A spectral envelope estimation method based on F0-adaptive multi-frame integration analysis

This paper presents a novel method of spectral envelope estimation and representation. Despite much sophisticated work in this area, estimating an appropriate envelope is still difficult. We therefore propose an F0-adaptive multi-frame integration analysis method for estimating spectral envelopes with appropriate shape and high temporal resolution. The method does not use pitch marks or phoneme labels and can be used with various types of sound (speech, singing, and instruments). The basic idea is to use F0-adaptive window analysis with a small window length yielding high temporal resolution. The analysis is then extended by using neighboring frames to obtain a stable spectral envelope. In tests using synthesized sound and resynthesized natural sound samples, for 8 of 14 samples the log-spectral distances obtained with the proposed method were smaller than those obtained with well-known previous methods.