Analyzing Noise Robustness of Cochleogram and Mel Spectrogram Features in Deep Learning Based Speaker Recogntion

Abstract The performance of speaker recognition is very well in a clean dataset or without mismatch between training and test set. However, the performance is degraded with noise, channel variation, physical and behavioral changes with the speaker. The studies confirmed that the features which represent speech in the Equal Rectangular Band (ERB) scale are more robust than Mel Scale at low Signal to Noise Ratio (SNR) level. Gammatone Frequency Cepstral Coefficient (GFCC) which represents speech in ERB scale is widely used in classical machine learning based speaker recognition at noisy conditions. Recently, deep learning models are widely applied in speaker recognition and show better performance than classical machine learning. Previous deep learning based speaker recognition models used Mel Spectrogram as an input rather than hand crafted features. However, the performance of Mel spectrogram drastically degraded at low SNR level because Mel Spectrogram represents speech in Mel Scale. Cochleogram is another important input to develop deep learning based speaker recognition models. Cochleogram represents speech in ERB scale, which is more robust at low SNR level. However, none of the studies used the Cochleogram feature to develop deep learning based speaker recognition models. In addition, none of the studies conduct analysis of noise robustness of Cochleogram and Mel Spectrogram features in speaker recognition using deep learning. In this study, we conducted an analysis of noise robustness of Cochleogram and Mel Spectrogram features in speaker recognition by using additive noises (such as: babble, street and restaurant noises).The train-clean-100 part of the LibriSpeech dataset, which consists of 251 speakers (126 male and 125 female speakers) and 28,539 utterances is used for the experiment. CNN model is used for training and classification of speakers into different classes. The evaluation results show that Cochleogram is more robust than Mel Spectrogram at low SNR level. Both Cochleogram and Mel Spectrogram features show approximately equal accuracy at high SNR and without additive noise. In conclusion, the Cochleogram feature improves performance of deep learning based speaker recognition in noisy conditions.