Skip to main content

Table 4 Evolution of CER and acoustic model size according to the number of Gaussian components of the emiting states (context-free models). Acoustic vectors are composed of 13 coefficients (12 PLP plus energy). voice-command recognition tests performed on the corpus VODIS (voice command/noisy).

From: Compact Acoustic Models for Embedded Speech Recognition

No. of gauss/state No. of parameters CER
2 5 832 5,80%
4 11 664 4,80%
128 373 248 3,94%
full 1 092 096 1,80%