Robust word-spotting system using an intelligibility criterion for reliable keyword detection under adverse and unknown noisy environments

A method for calculating a word spotting confidence score for a given word, the method comprising the steps of:

• dividing a minimum value of a speech recognition score by an average value of the speech recognition score over a predetermined period of time such that a matching ratio results, the average value defining an estimated background score;
• and
• normalizing the matching ratio;
• said normalized matching ratio defining the confidence score.

A word spotting system comprising:

• a speech recognizer for generating recognition scores based on a speech signal and lexicon entries for a plurality of words, the recognition scores tracking absolute likelihoods that the words are in the speech signal;
• and
• a spotting module for estimating background scores based on the recognition scores;
• said spotting module calculating confidence scores on a frame-by-frame basis based on matching ratios between minimum recognition values and the background scores, the confidence scores tracking noise-corrected likelihoods that the words are in the speech signal.