Signal Separation Motivated by Human Auditory Perception: Applications to Automatic Speech Recognition
Richard M. Stern · Kluwer Academic Publishers eBooks · 2006
It is believed that computational auditory approaches are potentially extremely useful in ameliorating some of the most difficult speech recognition problems, specifically the recognition of speech presented at low SNRs, speech masked by other speech, speech masked by music, and speech in highly reverberant environments. The solution to these problems using CASA techniques is likely to depend on the ability to develop several key elements of signal processing, including the reliable detection of fundamental frequency for isolated speech and for multiple simultaneously-presented speech sounds, the reliable detection of modulations of amplitude and frequency in very narrowband channels, and the development of across-frequency correlation approaches that can identify frequency bands with coherent microactivity as they evolve over time. I am extremely optimistic that effective solutions for these problems are within reach in the near future.