Acoustic Models with the Decoder SPOJUS Sample Clauses

Acoustic Models with the Decoder SPOJUS. The acoustic models used with the decoder SPOJUS are based on syllable HMMs, which have been developed in our laboratory [7]. The number of Japanese syllables for the syllable HMMs is 116. The sampling frequencies are 12 kHz / 16kHz and the frame shift lengths are 8 msec / 10 msec. The following three types of the sets of feature parameters are evaluated: • 2m dimensional xxx frequency cepstrum coefficients (MFCC) segmented from 4 successive frames, delta m dimensions calculated over 9 successive frames, delta delta m dimen- sions and delta, delta delta powers (m = 10, 12). • 12 dimensional xxx frequency cepstrum coefficients (MFCC), delta 12 dimensions, delta delta 12 dimensions and delta, delta delta powers The acoustic models are gender-dependent (male) syllable unit HMMs that have 5 states 4 densities, 4 Gaussian mixtures mod- els per density with full covariance / diagonal covariance matrices. We also switch between conventional HMM with self loop transi- tion and HMM with duration control, and evaluate both of them. Among various combinations of features such as the sampling frequencies, frame shift lengths, feature parameters, covariance matrices, and self loop transition / duration control, we carefully choose nine acoustic models so that they include the best perform- ing ones as well as a sufficient number of minimal pairs which have difference in only one feature. Then, as in the case of the acoustic models used with the decoder Xxxxxx, for each of the nine models, we evaluate both HMMs with and without the short pause states2, which amount to 18 acoustic models in total.