On-Line Feature and Acoustic Model Space Compensation for Robust Speech Recognition in Car Environment

In order to develop a robust man-machine interface based on speech for cars, the speaker variability and the acoustic environment effects have to be compensated. In this work, an on-line feature and acoustic model compensation (MATE-MEMLIN) is proposed to compensate the speaker variability and the acoustic car environment. MATE-MEMLIN consists on the combination of the techniques augMented stAte space acousTic modEl (MATE) and Multi-Environment Model based Linear Normalization (MEMLIN). MATE defines expanded acoustic models to compensate the speaker frequency variability using data driven estimated linear transformations. On the other hand, MEMLIN, an empirical feature vector normalization technique, was also presented and it was proved to be effective to compensate environment mismatch. Some experiments with Spanish SpeechDat Car database were carried out in order to study the performance of the proposed technique in a real car environment, reaching an important mean improvement in Word Error Rate, WER.