DESIGN OF AN AUTOMATIC WORD BOUNDARY DETECTION SYSTEM USING THE COUNTING RULE

Abstract

Word boundary detection is the stepping stone for many applications like keyword spotting, speech recognition, etc. It is proved that fifty percent of the speech recognition errors are due to the errors in the word boundary detector. Efficient word boundary detection can reduce the recognition errors and improve the performance of keyword spotting algorithms. Word boundary detection also helps in reducing the search space in the keyword spotting algorithm. Speech is non-stationary in nature and most of the time no utterance of the same word will be same as another utterance of same word. This makes it challenging to develop any speech processing algorithm. Many algorithms, to detect word boundaries, use acoustic features, lexical cues, energy, pitch etc. Acoustic features of energy, pitch and Teager Energy were used in this research to detect word boundaries. The strengths and drawbacks of each of the techniques are identified and the information from all the techniques was fused to obtain improved word boundary detection. Energy was able to detect word boundaries with 56% of the time, pitch with 68% of the time and Teager Energy with 72% of the time. Simple counting rule, which is based on reinforcement learning, was used in this research to fuse the detections from the three techniques to make a final decision on the word boundaries. This system does not need prior knowledge about the detection and false alarm probabilities of the techniques. The weights are adapted with the outcome in every iteration. Fusion of the decisions from energy, Teager Energy and pitch yielded a 79% hit rate on spontaneous speech using counting rule whereas linear opinion pool and log opinion pool produced 73% and 74% hit rate respectively.