Multichannel Speech Enhancement in Cars: Explicit vs. Implicit Adaptation Control

Speech-based command interfaces are becoming more and more common in cars. Applications include automatic dialog systems for hands-free phone calls as well as more advanced features such as navigation systems. However, interferences, such as speech from the codriver, can significantly hamper the performance of the speech recognition component, which is crucial for those applications. This issue can be addressed with adaptive interference cancellation techniques such as the Generalized Sidelobe Canceller (GSC). In order to cancel the interference (codriver) while not cancelling the target (driver), adaptation must happen only when the interference is active and dominant. This paper proposes a novel approach for pre-estimation of target and interference energies, along with its application to “explicit” adaptation control: a hard decision whether the filter(s) should be updated or not. It is compared with an “implicit” adaptation control method that does not require such pre-estimation. Two physical setups S1 and S2 are considered, with respectively 2 directional microphones, or 4 directional microphones in the rear-view mirror, as shown in Fig. 1. The spacing between microphones is 17 cm in S1, and 5 cm in S2. Section 1 presents an adaptation algorithm for each setup. Section 2 describes the two adaptation control methods. Experiments on real in-car recordings are presented in Section 3. Section 4 concludes. A detailed version of this paper can be found in [6], and will include additional experimental results in the future.