基于离散余弦变换与Transformer的语音增强

Speech enhancement based on discrete cosine transform and Transformer

为更有效提取低信噪比下含噪语音中的语音信息,受到基于多头自注意力机制的两阶段Transformer的启发,提出结合离散余弦变换与Transformer的语音增强算法。语音信号经离散余弦变换后的频谱为实数谱,将其作为深度神经网络的输入时,无需考虑实部、虚部的计算问题。使用两阶段Transformer提取局部和全局特征,在此基础上,通过编解码器网络得到增强后的语音。实验结果表明,所提方法与各参考方法相比,在-9到5的信噪比下,PESQ指标平均提升了9.92%,STOI指标平均提升了10.19%,在较低信噪比下具有良好的降噪能力。

To more effectively extract speech information from noisy speech under low signal-to-noise ratio,inspired by the two-stage Transformer based on the multi-head self-attention mechanism,a speech enhancement algorithm combining discrete cosine transform and Transformer was proposed.The frequency spectrum of the speech signal after the discrete cosine transform was a real number spectrum,so when it was used as the input of a deep neural network,there was no need to consider the calculation of the real and imaginary parts.A two-stage Transformer was used to extract local and global features.On this basis,the enhanced voice was obtained through the codec network.Results of the simulation indicate that compared with the reference methods,the proposed method has an average increase of 9.92%in PESQ index and an average increase of 10.19%in STOI index at a signal-to-noise ratio of-9 to 5,and has good noise reduction ability under low signal-to-noise ratio.