Signal resynthesis from auditory Mellin image using a high-quality VOCODER, STRAIGHT

摘要

We propose a method for resynthesizing sounds from auditory representations, Auditory Mellin Images, by using a high-quality VOCODER, STRAIGHT. Analysis/synthesis systems for speech sounds have been studied extensively until the VOCODER system was developed in 1939. There is, however, no system involving a reallistic auditiory model while human sound perception is known as an important facter to develop the system. We combined Auditory Mellin Image model and STRAIGHT into a new "auditory" VOCODER system by introducing a mapping function including frequency-warping Discrete Cosine Transform and nonlinear multivariate analysis. By using this system, we expect to include auditory functions such as noise-robustness and sound-source separation which have been problems for conventional VOCODERs.