论文部分内容阅读
针对目前语言辨识系统所采用的特征参数没有充分考虑人耳听觉机制、鲁棒性较差的问题,提出一种符合人耳听觉感知特性的鲁棒语言辨识参数提取算法.该算法主要从两个方面提高特征参数的鲁棒性:在计算各子带能量时采用更符合人耳感知特性的Gammachirp滤波器组代替常用的三角滤波器组;为每一子带通道设计一个补偿滤波器.子带补偿滤波器的设计采用数据驱动的策略,通过补偿使得各子带滤波器输出信号的失真及环境噪音导致的失真同时达到最小.实验表明,文中所提出的特征在常见噪声环境下,性能均优于目前普遍使用的Mel频率倒谱系数特征及其衍生参数.
In view of the fact that the characteristic parameters adopted by the speech recognition system do not fully consider the auditory mechanism of the human ear and have poor robustness, a robust speech recognition parameter extraction algorithm is proposed which is in line with human auditory perception. Improve the robustness of the characteristic parameters: replace the commonly used triangular filter bank with the Gammachirp filter bank which is more in line with the human ear perception when calculating the energy of each sub-band; design a compensation filter for each sub-band channel. The design of the compensation filter adopts data-driven strategy, and the distortion caused by the output signal of each sub-band filter and the noise caused by the environmental noise are minimized simultaneously through the compensation.Experiments show that the proposed features in the common noise environment are superior Mel Frequency Cepstral Coefficients and Their Derivative Parameters.