Speech Transduction Based on Linguistic Content

Digital hearing aids use a variety of advanced digital signal processing methods in order to improve speech intelligibility. These methods are based on knowledge about the acoustics outside the ear as well as psychoacoustics. This paper investigates the recent observation that speech elements with a high degree of information can be robustly identified based on basic acoustic properties, i.e., function words have greater spectral tilt than content words for each of the 18 Danish talkers investigated. In this paper we examine these spectral tilt differences as a function of time based on a speech material six times the
duration of previous investigations. Our results show that the correlation of spectral tilt with information content is relatively constant across time, even if averaged across talkers. This indicates that it is possible to devise a robust method for estimating information density in the speech signal based on
computationally simple short-term band-level differences. The principle described here has the potential to improve speech transduction in hearing aids and cochlear implants. In addition, the concept of information-based speech transduction may also be applicable in automatic speech recognition systems.