Abstrakt

W artykule przedstawiono system automatycznego rozpoznawania jakości i typu głosu śpiewaczego. Przedstawiono bazę danych oraz zaimplementowane parametry. Algorytmem decyzyjnym jest algorytm sztucznych sieci neuronowych. Wytrenowany system decyzyjny osiąga skuteczność ok. 90% w obydwu kategoriach rozpoznawania. Dodatkowo wykazano przy pomocy metod statystycznych, że wyniki działania systemu automatycznej oceny jakości technicznej głosów śpiewaczych są zgodne z wynikami osiągniętymi przez ekspertów. A system designed to recognize automatically the quality and type of a singing voice is presented. A database containing 2690 sample recordings of trained and untrained singers was first constructed. A set of parameters was then derived on the basis of these samples. Artificial neural networks (ANNs) were trained and tested to show that they can recognize a singing voice category automatically on the basis of the defined set of parameters. The results show that in 90% of its decisions the system was able to assign the sample correctly to either an adequate voice quality or voice type. In addition each of the singers' voice samples was judged by six experts, and a parametric technical quality score was assigned to every sample. Next the voice samples, along with their scores, were fed to the input of the ANN. It has been shown that the ANN can be trained effectively to determine the technical quality of singing voices, very similarly to experts. In order to prove their similarity the automatic recognition error distribution and the experts' precision plots were compared statistically. The Pearson's autocorrelation measure was used. The results showed that the critical value of 0.834 (for 0.005) was not reached, thus proving that differences between these results are statistically nonrelevant.

Autorzy (2)