Speaker-dependent automatic helium speech normalisation

Previous attempts to solving helium speech problem were oriented towards development of a helium speech production model. The aim was to obtain straightforward transformations which could be applied to helium speech, or its parametric representation, in order to obtain normal speech. However it has never been investigated if the poor quality of the resulting speech might be due to speaker dependance of voice distortion. This paper presents a non-model based approach in which all transformations are automatically calculated from formant parameters and F₀ extracted directly from the normal and helium speech signals of the same speaker. Results show considerable differences among speakers in the distortion of formant frequencies, bandwidths and amplitudes and a large depth-uncorrelated variation of F₀. The proposed unscrambling system results in a more natural and more intelligible sounding speech and it also allows to observe adaptation of divers to the pressurized helium environment over time