低信噪比环境下基于PR的音频分割

Composed Speech and Music Sound Separation Based on PR under Low SNR

针对当前不同的非白噪声背景研究很少,根据噪音、语音和音乐的性质并且结合统计学理论,提出一种在不同噪声背景下低信噪比的语音/音乐分割算法。以往的检测算法很少考虑低信噪比的环境,首先从音频数据中提取新的特征参数概率密度比(probability density ratio,PR)和概率密度比过零率(probability density ratio crossing rate,PRCR),特征参数在低信噪比环境下亦能明显表征语音和音乐的不同特性,然后根据音频的特性对PRCR进行修正,再基于此修正的特征参数对语音和音乐进行改变点检测,最后得到分割结果。实验结果显示,在信噪比达到5dB时分割点准确率达到85%以上,具有良好的鲁棒性。

In this paper, a detection algorithm for composed speech and music sound under low SNR noisy environment was adopted. Nevertheless, most of the algorithms proposed before did not consider the audio signals under a low SNR noisy environment, especially under different noise which is not white noise. The algorithm, which is based on the character among noisy, speech and music and combined with the statistical theory, firstly extracted the new characteristic parameters of probability density ratio （PR） and probability density ratio crossing rate （PRCR） from the audio, which can attribute the difference between speech and music even in low SNR, and then modified the PRCR according the property of audio, detected the change - points of speech and music based on these characteristic parameters, eventually the segmentation can be showed from the change - points. The experimental result revealed that the rate of accurate can reach to more than 85% when the SNR equals to 5dB, which shows the advantages of robust.