Investigating deep neural network based transforms of robust audio features for LVCSR

Micro-modulation components such as the formant frequencies are very important characteristics of spoken speech that have allowed great performance improvements in small-vocabulary ASR tasks. Yet they have limited use in large vocabulary ASR applications. To enable the successful application, in real-life tasks, of these frequency measures, we investigate their combination with traditional features (MFCC's and PLP's) by linear (e.g. HDA), and non-linear (bottleneck MLP) feature transforms. Our experiments show that such integration, using non-linear MLP-based transforms, of micro-modulation and cepstral features greatly improves the ASR with respect to the cepstral features alone. We have applied this novel feature extraction scheme onto two very different tasks, i.e. a clean speech task (DARPA-WSJ) and a real-life, open-vocabulary, mobile search task (Speak4it^SM), always reporting improved performance. We report relative error rate reduction of 15% for the Speak4it^SM task, and similar improvements, up to 21%, for the WSJ task.