Análise comparativa entre os métodos HMM e GMM-UBM na busca pelo a-ótimo dos locutores crianças utilizando a técnica VTLN

Abstract

The aim of this work is to find means to minimize the high error rate found in speech recognition systems which are trained on adult speakers and tested on children speakers. In this regard, we propose the use of the GMM-UBM method as an alternative to the HMM method to find the optimal warping factor (α-optimal) for children speakers when the speaker normalization technique is used. The adopted normalization technique was VTLN, which normalizes the vocal tract of different children speakers through the use of mel filterbank frequency warping. The assessment of this technique also aimed to find the optimal mixture quantity that improves the system performance. Thus, the error rate in the system trained with adults and tested on children was reduced from 4,95% to 1,88% when VTLN was used with α-optimals found by HMM and to 1,92% when VTLN was used with α-optimals found by GMM. It was noticed that the application of VTLN technique using α-optimals found by GMM-UBM method achieved a similar performance to HMM in the experiments. From the experiments it was observed that choosing GMM-UBM method turns to be more suitable due to its implementation simplicity and to the need of lower computational cost, being thus an alternative to HMM in the use of VTLN in Speech Recognition Systems for children speakers.