Keyword spotting using normalization of posterior probability confidence measures
Vargiya, Rachna Vijay
MetadataShow full item record
Keyword spotting techniques deal with recognition of predefined vocabulary keywords from a voice stream. This research uses HMM based keyword spotting algorithms for this purpose. The three most important components of a keyword detection system are confidence measure, pruning technique and evaluation of results. We suggest that best match for a keyword would be an alignment in which all constituent states have high emission probabilities. Therefore score of even the worst subsequence must also be better than a threshold and a path can be represented by the score of its worst subsequence match. This confidence measure is called Real Fitting. The harsher the pruning in a technique, the fewer paths survive. This increases the speed as well as the risk of pruning the best match. Three levels of pruning are explored and results and performance are compared. Since the proposed algorithms do not follow the principle of optimality, possible optimal paths could be pruned. Therefore we propose a pruning that permits a set of paths to be retained in every frame instead of a single path. Finally, the evaluation of these algorithms plays an important role in assessing the performance of these algorithms. Since the choice of keywords can affect the results, an equal opportunity evaluation is proposed which evaluates the algorithm on their respective best keywords.