摘要
语音端点检测的检测结果好坏对后续的语音处理起着决定性的作用。为了解决语音端点在低信噪比情况下检测率不高的问题,该文提出了基于深度置信网络去噪的语音增强方法与传统的端点检测方法相结合的方法。该方法首先由大量的语音数据训练深度置信网络模型,使其能够很好地映射带噪与无噪语音之间的非线性关系,进而使其成为一个良好的降噪滤波器,再对比带噪与去噪后语音对端点检测准确率的影响,以及不同信噪比的端点检测的正确率。从该实验结果可以得到,该方法在平稳噪声和非平稳噪声的低信噪比情况下都可以提高语音端点检测的准确率。
The test results of voice activity detection(VAD)play a decisive role in the subsequent speech processing.To resolve the problem of low detection rate of speech endpoints at low signal-to-noise ratio(SNR),a method of combing speech enhancement method based on deep belief network denoising with the traditional endpoint detection method is proposed.The deep belief network model is trained by large volumes of speech data to effectively map the nonlinear relationship between noisy speech and noise-free speech,and is made to become a good noise reduction filter.The effects of noisy speech and denoised speech on endpoint detection accuracy,and the correctness of endpoint detection at different SNRs are compared.The experimental results show that the method can improve the accuracy of VAD in the case of both stationary noise and non-stationary noise with low SNR.
出处
《现代电子技术》
北大核心
2017年第22期1-4,9,共5页
Modern Electronics Technique
基金
国家自然科学基金(61365005
60965002)
