谱跟踪和噪声模型语音信号分析/合成方法

摘要

把语音信号分解为确定和随机两部分是非常典型的语音信号分析与合成方法。针对目前主流研究的谐波加噪声模型( Harmonic plus Noise Model, HNM)和准谐波模型( Quasi-Harmonic Model, QHM)法等存在冗余大的问题,提出了一种基于谱跟踪和噪声模型的语音信号分析/合成方法。以音素为基本单元进行编解码,用谱跟踪法提取时变特征合成语音信号的确定部分,用AR模型合成语音信号的随机部分。实验发现,合成语音与原语音波形上比较接近,MOS分约为3.24,冗余也明显减少。%Decomposition of speech signal into deterministic part and stochastic part is a typical method for speech analysis and synthesis. Aiming at the redundancy of current mainstream methods, such as HNM ( Harmonic plus Noise Model ) and QHM ( Quasi-Harmonic Model ) , a novel speech analysis/synthesis method based on spectral tracking and noise model is proposed. The basic coding units are phonemes, and the deterministic part of speech signal is synthesized via time-varying characteristics extracted by spectral tracking method, and the stochastic part of speech signal is synthesized by AR model. Experiments show that the reconstructed signal is quite close to the original, and the MOS is 3. 24, with obvious reduction of redundancy.