Signal Separation Motivated by Human Auditory Perception: Applications to Automatic Speech Recognition

The human auditory system uses a number of well-identified cues to segregate and separate individual sound sources in a complex acoustical environment. For example, researchers in auditory scene analysis have long identified cues such as common onset, correlated fluctuations in instantaneous amplitude and frequency, harmonicity, and common interaural time and amplitude differences as ways of identifying which components of a complex signal are derived from a common source. It is widely believed that the use of these cues to achieve such “grouping” and signal separation should be very useful in improving the accuracy of automatic speech recognition in very difficult environments such as competing speech, background music, and transient noise, and this has been a goal of several research groups in computational auditory scene analysis. This talk describes and discusses several ways in which signals can be separated using physiologically-motivated cues, along with the potential benefit to be derived from such separation for automatic speech recognition.