电话线质量语音的基音周期提取算法

Pitch determination algorithm for telephone speech

基音周期提取算法是语音编码中至关重要的组成部分,针对传统自相关方法在电话线质量语音中提取准确度不高的现状,该文提出了一种新的时频结合的基音周期提取算法。时域上,引进一个新的参数——长时基音周期,并根据语音短时平稳特性,对自相关函数进行时域修正,去除不可能成为基音周期的延时值。频域上,计算频域自相关函数,将基音周期候选值所对应的频域自相关值也作为候选值权重的一部分,以增大真正基音周期的权重。通过Keele语音库进行性能测试表明:该算法对电话线质量语音的严重错误率比传统自相关方法的降低46.8%,极大地提高了电话线质量语音的基音周期判决准确度,同时对正常语音的严重错误率也降低了31.2%。

Pitch determination algorithm is a critical part in speech coding systems. Considering the poor performance of traditional autoeorrelation function （ACF） based algorithm in telephone speech, a new time-frequency based pitch determination algorithm is proposed. In time domain, a new parameter, long-time average pitch （I.TAP） is introduced. Based on both LTAP and short-time stationary property of speech, the time-domain ACF is revised to eliminate the lag values that are impossible to be true pitch. In frequency domain, a frequency-domain based ACF is calculated and incorporated into the weight calculation for each pitch candidate. This processing has the potential to increase the weight of the true pitch. The experiments results on Keele speech database show that the proposed algorithm reduces the gross pitch error rate by 46.8% for telephone speech compared with the traditional ACF based algorithm, thus improves the performance of pitch determination on telephone speech significantly, and it also reduces the gross errorrate by 31.2% for normal speech.