Publicação
Low band continuous speech system for voice pathologies identification
| Resumo: | This paper describes the impact of the signal bandwidth reduction in the identification of voice pathologies. The implemented systems evaluate the identification of 3 classes divided by healthy subjects, subjects diagnosed with physiological larynx pathologies and subjects diagnosed with neuromuscular larynx pathologies. Continuous speech signals are down-sampled to 4 kHz and the extracted spectral parameters are applied to a GMM classifier. No significant change in accuracy occurs, being possible to conclude that the low frequencies contain sufficient information to allow the classification of pathologies. A second objective is to test the effects of suppressing the voice activity detection and the increasing the analysis window length. In both cases the accuracy increases. In conclusion, a pathological voice identification system based on signals sampled at 4 kHz, without voice activity detection and with an analysis window length of 40 ms is proposed, getting 81.8% accuracy. The proposed system has also the advantage of reduces the storage memory and the processing time. |
|---|---|
| Autores principais: | Cordeiro, Hugo |
| Outros Autores: | Meneses, Carlos |
| Assunto: | Voice pathologies identification Low band speech analysis Spectral parameters Voice activity detection |
| Ano: | 2018 |
| País: | Portugal |
| Tipo de documento: | documento de conferência |
| Tipo de acesso: | acesso restrito |
| Instituição associada: | Instituto Politécnico de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Científico do Instituto Politécnico de Lisboa |
| Resumo: | This paper describes the impact of the signal bandwidth reduction in the identification of voice pathologies. The implemented systems evaluate the identification of 3 classes divided by healthy subjects, subjects diagnosed with physiological larynx pathologies and subjects diagnosed with neuromuscular larynx pathologies. Continuous speech signals are down-sampled to 4 kHz and the extracted spectral parameters are applied to a GMM classifier. No significant change in accuracy occurs, being possible to conclude that the low frequencies contain sufficient information to allow the classification of pathologies. A second objective is to test the effects of suppressing the voice activity detection and the increasing the analysis window length. In both cases the accuracy increases. In conclusion, a pathological voice identification system based on signals sampled at 4 kHz, without voice activity detection and with an analysis window length of 40 ms is proposed, getting 81.8% accuracy. The proposed system has also the advantage of reduces the storage memory and the processing time. |
|---|