基于听觉模型的小波包变换的语音增强

由于人耳频率分辨率是非线性的 ,用传统的线性信号处理方法 (如FFT)来模拟人耳基底膜的频率分析特性是比较困难的 .小波包算法有灵活的时频分析能力 ,可较好地符合人耳基底膜的频率分析特性 .在模拟人耳的听觉机理方面 ,用动态阈值法成功地对含噪语音进行了去噪处理 ,在去噪处理中引入音乐噪声的问题也较好地得到解决 .实验表明 :在单声道的条件下 ,其语音增强效果比传统的频谱减法有更高的清晰度和可懂度

Speech Enhancement Using Wavelet Packet Transform Based on Auditory Model

Because of the non?linear characteristics of human auditory,the auditory frequency bands cannot be approached by the conventional signal analysis methods such as FFT,etc.This paper proposed a new approaching method using wavelet packet transform. Wavelet packet transform has flexible frequency bands,so it is more compatible to simulate the human auditory model.The proposed method decompresses speech signal to 52 wavelet packet bands according to the auditory bandwidth,three wavelet bands for each critical band. The dynamic?threshold denoising method is also based on the speech enhancement mechanism of human auditory perception system. It can decrease the threshold when speech is detected;increase it when noise is detected,just like what the perception systems do.This enhancement method has better SNR improvement than other single?channel speech enhancement methods.It also can significantly reduce the unnatural structure of residual noise ("musical tones") even at very low signal?to?noise rations (SNR's).The presented experimental results show that,using this method,better clearness and higher intelligibility of speech can be achieved than using spectral subtraction processing.