psola算法c语言实现代码,【图片】重新理解PSOLA算法【c语言吧】_百度贴吧

该楼层疑似违规已被系统折叠 隐藏此楼查看此楼

1. Introduction

PSOLA (Pitch Synchronous OverLap and Add)[1], an algorithm able to change the duration or shift the pitch of speech signals, invented around 1990 (arguably) by France Telecom, remains being the most widely used speech modification algorithm as of 2015. Staying in time-domain from the beginning to its end, this incredibly simple algorithm outputs speech at pretty good quality and blazingly fast speed. Given such a pleasant compromise between complexity, speed and quality, no wonder those top-tier softwares (e.g. Melodyne and Auto-Tune) in music production industry are using PSOLA as their `secret' weapon to conquer the market; no wonder those top-tier universities all over the world more or less put PSOLA in their speech processing courses.

What leaves me wondering is: seems like there's always a blind spot in all tutorials, slides and papers about PSOLA. In a few sentences they tell you something like,

> PSOLA basically marks the "epochs" (some call them Glottal Closure Instants or Instants of Significant Excitation), puts a bunch of 2-period-long hanning windows on these epochs, cuts the signal into pieces, and finally adds the pieces back together (at some different epoch positions determined by pitch and time stretch).

Now you can clutch onto your favourite editor, implement the whole thing in a few hundred lines. They've taught you "how does it work", but they hide the more intriguing fact from you - "what does this mean?", or they themselves either don't know what PSOLA does in frequency domain.

The goal of this quite informal article is to fill in the blank, i.e., to gain a deep understanding of PSOLA, discover a few interesting properties, and see how to ultimately improve the algorithm using these properties.