next up previous contents
Next: WP4 Future Development Up: Task 4.1: Technical Description Previous: Experiments

Results


 
Table 4.3: Error rates on isolated word recognition (75 and 600 lexicon words never seen in the training data) with hybrid HMM/ANN systems and RASTA-PLP features. Comparison is performed between context independent phoneme models and combination of context independent and context dependent models. Only automatic data driven clustering is demonstrate in this table.
Model Type 75 words 600 words
CI 2.2% 7.8%
CI&CD(III) 1.7% 5.7%
 


 
Table 4.4: Error rates on isolated word recognition (75, 600 and 8000 lexicon words) with hybrid HMM/ANN systems and CMS features. Comparison between context independent and combination of context independent and context dependent models. The 3 kinds of phone to phone transition clustering are presented here.
Model type 75 words 600 words 8000 words
CI 1.5% 5.3% 17.3%
CI&CD(I) 1.1% 3.9% 13.8%
CI&CD(II) 1.1% 3.3% 12.0%
CI&CD(III) 1.1% 4.2% -
 

As shown on Table 4.3 and Table 4.4 (which present results of experiments carried out on both rasta-plp and cms feature sets), using CD&CI models yielded up to 38% reduction of the error rate with only a slight increase of the computational load due to the modified word topologies.


next up previous contents
Next: WP4 Future Development Up: Task 4.1: Technical Description Previous: Experiments
Christophe Ris
1998-11-10