DNN-based speech enhancement in ad-hoc microphone arrays

Ad-hoc microphone arrays (AHMA) are formed by the network of microphones embedded in our daily-life devices like phones, computers or smart speakers. They offer a big potential for natural language processing, especially speech enhancement, which consists in removing the noise from the recordings. However, this processing step is hindered by the small size of the devices and by their asynchronous recordings.

This thesis proposes a speech enhancement solution adapted to the usage conditions of ad-hoc microphone arrays, by combining the modelling power of deep neural networks (DNNs) with the flexibility of use of AHMAs. We conduct a detailed empirical analysis of our system to validate its efficiency and to highlight the benefits of jointly using DNNs and distributed speech enhancement algorithms.

Besides, we develop our solution in order to make it robust to the challenges of the usage conditions of ad-hoc microphone arrays. We study its behaviour when the number of devices in the AHMA varies, and when the signals of different devices are not synchronised. We propose a solution to adapt our system to each of these conditions. Lastly, we propose an extension of our system to make it a suitable solution to source separation in a meeting scenario.