Next: CONCLUSION Up: A COMPARISON OF Previous: TUNING THE RECURRENT

COMPARISON WITH HMM

The best Hidden Markov Model phoneme recognition results are currently obtained by using multiple codebooks and smoothed triphone models and the best published results to date for the TIMIT task come from the SPHINX recognition system [2]. By simply mapping the 61 TIMIT symbol set onto the 39 CMU/SPHINX symbol set for the pd4 configuration it is possible to make a comparison between the Hidden Markov Model (HMM) and the Artificial Neural Network (ANN) techniques. These results are shown in table 11, the difference in performance between the two techniques is believed to be significant.

  
Table 11: Comparison with the SPHINX phone recogniser