从存在伴奏的歌曲中提取歌声基音的时域算法

提取歌曲中的唱者基音拥有广泛的用途,如可用于基于内容的音频检索等。在歌曲中提取唱者基音存在许多与普通语音处理不同的问题,传统的时域算法在强音乐背景的干扰下很难正确提取唱者人声基音,近年来研究歌声特殊性的各系统则采用较为复杂的训练模型和频域算法。本系统为改进传统时域算法,提高歌声基音提取的准确度,同时为降低算法复杂度,利于用硬件实现基于内容的音频检索系统,提出一种以人声特征着眼、以开环-闭环基音提取为框架的时域算法。实验证明此算法在歌声基音提取上相对传统时域算法准确度有显著提高。

A Time-domain Algorithm of Extraction Singing Voice''''s Pitch with Musical Instrument Accompaniment

Extraction singing voice from music with music instrument accompaniment can be applied in many areas such as content-based music retrieval. Conventional time-domain algorithms perform poor in strong noise circumstance due to its remarkable diversity from ordinary speech processing. Newly-emerging systems dealing with singing voice and background music employ complex training models and realize it in frequency domain. With the consideration of hardware implementation and resource saving, our system proposes a new time-domain algorithm, which is based on human voice feature. The system adopted an open-loop and close-loop framework. Simulations on MATLAB show that accuracy of pitch extraction is superior to conventional time-domain algorithms and some other modified ones.