Perception and use of high-frequency energy in defining indexical and segmental features in speech

Abstract

Historically, the focus of speech perception research has been on spectral energy below 4-5 kHz. Recent research has shown a substantial amount of useful perceptual information above this frequency cutoff. Our studies at the Auditory Perception Laboratory at West Virginia University have revealed usable information related to talker sex, talker identity, vowel identity, and listening effort for human and machine applications. These benefits have been found in quiet and noisy listening conditions. Current projects are examining the use of high-frequency energy (above 4 kHz) for automated recognition tasks in quiet and noise. Using various spectral and temporal features extracted from high-frequency energy from a large database of signals (approximately 7000 vowels), classification accuracy for vowel identity, speaker sex, and individual speaker identify has been found to be significantly above chance. Specific details of this project as well as additional projects will be included in this presentation. Implications of this line of research will also be discussed.