Experiments of in-car audio compensation for hands-free speech recognition

The work presented in this paper aims at improving in-car speech recognition performance in presence of interfering signals diffused by the loudspeakers, addressing the well known problem of acoustic echo cancellation. A NLMS-based technique has been tested with real in-car speech recordings and issues like algorithm parameters and convergence or test/training mismatch have been investigated. The echo canceller shows good properties such as acceptable adaptation speed and negligible speech distortion: in the best configuration the WER was lowered from 18.3% to 5.5%. Beneficial effects have also been observed by adopting a contamination technique in the acoustic model training phase.