SPECTRAL PATTERN-RECOGNITION OF IMPROVED VOICE QUALITY

The self-organizing map (a neural network) was applied to the spectral pattern recognition of voice quality in 34 subjects: 15 patients oper ated on because of insufficient glottal closure and 19 subjects not tr eated for voice disorders. The voice samples, segments of sustained /a /, were perceptually rated by six experts. A self-organized acoustic f eature map was first computed from tokens of /a/ and then used for the analysis of the samples. The locations of the samples on the map were determined and the distances from a normal reference were compared wi th the perceptual ratings. The map locations corresponded to the degre e of audible disorder: the samples judged as normal were overlapping o r close to the normal reference, whereas the samples judged as dysphon ic were located further away from it. The comparison of pre- and posto perative samples of the patients showed that the perceived improvement of voice quality was also detected by the map.