A study of pitch period extraction based on the multiresolution analysis of speech signals

摘要

This paper focuses on the pitch extraction technique based on the multiresolution analysis of speech signals. The conventional method potentially issues erroneous results for voiced consonants since it does not consider the case in which speech signals simultaneously show both voiced and unvoiced feature. The speech formants can also potentially cause the errors. In order to mitigate these problems, the proposed method employs the LPC inverse filtering and makes voiced/unvoiced decisions by evaluating the periodicity of extracted the local peaks. The proposed method was evaluated as much useful than the conventional method due to the above improvement.