Low-Complexity Variable Frame Rate Analysis for Speech Recognition and Voice Activity Detection

Frame-based speech processing inherently assumes a stationary behavior of speech signals in a short period of time. Over a long time, the characteristics of the signals can change significantly and frames are not equally important, underscoring the need for frame selection. In this paper, we present a low-complexity and effective frame selection approach based on a posteriori signal-to-noise ratio (SNR) weighted energy distance: The use of an energy distance, instead of, e.g., a standard cepstral distance, makes the approach computationally efficient and enables fine granularity search, and the use of a posteriori SNR weighting emphasizes the reliable regions in noisy speech signals. It is experimentally found that the approach is able to assign a higher frame rate to fast changing events such as consonants, a lower frame rate to steady regions like vowels and no frames to silence, even for very low SNR signals. The resulting variable frame rate analysis method is applied to three speech processing tasks that are essential to natural interaction with intelligent environments. First, it is used for improving speech recognition performance in noisy environments. Second, the method is used for scalable source coding schemes in distributed speech recognition where the target bit rate is met by adjusting the frame rate. Third, it is applied to voice activity detection. Very encouraging results are obtained for all three speech processing tasks.