tolysz/speaker-identification

An attempted speaker recognition

Speaker Recognition

Speaker recognition is the task of recognizing a person automatically from his or her voice; we assume there is a set of known voices recorded prior to the recognition task. In addtion, we don't address the verification problem here.

For detailed discussion about the subject, please refer to the paper Automatic Speaker Recognition Using Gaussian Mixture Speaker Models.

The main technique we are using are MFCC and GMM.

MFCC

Mel-frequency cepstrum coefficients are common features that are used for voice/speech related tasks. It's a representation of the power spectrum, but based on a nonlinear mel scale.

There is an educational MFCC implementation (not focusing on , please check out the mfcc branch. The master branch uses BLAS to speed up feature extraction.

GMM

For text-independent speaker recognition, the sound source (human voice) can be modelled as a mixture models, each following the Gaussian distribution. Typically, 32-component GMM should be sufficent to distinguish a few number of speakers.

Code Structure

This repository mainly contains the C++ implementation. However, during the research, it was easier to use Matlab to do data and algorithm exploration. You can check out the matlab folder and specifically the script speaker_recognition.m. It currently uses data that's exported from GRT. It wouldn't be too hard to simply use audio files (you may check out the MOCHA-TIMIT dataset).

Dataset

1. TIMIT

Doesn't seem to be free.

2. MOCHA-TIMIT

Although small, two speakers but can be used for initial testing.