Multi-keyword spotting based on speech feature space trace matching

Keyword spotting has been an active research issue in recent years. An efficient multi-keyword spotting method based on the speech feature space trace matching is proposed. Let each template do the moving matching along test speech trace and take down each step´s matched score. If the score is lower than the template threshold, write the matched score to the distance matrix. In consideration during the spotting, once some template matched successfully the next matching template is on the left part or the right part of the previous matched template, so the bi-tree recursively-created algorithm is adopted. Using this algorithm, each time find a lowest matched score in the distance matrix to create the bi-tree node, and then output the bi-tree, which is the spotting result. Compared to other methods, this method can greatly reduce the computation in spotting. Experiment shows the recognition rate for speaker-independent is 90.8% and for speaker-dependent is 97.8%, having considerable practicability.