Helium speech normalisation by codebook mapping

In this paper we present a non-parametric approach to solving the helium speech problem. Properties of helium speech are replaced by those pertaining to normal speech by means of codebook mapping of spectral envelopes. This method eliminates the drawbacks inherent in the previous procedures of helium speech unscrambling as it requires neither model of helium speech production nor estimation of formant parameters. The only assumption is the general source-filter model required for linear prediction analysis. In the traditional approach spectral transformations were computed based on the assumed helium speech production model. And in the nonmodel approach it was assumed that helium speech distortion is speaker dependent, so all spectral transformations were calculated from formant parameters and F0 extracted directly from speech signals. In all previous methods the resulting speech was still retaining a nasal quality due to inaccurate modelling and speech processing schemes that were unable to guarantee independent manipulation of formant parameters. On the contrary our system results in speech that is completely free of the hyperbaric helium quality however its technical quality is still unsatisfactory as the mapping introduces noise into the corrected speech.