next up previous contents
Next: Hidden Neural Network (HNN) Up: Task 5.2: Technical Description Previous: Forward-Backward training


The performance of this algorithm was tested on the task-independent PHONEBOOK database [25]. Recognition experiments were performed on a 600 word lexicon using log-RASTA PLP as acoustic features. We defined two training sets, a small one (9,000 utterances) and a large one (19,000 utterances). Results are reported in table 5.2

Table 5.2: Error rates on isolated word recognition with hybrid system and log-rasta plp features. Comparison between Viterbi and full forward-backward training.
Word error rate Training set 1 Training set 2
Viterbi 13.7% 9.8%
F-B 12.2% 10.1%


Christophe Ris