Abstract
This study addresses the actual problem of improving the noise robustness of voice activity detection (VAD) systems. The aim of the work is to conduct a comparative analysis of hybrid neural network architectures, namely CNN-GRU and CNN-BiGRU, to evaluate their effectiveness under conditions of intensive acoustic noise. The experimental dataset was formed based on a Kazakh speech corpus with a total duration of 332 hours. In the course of the study, more than 2.4 million training segments were prepared from 30 speakers, ensuring speaker independence. Mel-frequency cepstral coefficients (MFCCs) with a temporal window of 120 ms were used as input features. The architectures CNN+BiGRU (11,106 parameters) and a lightweight CNN+GRU model (6,050 parameters) were compared. Experimental results across a signal-to-noise ratio (SNR) range from −10 to 30 dB demonstrated high accuracy for both models (up to 99%) under moderate noise conditions. However, under extremely low SNR (−10 dB), the BiGRU architecture outperformed the GRU model by 5–8%. A “spectral specialization” effect was observed, leading to a noticeable decrease in accuracy when testing on clean signals after training in noisy environments. The study confirms that bidirectional layers are more effective in reconstructing speech information by analyzing both forward and backward temporal sequences. It was also established that the CNN+GRU model is more suitable for mobile applications due to its significantly lower computational requirements.