Construction of computational theories of audition and the application to sound signal processing

点击翻译按钮获取中文摘要

The purpose of this project was to construct computational theories of audition and to apply them to various sound signal processing techniques. There were four sub-goals : 1) The development of reliable auditory filterbanks ; 2) The experimental verification of a computational theory of auditory size-shape perception ; 3) The development of speech processing methods using vocoders ; and 4) Examining various speech applications and surveying their principles.1)We determined a reasonable parameter set of compressive gammachirp (cGC) auditory filters by quantitatively fitting them to a thousand of human masking data. A dynamic version of the cGC filterbank (dcGC) was designed to account for the nonlinear characteristics of the cochlea within an analysis/synthesis framework, which is essential for various applications. The dcGC software will be available to the public on a web page as a de facto standard for auditory peripheral processing.2)We performed perceptual experiments supporting t … More he computational theory of auditory size-shape perception. The experiments would be a milestone for the fixture studies in this new field. We organized a special session entitled "Size information in speech and animal calls" at the spring meeting of the Acoustical Society of America (ASA) in May 2005. We invited twelve speakers (seven from our group) to present the related works.3)We improved an auditory vocoder for speech segregation and evaluated it with both objective and subjective measurements. We developed accurate and robust fundamental frequency methods that are important for this system. We improved the high quality analysis/synthesis system, STRAIGHT, applied it to auditory morphing, and developed a real-time version.4)We developed several speech/acoustic processing methods : a speech recognition model using phoneme hypotheses ; a multi-microphone speech enhancement system ; a new robust acoustic measurement signal called "warped-TSP" ; and methods to deal with emotional speech signals and unexpected noise in a speech dialog system. We began using a statistical method to identify the neural circuits for directional sensitivity. Less