A psychoacoustically-motivated conceptual model for automatic speech recognition

摘要

This paper describes a conceptual model of human auditory perception processes for automatic speech recognition (ASR) systems. Even in complex noisy environments, humans can perceptually segregate the target sound that they apply attention to from an acoustic mixture. This paper names the primary advantage of human auditory perceptual process 慳ttentive ear.?We propose a psychoacoustically-motivated conceptual model of attentive ear, and introduce an ASR system as an application of our proposed model. Although psychoacoustically-motivated conceptual models based on human auditory processes for ASR system were proposed in decade, these are not as robust as the ASR systems using traditional noise reduction methods and adaptation methods based on statistical methodology. This is because properly, conceptual models based on human auditory processes were used as a preprocessor, connecting with a statistical recognizer in traditional ASR systems though a psychoacoustically-motivated conceptual model and a model based on statistical methodology in the ASR systems are di.erent theoretically. The ASR system based on our model is not a preprocessor for ASR system but an ASR system itself. To evaluate the ASR system based on our model, we carried out Japanese digit recognition experiments in six typical noisy environments. Results showed that our ASR system is more robust than traditional ones in experimental conditions of 0 dB SNR. These results suggest that our psychoacoustically-motivated conceptual model based on human auditory perceptual process, attentive ear, is e.ective for ASR, and robust in adverse noisy environment.