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

The performance of speaker recognition systems is very well on the datasets without noise and mismatch. However, gets degraded with environmental noises, channel variation, physical behavioral changes in speaker. types Speaker related feature play crucial role improving systems. Gammatone Frequency Cepstral Coefficient (GFCC) features has been widely used to develop robust conventional machine learning, it achieved better compared Mel (MFCC) noisy condition. Recently, deep learning models showed learning. Most previous learning-based Spectrogram similar inputs rather than a handcrafted like MFCC GFCC features. high ratio mismatch utterances. Similar Spectrogram, Cochleogram another important for models. Like features, represents utterances Equal Rectangular Band (ERB) scale which none studies have conducted analysis robustness recognition. In addition, only limited speech-based condition using this study, model at Signal Noise Ratio (SNR) level from −5 dB 20 dB. Experiments are VoxCeleb1 added dataset by basic 2DCNN, ResNet-50, VGG-16, ECAPA-TDNN TitaNet Models architectures. identification verification both evaluated. results show that