摘要
利用FFT-ACF算法进行基音周期候选值估计,减少在语音基音周期提取中常见的倍频和半频错误,提出针对候选值的多重后处理算法.后处理过程:首先运用峰值筛选法进行初选,接着利用一次均值法将语音分为不同的音高段,再使用二次均值法为每个音高段确定合适的频率范围,最后精确提取出基音周期.实验结果表明,基音周期后处理算法有效,在音乐哼唱识别应用中收到良好效果.
To reduce the halving and doubling errors in pitch tracking, the FFT-ACF algorithm was used to estimate the candidates of pitch, and a new multi-post-processing algorithm to process the candidates of pitch was proposed. Firstly, the peek-selecting method was used to detect the right candidates of pitch. Secondly, the first-mean method was used to divide the singing speech into different pitch segment. Thirdly, the second-mean method was used to obtain the optimal frequency range for every pitch segment. As a result, the precise pitch is determined from speech signal. Experiments show that the proposed multi-post-processing algorithm outperforms other algorithms, demonstrating desirable performance in query by singing system.
出处
《深圳大学学报（理工版）》
EI
CAS
北大核心
2007年第4期388-392,共5页
Journal of Shenzhen University(Science and Engineering)
基金
深圳市科技计划资助项目(QK200601)
