Joint optimization of spectro-temporal features and deep neural nets for robust automatic speech recognition

Abstract

In speech recognition, feature extraction and acoustical model training are traditionally done in two separate steps. Here, instead, we use a framework that combines spectro-temporal feature extraction and the training of neural network based acoustic models into a single process. We found earlier that this approach can be successfully applied for the recognition of speech. In this paper, we propose two further improvements to our method based on recent advances in neural net technology and extend our evaluation to speech contaminated with new types of noise. By repeating our experiments on TIMIT phone recognition tasks using clean and noise contaminated speech, we can compare the recognition performance of the original framework with our new, modified framework. The results indicate that both these modifications significantly improve the recognition performance of our framework. Moreover, we will show that these modifications allow us to achieve a substantially better performance than what we got earlier.