Table 3 reports on the efficiency of the NLP components for the set of 1000 wordgraphs and test utterances. The first two rows present the results for sentences; the remaining rows provide the results for word-graphs. Listed are respectively the average number of milliseconds per input; the maximum number of milliseconds; and the maximum space requirements (per word-graph, in Kbytes).
For most word-graphs we used the nlp_speech_trigram method as described above. For large word-graphs (more than 100 transitions), we first selected the best path in the word-graph based on acoustic scores and N-gram scores only. The resulting path was then used as input for the parser. In the case of these large word-graphs, N=2 indicates that bigram scores were used, for N=3 trigram scores were used.
CPU-time includes tokenizing the word-graph, removal of pause transitions, lexical lookup, parsing, the robustness/disambiguation component, and the production of an update expression. 10
For word-graphs the average CPU-times are actually quite misleading because CPU-times vary enormously for different word-graphs. For this reason, we present in table 4 the proportion of word-graphs (in %) that can be treated by the NLP component within a given amount of CPU-time (in milliseconds).