Data-driven modulation filter design under adverse acoustic conditions and using phonetic and syllabic units

Constructing speech feature extraction methods that are robust to many types of corrupting acoustic environments remains a daunting task and it is instructive to investigate which properties of the speech carry the discriminative information for recognition under a variety of conditions. In this paper we describe results for generating RASTA-style modulation filters under a number of acoustic environments. We utilize Linear Discriminant Analysis in a manner previously described by van Vuuren and Hermansky to automatically generate discriminant filters for speech with artificially added background noise and reverberation. We also generate the filters using both phonetic and syllabic classification targets. Trends in the responses of the discriminant filters lend support to feature extraction design decisions employed by RASTA-PLP and Modulation-filtered Spectrogram features. Further, tests with added reverberation corroborate views on the perceptual stability of syllabic rates.