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

This was my Mechatronics Engineering graduation project at Afyon Kocatepe University: a DSP, FFT and Machine-Learning-Based Speech Separation and Filtering System.

The engineering goal was not simply to make noisy recordings sound cleaner to a listener. The processing had to reduce unwanted components while preserving information useful to the downstream speaker-recognition stage.

DSP and FFT Filtering

The first layer used digital signal-processing methods and the Fast Fourier Transform to inspect and filter audio in the time/frequency domain.

Frequency-domain analysis made it possible to observe how speech and environmental noise occupied the spectrum in the tested examples and to apply filtering decisions accordingly.

A key boundary was avoiding excessive suppression of speaker-related spectral information. Audio that sounds subjectively cleaner is not automatically a better biometric feature source.

Neural-Network Separation

The second layer introduced an artificial-neural-network approach to separating speech from environmental sound. This moved part of the decision process from fixed filtering rules toward behavior learned from data.

The implementation belongs to the neural-network methods and C++ environment used at the time; the models and runtimes I used in later years were not part of this project.

C++ Implementation and Real-Time Constraint

The software was implemented in C++ with an object-oriented structure. Acquisition/input, transforms, filtering, separation and evaluation were treated as distinguishable stages.

The real-time requirement meant that correctness alone was not enough: processing also had to keep up with the audio stream.

MFCC/GMM Speaker Recognition

To evaluate the effect beyond waveform inspection, I implemented an MFCC- and GMM-based speaker-recognition stage.

The resulting pipeline can be summarized as:

audio -> enhancement/separation -> MFCC -> GMM -> speaker decision

This made it possible to compare recognition behavior with and without the preprocessing stage.

VoxCeleb Evaluation

The historical project record states that the system was evaluated with VoxCeleb speech data and that noisy recordings showed improved speaker-recognition behavior without observed loss on clean examples.

The surviving project record establishes that filtered and unfiltered speaker-recognition pipelines were compared on VoxCeleb speech data. Because the complete split, noise model and numerical metric table are unavailable, the supported result is limited to that comparison.

Connection to Later Work

The project became an early engineering basis for my later speech, speaker-recognition and real-time audio systems. Later systems use different models and runtimes and treat VAD, speaker embeddings, segmentation and capacity engineering as separate system concerns.

The graduation project was supervised by Assoc. Prof. Güray Sonugür.