Transformation of Jacobian matrices for noisy speech recognition
Abstract
Robustness of automatic speech recognition systems to noise is still an issue, especially for small footprint systems. This paper addresses the problem of noise robustness using model compensation. Jacobian adaptation is an example of model compensation, which is also computationally efficient. However, its performance drops when the mismatch between training and testing environments increases. Our new adaptation algorithm is based on finding the best linear transformation for compensating the mismatch between training and testing conditions. In this paper, it is shown that, using a dimensionality reduction technique, time and memory requirements of the algorithm can be further decreased compared to the classical Jacobian adaptation. Furthermore, experimental results for an English digit recognition task, recorded in a car at 60 mph, show that the error rate decreases by 66 %, compared to the classical Jacobian adaptation.