Table 3 DER and acoustic model size according to the number of Gaussian components per state (context-free models). The acoustic vectors are composed of 13 coefficients (12 PLP plus energy). isolated-digit recognition tests were performed on the corpus BDSON (digit/clean).