EUROSPEECH 2003 - INTERSPEECH 2003
Handsfree speaker input is mandatory to enable safe operation in cars. In those scenarios robust speech recognition emerges as one of the key technologies to produce voice control car devices. Through this paper, we propose a method of processing speech degraded by reverberation and noise in an automobile environment. This approach involves analyzing the linear prediction error signal to produce a weight function suitable for being combined with spectral subtraction techniques. The paper includes also an evaluation of the performance of the algorithm in speech recognition experiments. The results show a reduction of more than 30% in word error rate when the new speech enhancement frontend is applied.
Bibliographic reference. Alvarez, A. / Nieto, V. / Gomez, P. / Martinez, R. (2003): "Speech enhancement for a car environment using LP residual signal and spectral subtraction", In EUROSPEECH-2003, 1373-1376.