Contributions to speech processing and ambient sound analysis
Serizel, Romain · HAL (Le Centre pour la Communication Scientifique Directe) · 2022
We are constantly surrounded by sounds that we continuously exploit to adapt our actions to situations we are facing. Some of the sounds like speech can have a particular structure from which we can infer some information, explicit or not. This is one reason why speech is possibly that is the most intuitive way to communicate between humans. Within the last decade, there has been significant progress in the domain of speech andaudio processing and in particular in the domain of machine learning applied to speech and audio processing. Thanks to these progresses, speech has become a central element in many human to human distant communication tools as well as in human to machine communication systems. These solutions work pretty well on clean speech or under controlled condition. However, in scenarios that involve the presence of acoustic perturbation such as noise or reverberation systems performance tends to degrade severely. In this thesis we focus on processing speech and its environments from an audio perspective. The algorithms proposed here are relying on a variety of solutions from signal processing based approaches to data-driven solutions based on supervised matrix factorization or deep neural networks. We propose solutions to problems ranging from speech recognition, to speech enhancement or ambient sound analysis. The target is to offer a panorama of the different aspects that could improve a speech processing algorithm working in a real environments. We start by describing automatic speech recognition as a potential end application and progressively unravel the limitations and the proposed solutions ending-up to the more general ambient sound analysis.