基于时域和时频域联合优化的语音增强算法

Speech Enhancement Algorithm Based on Time-Domain and Time-Frequency Joint Optimization

基于深度学习的语音增强算法从带噪语音信号的时域或者时频域中恢复出干净的语音信号。然而,时域和时频域的增强算法都有自己的优点和不足。针对这一问题,本文提出了一种基于时域和时频域联合优化的语音增强算法。在生成对抗网络框架下,分别构建了时域和时频域的学习目标。在训练过程中,利用跳跃连接搭建了深层次的网络结构,通过对不同领域学习目标的联合优化,获得了语音增强性能的改善。实验结果表明:相比基线模型,本文提出的算法在多个客观评价指标上都具有更好的表现。

Deep learning-based speech enhancement algorithm recovers clean speech signals from the time domain or time-frequency domain of noisy speech signals.However,both time-domain and time-frequency-domain methods have their own advantages and disadvantages.To solve this problem,this paper proposes a speech enhancement algorithm based on joint optimization of time domain and time-frequency domain.In the framework of generative adversarial network,the learning objectives in time domain and time-frequency domain are constructed respectively.In the training process,the skip connection is used to build a deep network structure,and the improvement of speech enhancement performance is obtained through joint optimization of learning goals in different fields.The experimental results show that:compared with the baseline models,the proposed algorithm has better performance on multiple objective evaluation metrics.