Continuous speech recognition by sentence spotting applying the Vector Continuous Dynamic-programming to output of Partial Matching Method

This paper proposes a speaker-independent continuous speech recognition system on sentence spotting. The system is composed of the following three parts: (1) construction of three-connected demi-phoneme standard patterns; (2) description of sentence set by demi-phoneme network; and (3) sentence spotting by ‘Partial Matching Method’ and ‘Vector Continuous DP’. A sentence spotting recognition experiment was conducted in which each of 10 open speakers uttered 11 sentences (110 sentences in total, which form a task of the average branching factor of 4.1 and the number of words 113). The sentence recognition rate of 76.4 percent was obtained (word recognition rate of 94.5 percent in the sentence). In the proposed system, three connected demiphoneme standard patterns are constructed using only 492 kinds of balanced word sets, and no training by the continuous speech is applied. The complete frame synchronization is established in the system from the waveform analysis to the sentence spotting.