Noisy Phoneme Recognition Using 2D Convolution Neural Network

Speech is one of the most important parts of everyday life, thus it has been investigated from various standpoints, however, there is still room for exploration within noisy speech signals. This study examines how speech signals are recognized in the presence of noise by conducting a recognition process using both clean speech and speech data with additive noise. Spectrograms and Mel Spectrograms have been extracted and tested using a Convolutional Neural Network. Training on noise-free data and on mixed data which has been composed of clean and noisy phoneme signals has been considered. The experimental results showed that model trained with set which includes noisy samples gives better results when classifying signals with noise present compared to noise-free trained model. It was also revealed that Mel Spectrograms represent noisy signals better than Spectrograms.