Robust noise reduction by novel means of incorporating phase processing

今天，包括智能手机、助听器和声学人机界面在内的语音通信设备无处不在。然而，在许多日常情况下，语音信号会被噪音扭曲，例如在自助餐厅或繁忙的街道上。为了减少这些干扰对语音通信的负面影响，使用了语音增强算法。然而，在最需要语音增强算法的声学挑战场景中，性能增益仍然有限。因此，在这个项目中，我们的目标是使语音增强算法更可靠地工作，以便在声学困难的环境中简化语音通信。语音增强通常应用于频谱变换域，其中信号系数是复值的，即它们由频谱幅度和相位表示。然而，大多数关于单通道语音增强的研究都集中在频谱幅度上，而忽略了频谱相位。然而，最近的研究，包括我们在该项目第一个资助期的工作，表明频谱相位对语音增强的重要性可能被低估了：通过仪器测量和听力实验，我们能够证明对干净语音频谱相位的估计可以用来提高语音增强性能，特别是在具有挑战性的声学场景中。现在，我们的目标是在这些结果的基础上，推导出新的和改进的干净语音频谱相位估计器，以进一步提高性能。为此，我们将开发并结合浊音、浊音和瞬态音的单独相位估计器。我们还将把预训练语音增强的最新进展，例如，基于深度神经网络（dnn），转化为阶段处理。目前，大多数预训练方法仅依赖于幅度特征，并且只修改谱幅度。在这里，我们的目标是克服这两个限制，为此我们将研究新的相位特征以及如何使用它们，例如构建新的基于深度神经网络的相位感知语音增强系统。我们的研究将为相位处理在语音增强中的作用和相关性以及将提高语音通信设备性能的新算法提供新的和有价值的见解。

Today, speech communication devices including smart-phones, hearing aids and acoustic human-machine interfaces are ubiquitous. However, in many everyday situations, speech signals are distorted by acoustic noise, for example in a cafeteria or on a busy street. To reduce the negative impact of these disturbances on speech communication, speech enhancement algorithms are used. However, in the acoustically challenging scenarios in which speech enhancement algorithms would be needed most, the performance gain is still limited. Therefore, in this project we aim at making speech enhancement algorithms work more reliably to ease speech communication in acoustically difficult environments.Speech enhancement is usually applied in a spectral transform domain, in which the signal coefficients are complex-valued, i.e. they are represented by spectral amplitudes and phases. Still, most research on single-channel speech enhancement focused on spectral magnitudes while the spectral phase was largely ignored. However, recent research, including our work in the first funding period of this project, indicate that the importance of the spectral phase for speech enhancement might have been underestimated: with instrumental measures and in listening experiments we were able to show that an estimate of the clean speech spectral phase can be used to improve the speech enhancement performance, especially in challenging acoustic scenarios. Now, we aim at building up on these results and derive new and improved estimators of the clean speech spectral phase to improve performance further. For this, we will develop and combine individual phase estimators for voiced, unvoiced, and transient sounds. We will also translate recent advances in pre-trained speech enhancement, e.g., based on deep neural networks (DNNs), to phase processing. Currently, most pre-trained approaches rely only on magnitude features and also only modify the spectral magnitudes. Here, we aim at overcoming both limitations, for which we will investigate new phase features and how they can be employed, e.g. to build novel DNN based phase-aware speech enhancement systems. Our research will provide new and valuable insights into the role and relevance of phase processing for speech enhancement as well as novel algorithms that will boost the performance of speech communication devices.