Single-Microphone Speech Separation: The use of Speech Models

Abstract

Separation of speech sources is fundamental for robust communication. In daily conversations, signals reaching our ears generally consist of target speech sources, interference signals from competing speakers and ambient noise. Take an example, talking with someone in a cocktail party and making a phone call in a train compartment. Fig. 1 shows a typical indoor environment having multiple sound sources, such as speech from different speakers, sounds from a television set and telephone ringing, etc. These sources are often overlapped in time and frequency. While human attends to individual sources without difficulty, most speech applications are vulnerable and resulted in degraded performance. This chapter focuses on speech separation for single microphone input, in particular, the use of prior knowledge in the form of speech models. Speech separation for single microphone input refers to the estimation of individual speech sources from the mixture observation. It remains important and beneficial to various applications, namely surveillance systems, auditory prostheses, speech and speaker recognition. Over the years, extensive effort has been devoted. Speech enhancement and separation are two popular approaches. Speech enhancement (Lim, 1983; Loizou, 2007) generally reduces the interference power, by assuming that certain characteristics of individual source signals are held. There is one speech source at most. In contrast, speech separation (Cichocki & Amari, 2002; van der Kouwe et al., 2001) extracts multiple target speech sources directly.