Source separation and noise reduction for automatic speech recognition in dynamic acoustic scenarios

该项目的目标是开发一种自动语音识别系统，用于智能家居环境中的免提语音控制。从DFG资助的前期项目中开发的多通道盲源分离和降噪技术开始，这些算法将在资源受限的嵌入式硬件上进一步开发和优化延迟和可实现性。与工业伙伴现有的语音识别系统相比，将开发的Prototyp将以低延迟的方式实时执行识别，并在智能家居应用的典型声学环境中实现更高的识别率。作为上述参数源分离和语音增强算法的替代方案，我们还将研究和开发基于神经网络的解决方案。我们将在增强/识别性能、延迟、计算和内存需求以及对不同声学环境条件的稳健性方面对这两种方法进行彻底的比较。

The goal of the project is the development of an automatic speech recognition system to be used for hands-free voice control in a smart home environment. Starting with the multi-channel blind source separation and noise reduction techniques developed in the DFG-funded preceding project, these algorithms will be further developed and optimized with respect to latency and realizability on resource-constrained embedded hardware. The prototyp to be developed shall perform recognition in real-time with low latency and achieve superior recognition rates in acoustic environments typical of smart home applications compared to an existing speech recognition system available at the industrial partner.As an alternative to the above parametric source separation and speech enhancement algorithms, we will also research and develop neural network based solutions. We will carry out a thorough comparison between the two approaches with respect to enhancement/recogntion performance, latency, computational and memory requirements and robustness towards varying acoustic environmental conditions.