Turning wine into water: Can ordinary speech be artificially nasalized?

Synthetic speech has recently been used to study resonance and voice disorders. The advantage of synthetic speech is that it becomes possible to control and artificially manipulate acoustic variables related to a specific perceptual feature without confounding effects from other co-occurring problems often found in disordered speech. To date, synthesis has been restricted to vowels in isolation or in single words. An analysis-synthesis technique has been used that starts by making a spectral match to actual disordered speech. The study described here has investigated the feasibility of using another technique that may be suitable for connected speech contexts. The technique involves spectral manipulation of vowels within samples of connected natural speech. Several variations were tested to mimic the nasal cavity’s effect on the speech signal: direct manipulation of spectral features extracted from linear predictive coding analysis, filtering with a pole-zero filter and adding a pole-zero pair to the transfer function. All spectral modifications were applied to vowels within sentences of natural speech. Problems that were encountered with the techniques are discussed. Further a database of pilot sentences has been prepared for perceptual evaluation of their resonance quality and comparison of the different types of modification.