Speech-Signal Separation System (Advisor: Assoc. Prof. Güray Sonugür)

Speech-Signal Separation System (Advisor: Assoc. Prof. Güray Sonugür)

A mechatronics-engineering graduation project on real-time speech separation and filtering using DSP, FFT, artificial neural networks, MFCC and GMM.

QR code for this page

Accuracy in speaker-recognition processes has an important place in digital forensics. This study aimed to increase that accuracy by improving speech recordings in real time.

Two speech-enhancement layers were implemented. The first layer filtered noise in the time-frequency domain with digital signal-processing techniques and the Fast Fourier Transform. The second layer used artificial neural networks to separate speech from other environmental sounds.

The software was implemented in C++ with an object-oriented approach. Tests on sample recordings showed that it could improve the audio with high accuracy.

A subsequent MFCC- and GMM-based speaker-recognition application used the same filter. The system was evaluated with the large-scale VoxCeleb speech dataset. The results indicated improved recognition accuracy for noisy recordings and no loss for clean recordings.