Speaker: María Pilar Fernández Gallego
Abstract: Nowadays a large amount of companies record conversations, calls, sales or even meetings, in many cases to comply with the current legislation. Apart from the legal need, these recordings constitute an invaluable source of information about clients, call center operators, marketing campaigns, markets trends, etc. The current state of the art in Automatic Speech Recognition (ASR) allows to exploit this information in a very efficient way. However, the recordings at these repositories tend to present very low quality because the audio is typically recorded in a highly compressed way to save storing space. Besides, since it is very common to use Voice over IP (VoIP) in these systems, it is usual to have short interruptions in the speech signal due to packet losses. Both effects, and particularly the last one, have an impact in ASR performance. This paper presents an extensive study of the influence of these effects and the effectiveness of different data augmentation strategies to increase the robustness of ASR systems in these circumstances, and in particular when packet losses degrade the speech signal.