摘要
为了克服低信噪比输入下,语音增强造成清音弱分量损失,导致信号重构失真的问题,提出了一种新的语音增强方法。该方法采用小波包拟合语音感知模型的临界带,按子带能量对语音清浊音分离,然后对清音和浊音信号分别作8层和4层小波包分解,在阈值计算上采用Bark子带小波包自适应节点阈值算法,在Bark子带实时跟踪噪声水平,有效保护清音中高频弱分量,减少失真。通过与传统语音增强方法的仿真对比实验,证实该方法在低信噪比输入时,具有明显优势,输出信噪比高,语音失真度低。将该方法与谱减法相结合,进行语音二次增强,能进一步提比输入时,具有明显优势,输高增强语音质量。
When input signal has low Signal-to-Noise Ratio (SNR), the commonly used speech de-noising algorithm will cause distortion for reconstructed signal because of unvoiced sounds weak information losses. In order to overcome this, this paper presented a new method for speech enhancement. Wavelet packet decomposition was used to fit speech critical band, and the voiced and unvoiced sounds were processed separately based on sub-band energy ratio. Then, eight scales of wavelet packet decomposition and four scales of wavelet packet decomposition were employed for the unvoiced and the voiced sounds. A new wavelet adaptive threshold algorithm was obtained based on Bark sub-band, in Bark frequency domain real-time tracking noise level and the adaptive adjustment of coefficient can increase the accuracy of threshold value judgment, and effectively reduces signal reconstruction distortion. The computer simulation results indicate that the new method compared to traditional algorithm has obvious advantages in improving output SNR and effectively reducing the speech distortion. When this new algorithm is combined with spectral subtraction, it can further improve the quality of speech de-noising.
出处
《计算机应用》
CSCD
北大核心
2010年第11期3111-3114,共4页
journal of Computer Applications
关键词
小波包
听觉掩蔽
语音增强
清音分离
自适应阈值
wavelet packet
hearing masking
speech enhancement
separation of unvoiced sound
adaptive threshold