ISCA Archive Eurospeech 1993
ISCA Archive Eurospeech 1993

Feature extraction for profoundly deaf people

U. Hartmann, K. Hermansen, F. K. Fink

Profoundly deaf people only have a minor frequency range available for reception of information in speech signals. These people do not benefit sufficiently from standard hearing aids. To overcome this problem we have developed a technique for transforming speech signals from one frequency range to another maintaining as much information and "speech likeness" as possible. Our concept includes 1) parametric modeling of the speech production system, 2) transformation of the speech production model to match the available frequency range, 3) resynthesis of the speech using the transformed model. In this way we can present the speech information of interest in a frequency range at choise. Supplementary, this concept is believed to reduce wideband background noise which is a problem for hearing disabled as well as for people with normal hearing ability. The technique is well suited for real time implementation (VLSI), primarily caused by numeric robustness of the algorithms. An example - a transformed danish token "Polsevognen" - is used for highlighting the benefits from using the technique.

Keywords: Parametric modelling, Pseudodecom-position, Transposer, Speech enhancement, Noise reduction, Speech intelligibility