Some experimental comparisons.

According to the ASR experiments reported in Publication 5 and recent checks with 80-dimensional context vectors [Kurimo, 1997], the obtained speed-up compared to the unordered complete search (optimized by PDC as well) is about 26-42 % depending on the codebook size and the feature vectors. The corresponding increase of the average number of errors is about 4-10 %. For codebooks trained with the segmental LVQ3 the same relative speed-up is obtained and the amount of errors increases by about 10 %. It seems that here the further training without neighborhood updates has not excessively damaged the topology necessary for speeding up the search.

To reflect the real significance of the speed-up, the relative improvements above were measured for the whole ASR system (Figure 2) except the mel-cepstrum computation (less than 10 % of the total time). The output density estimation which is the approximated part here, consumes originally about 50 % of the total time. The largest codebooks tested here have, however, contained only 140 units per phoneme for 80-dimensional inputs, so gains might be more significant, if really large SOMs would be required.