论文部分内容阅读
基于计算听觉场景分析,对基于能量的二值掩蔽语音分离算法的性能进行分析,证明了理想二值掩蔽算法在信噪比下具有最佳的单元分离性能,并通过3种类型带噪语音的分离实验证实了该结论。采用理想二值掩蔽算法对8种噪声类型的低信噪比带噪语音进行了分离实验,信噪比平均提升幅度大于10dB,表明算法对低信噪比语音分离的有效性和普遍适用性;采用非均匀、均匀两种多子带分析滤波器组进行分离性能对比测试,结果表明子带均匀性对信噪比提升影响不大。分析滤波器组的子带数量应大于32以实现较好的分离性能。
Based on the computational auditory scene analysis, the performance of the energy-based binary masking speech separation algorithm is analyzed, and it is proved that the ideal binary masking algorithm has the best element separation performance under the signal-to-noise ratio. Through the three types of noisy speech The separation experiment confirmed this conclusion. The ideal binary masking algorithm is used to separate the low noise and noisy speech of 8 noise types. The average signal to noise ratio (SNR) increase range is more than 10dB, which shows the effectiveness and general applicability of the proposed algorithm for low signal to noise ratio speech separation. The non-uniform and even two multi-subband analysis filter banks are used to test the separation performance. The results show that the subband uniformity has little effect on the signal to noise ratio. The number of subbands in the analysis filter bank should be greater than 32 for better separation performance.