Perturbation-Correlation to Measure the Relative Importance of Frequency Bands in Speech Recognition

Explore the perturbation-correlation method for measuring frequency band importance in speech recognition through this research lecture from the Center for Language & Speech Processing at Johns Hopkins University. Learn how human listeners utilize cues distributed across different frequencies to recognize speech and discover an alternative approach to traditional low- and high-pass filtering methods. Examine the perturbation-correlation technique pioneered by Doherty and Turner, which applies random perturbations independently across frequency bands in wide-band speech and estimates band importance through correlation analysis between signal-to-noise ratios and binary recognition scores across thousands of trials. Understand how this method reveals listener strategies for speech recognition in various acoustic backgrounds, including noise versus competing speech environments. Investigate inter-listener variability patterns and the effects of hearing loss on frequency importance functions. Discover recent theoretical developments aimed at improving the mathematical foundation of the perturbation-correlation method as applied to speech processing, with insights into frequency-importance functions that are essential for predictive models of speech intelligibility like the articulation index and optimization of multi-band speech-processing devices such as hearing aids.