A Comparison between Mel-Scale Cepstrum and Auditory Model Representation for Noisy Speech Recognition

A joint synchrony/mean-rate auditory model, recently proposed by Seneff[6], is embedded into a classical DTW-based system for the recognition of Italian digits. Its performances are evalu¬ated in both clean and noisy speech and compared with those of a system based on the Mel¬cepstrum representation. Experimental results show that the Mel representation outperforms the auditory model. Problems encountered by the auditory model in noisy speech are outlined and suggestions for noise compensation techniques both inside and outside the model are given. Simple image processing techniques aiming to clean up the synchrony spectrogram in noisy speech are suggested and some promising preliminary results are presented.