Time-domain beamforming and blind source separation: speech input in the car environment

Speech is a natural and therefore privileged communication modality. Safety and convenience issues require hands-free, eyes-free speech-based human-computer interfaces to manipulate complex functionalities and devices. For example, in cars, applications include entertainment, telephony as well as more advanced functions such as automatic spoken language dialog systems for in-vehicle navigation. With a seamless speech input, such interfaces bring an increased comfort but have to face several issues: degradation of the signal-to-noise ratio (SNR) at the microphone, reverberated speech signal, and, above all, the presence of interferences. The interferences, such as speech from the co-driver, can greatly hamper the performance of the speech recognition component, which is crucial for dialog applications. Especially for overlaid speech, the separation of the target speaker from the interferer represent a particular challenge.

Time-domain Beamforming and Convolutive Blind Source Separation addresses the problem of separating spontaneous multi-party speech by way of microphone arrays (beamformers) and adaptive signal processing techniques.

While existing techniques requires a Double-Talk Detector (DTD) that interrupts the adaptation when the target is active, the described method addresses the separation problem using continuous, uninterrupted adaptive algorithms. The advantage of such an approach is twofold: Firstly, the algorithm development is much simpler since no detection mechanism needs to be designed and no threshold to be tuned. Secondly, the performance can be improved due to the adaptation during periods of double-talk.

The book is organized in three parts, roughly described as follows:

The first line of attack, termed implicit beamforming, is built upon the classical supervised beamforming, i.e. it requires the position of the target speaker to be known. Using a time-varying pseudo-optimal step-size that takes over the adaptation control, a continuous adaptive algorithm is obtained. Experimentally, the performance of this algorithm appears to be sufficient if the microphones are oriented adequately. However, in general, more sophisticated Blind Source Separation (BSS) techniques are required.

In the second part, the time-domain BSS method (Buchner et al., 2005) exploiting second-order statistics of the source signals is considered. This method is based on the natural gradient and limited to square systems with an equal number of sources and microphones. Introducing the concept of partial separation, a novel approach is proposed to remove this restriction of the natural gradient. The Sylvester-based representation of the separation system allows a very concise derivation of second-order BSS algorithms in the time-domain but cannot be directly implemented. Revisiting the natural gradient in the z-domain, this implementation issue is clarified. Furthermore, the convergence and stability of BSS is discussed from a theoretical point of view, and its properties are compared to those of supervised beamforming.

Finally, combinations of beamforming and BSS are presented leading to already known, but also novel algorithms. The underlying idea is the following: if the position of the target speaker (the driver) is known in advance, a purely blind approach, which does not exploit this information, seems sub-optimal. Therefore, an emphasis is placed on the development of an algorithm that combines the benefits of both approaches. It outperforms BSS and removes the need for a DTD and allows for a continuous adaptation, even during double-talk.

 The book is written is a concise manner and an effort has been made such that all presented algorithms can be straightforwardly implemented by the reader. All experimental results have been obtained with real in-car microphone recordings involving simultaneous speech of the driver and the co-driver, as opposed to computer-generated simulations. Experiments with background noise have been carried out in order to assess the robustness of the considered methods in noisy conditions.