Comparative analysis between HMM and GMM-UBM methods:

This research addresses ways to minimize the high error rate in speech recognition systems trained with adult speakers and tested with child speakers. The GMM-UBM method is used as an alternative to the HMM method in the search for the optimal scaling factor (α-optimal) for child voiceovers when using the speaker standardization technique. The normalization technique adopted is the VTLN which normalizes the vocal tract of different child speakers through the frequency scaling of the honey filter bank. In the evaluation of this technique we also looked for the amount of optimal mixtures that improve the performance of the system.