Identity vector computation using statistical methods
- Verification: Distance-based speaker verification
- Evaluation Module: Assesses model performance with various metrics
The system uses a statistical approach to create identity vectors that capture speaker-specific characteristics:
- Statistical Feature Modeling: Custom algorithms for feature distribution analysis
- Dimensionality Reduction: Mathematical techniques for efficient representation
- Speaker Modeling: Identity vectors as compact speaker representations
- Verification Metrics: Statistical distance measures for speaker comparison
- Language-Specific Tuning: Optimized for Romanian phonetic characteristics
- Corpus Analysis: Statistical analysis of Romanian speech patterns
- Cultural Adaptation: Accounting for Romanian accent variations
- Forensic Applications: Suitable for legal proceedings in Romanian courts
This project uses the “RoDigits – a Romanian connected-digits speech corpus for automatic speech and speaker recognition” for speaker recognition training and evaluation.
@article{georgescu2018rodigits,
title={Rodigits-a romanian connected-digits speech corpus for automatic speech and speaker recognition},
author={Georgescu, Alexandru Lucian and Caranica, Alexandru and Cucu, Horia and Burileanu, Corneliu},
journal={University Politehnica of Bucharest Scientific Bulletin, Series C},
volume={80},
number={3},
pages={45--62},
year={2018}
}There is a demo version of the application in the demo directory that allows enrolling a new speaker in the system, as well as verification. For creating models, all methods are provided in the code directory.
- Spectral Features: Hand-implemented FFT-based spectral analysis
- Cepstral Analysis: Custom cepstral coefficient computation
- Temporal Features: Statistical temporal pattern analysis
- Normalization: Custom normalization techniques for Romanian speech
- Statistical Modeling: Gaussian mixture parameter estimation
- Dimensionality Reduction: Principal component analysis implementation
- Vector Quantization: Custom clustering algorithms
- Model Adaptation: Speaker-specific model adaptation techniques
- Novel Statistical Approach: Custom implementation of identity vector methods
- Romanian Language Focus: Specialized optimization for Romanian speakers
- Mathematical Rigor: Complete mathematical foundation and proof of concepts
- Forensic Applications: Practical applications in legal contexts
- Advanced Statistics: Deep understanding of statistical modeling
- Signal Processing: From-scratch implementation of audio processing
- Algorithm Design: Custom algorithm development and optimization
- Research Methodology: Scientific approach to problem-solving
- Academic Writing: Peer-reviewed publication skills
O. -M. Novac, S. -A. Toma and E. Bureaca, "Speaker Verification Experiments using Identity Vectors, on a Romanian Speakers Corpus," 2021 International Conference on Speech Technology and Human-Computer Dialogue (SpeD), Bucharest, Romania, 2021, pp. 162-166, doi: 10.1109/SpeD53181.2021.9587396. keywords: {Training;Forensics;Speaker recognition;Testing},
@INPROCEEDINGS{9587396,
author={Novac, Oana-Mariana and Toma, Stefan-Adrian and Bureaca, Emil},
booktitle={2021 International Conference on Speech Technology and Human-Computer Dialogue (SpeD)},
title={Speaker Verification Experiments using Identity Vectors, on a Romanian Speakers Corpus},
year={2021},
volume={},
number={},
pages={162-166},
keywords={Training;Forensics;Speaker recognition;Testing},
doi={10.1109/SpeD53181.2021.9587396}}This project is licensed under the MIT License - see the LICENSE file for details.
Academic use encouraged with proper citation of the IEEE publication.
⭐ If you find this research useful, please star the repository and cite our IEEE paper!
📄 IEEE Publication: This work has been peer-reviewed and published in IEEE Xplore Digital Library, demonstrating its academic rigor and technical merit.