卡尔曼滤波语音增强算法的CUDA实现

摘要

提出一种卡尔曼滤波语音增强算法的统一计算设备架构并行实现方案。该方案通过离散余弦变换把含噪语音分解为不相关的 DCT 系数，使原来的时域串行处理转化为统一计算设备架构并行处理，只需较少的线性预测阶数，节省运算时间。实验结果表明：与时域卡尔曼滤波比较，该方法有更高的输出分段信噪比，其统一计算设备架构加速方案可有效缩短语音增强运算的时间。%A Compute Unified Device Architecture (CUDA) parallel implementation scheme for Kalman speech enhancements is proposed. That scheme uses Discrete Cosine Transformation (DCT) to de-correlate the noisy speech frame and divide it into multichannel parallel DCT coefficients. That transforms time domain cascaded Kalman filter processing to CUDA parallel one and needs less linear predictive order to model the speech evolution, so it can save computational time. Experimental results show that it has higher output segment SNR and that this accelerating scheme can reduce operation time.