Performance enhancement of text-independent speaker recognition in noisy and reverberation conditions using Radon transform with deep learning

Automatic Speaker Recognition (ASR) in mismatched conditions is a challenging task, since robust feature extraction and classification techniques are required. Long Short-Term Memory Recurrent Neural Network (LSTM-RNN) is an efficient network that can learn to recognize speakers, text-independently, when the recording circumstances are similar. Unfortunately, when the recording circumstances differ, its performance degrades. In this paper, Radon projection of the spectrograms of speech signals is implemented to get the features, since Radon Transform (RT) has less sensitivity to noise and reverberation conditions. The Radon projection is implemented on the spectrograms of speech signals, and then 2-D Discrete Cosine Transform (DCT) is computed. This technique improves the system recognition accuracy, text-independently with less sensitivity to noise and reverberation effects. The ASR system performance with the proposed features is compared to that of the system that depends on Mel Frequency Cepstral Coefficients (MFCCs) and spectrum features. For noisy utterances at 25 dB, the recognition rate with the proposed feature reaches 80%, while it is 27% and 28% with MFCCs and spectrum, respectively. For reverberant speech, the recognition rate reaches 80.67% with the proposed features, while it reaches 54% and 62.67% with the MFCCs and spectrum, respectively.