摘要
在语音识别系统中,对识别的准确性有很重要的作用。对于纯净语音信号,传统的端点检测算法能够很好地检测语音部分的起止点。由于在有噪声干扰的情况下,算法的检测准确度往往会急剧下降。为了改善噪声环境下的端点检测效果,从语音信号和噪声信号频域分布特性的差异出发,用频谱方差数值来区分语音和噪声,提出了基于频谱方差的端点检测算法,并进行了无噪声和噪声环境下的仿真,证明了这种算法在强噪声干扰的情况下也能够取得很好的效果。同时将新算法和传统的基于LPCC的端点检测算法进行了对比试验,实验结果表明,在噪声环境下,新算法的检测精度有明显提高。
Endpoint detection is a crucial technology of preprocessing step in speech recognition systems. It plays a very important part in the accuracy of the recognition results. To pure speech signal,traditional endpoint detection methods can also achieve fairly good results in detecting the start and endpoint. But when interfered by noise,the detection accuracy of such methods always falls sharply,or even loses effectiveness. There are differences in frequency domain between speech and noise distribution characters. To improve the performance of endpoint detection in noisy environment,this article takes such discrepancy as a starting point,uses spectrum variance to distinguish voice and noise,and proposes a new anti-noisy endpoint detection method base on spectrum variance. Then simulate the method in both pure speech signal and speech with white noise,prove that this new method can also perform well under the circumstances of serious noise. A contrast test is also done with the new method and a traditional method,based on LPCC. The outcome shows that the new method can achieve better effectiveness.
出处
《计算机仿真》
CSCD
北大核心
2010年第9期337-340,共4页
Computer Simulation
