论文部分内容阅读
提出了在语音信号对数功率谱域和功率谱域顺序滤波的新的增强ＲＡＳＴＡ滤波（ＥＲＡＳＴＡ）方法。语音识别和说话人识别实验表明，ＥＲＡＳＴＡ滤波能够有效地去除加性噪声和卷积噪声的干扰，ＥＲＡＳＴＡ算法与语音信号的失真过程和噪声的功率谱无关。ＥＲＡＳＴＡ方法性能同ＪＲＡＳＴＡ算法类似或更好，且不需要ＪＲＡＳＴＡ 算法中的实时语音信噪比估计。ＥＲＡＳＴＡ 滤波器的设计表明，低频率的谱调制分量可引起语音识别和说话人识别性能的下降，说话人识别较语音识别需要较小的谱时间调制带宽。
A new enhanced RASTA filtering (ERASTA) method is proposed for the sequential filtering of logarithmic power spectral domain and power spectral domain of speech signal. Speech recognition and speaker recognition experiments show that ERASTA filter can effectively remove the interference of additive noise and convolution noise. The ERASTA algorithm has nothing to do with the distortion of speech signal and the power spectrum of noise. The ERASTA method performance is similar or better than the JRASTA algorithm and does not require the estimation of real-time speech signal-to-noise ratio in the JRASTA algorithm. The design of the ERASTA filter shows that low-frequency spectral modulation components can cause degraded speech recognition and speaker recognition performance, and speaker recognition requires less spectral modulation bandwidth than speech recognition.