Next: ACKNOWLEDGEMENTS Up: A COMPARISON OF Previous: COMPARISON WITH HMM

CONCLUSION

From the previous sections it can be seen that most power spectrum based preprocessors give about the same performance. It would have been possible to run each preprocessor several times with different starting weights, so eliminating this source of variance and so obtaining a more accurate ranking. However, the difference between preprocessors was found to be small and changes to the network were found to yield far more significant improvements.

Is is perhaps not surprising that the conventional signal processing front ends perform better for speech recognition by machine than the auditory model, since signal processing methods have been intensively studied and optimised for just this purpose, whereas auditory models have been pursued mainly for psychophysical modelling results.

It is interesting that the autocorrelation function ( acf) and linear predictive filter ( lpf) performed worse than the other LPC techniques, even though the latter preprocessors could all be derived from the former. This demonstrates that although within the connectionist framework it is theoretically possible to perform any arbitrary mapping, the data representation, in the form of the choice of preprocessor, is important.

Unfortunately, the best preprocessor for phoneme recognition is not necessarily the best for word recognition, as has been demonstrated by a number of researchers including Russell et. al. [7]. As a result, this work is currently being extended to word recognition and preliminary results from the 1000 word DARPA Resource Management task will soon be presented [8].

This paper has presented the results of training 35 networks at about 17 hours each on a 64 processors array. This represents over 4 CPU years and it is disappointing that no significant improvement in preprocessors was found. However, changes to the recurrent network have yielded increases in performance, and the authors believe that the resulting technique yields the best results on this task to date.

Next: ACKNOWLEDGEMENTS Up: A COMPARISON OF Previous: COMPARISON WITH HMM