Introduction

Speech is a natural and therefore a privileged communication modality for humans. For example in cars, safety and convenience issues require hands-free (or “seamless”) speech-based human–machine interfaces for the driver to manipulate complex functionalities and devices while driving. Applications include hands-free phone calls as well as more advanced functions such as automatic dialog systems for in-vehicle navigation assistance systems [71]. With a seamless speech input, such interfaces increase comfort but have to face several issues:

• (i) The signal-to-noise ratio (SNR) at a given microphone can be weak relative to the background noise since the signal energy is inversely proportional to the square of the distance to the sound source [14]. Moreover, room acoustics leads to a reverberated speech signal.

• (ii) Interferences, such as speech from the codriver, may greatly hamper the speech recognizer performance, which is crucial for human–machine dialog applications. Separation of the target speaker during periods of competing speech from the codriver represent a particular challenge. This is because the characteristics of the interferer signals cannot be directly estimated from the microphone signals during these periods [50]. This problem is of particular importance since spontaneous multiparty speech contains lots of overlaps between the speech flows of the participants [43].

These issues make the seamless speech input a challenging problem. Before recognizing speech as a sequence of words, an important preprocessing step is to denoise the speech signal from its perturbations. In this book, we address the issue of separating the desired signal from interfering speech, i.e., the point (ii) above.