4.5.1 N-Best Generation
Until this point, we have only been determining the best hypothesis for an utterance. What if we need to determine the ten or twenty best hypotheses? Of course the software makes many other "guesses", but we usually only need to know the best guess. For this experiment, we will generate the five best hypotheses and store the results in a transcription database. Go to the directory
Command: isip_recognize -parameter_file params_decode.sof -list $ISIP_TUTORIA./databases/lists/identifiers_test.sof -verbose ALL Version: 1.16 (not released) 2002/09/25 00:20:53 loading front-end: ../../recipes/frontend.sof loading language model: ../../models/lm_model_update.sof loading acoustic model: ../../models/ac_model_update.sof loading audio database: ./audio_db.sof opening the output file: ./hypo.out processing file 1 (st_9z59362a): ../../features/st_9z59362a.sof ref:In this experiment, the output file contains N best hypotheses instead of just one best hypothesis. The score values indicate the likelihood of each of the N best hypotheses. The lower the likelihood, the lower the confidence is about a hypothesis. Note that the best hypothesis has the best overall likelihood. As discussed above, the output generated by N-Best begins with the best hypothesis followed by other possibilities in decreasing likelihood. This output format consisting of a list of hypotheses is commomly known as a N-best list. N-best lists are popular because they can be postprocessed by many natural language processing tools, and can be reordered based on their grammatical and semantic content. The N-best lists are also popular in the hybrid HMM-SVM based systems. In such systems, the first recognition pass is performed using the conventional continuous density HMM based system to generate the N-best lists. These N-best lists are then used to perform a second recognition pass using an SVM based system. |