Speech recognition system using modifiable recognition threshold to reduce the size of the pruning tree

A speech recognition system comprising:

• a speech analyzing unit for analyzing an input speech waveform to extract feature quantities and converting the extracted feature quantities to an input pattern;
• a reference pattern memory for storing reference patterns of words to be recognized;
• a temporary pattern memory for storing partial reference patterns as subjects of similarity computation with the input pattern among the stored reference patterns, and for storing cumulative similarities for each of the partial reference patterns, the cumulative similarities representing a measure of similarity between the partial reference patterns and the input pattern;
• a threshold memory for storing at least one threshold value;
• a counting unit for counting a number of partial reference patterns among the stored partial reference patterns that have corresponding cumulative similarities that exceed the at least one threshold value by comparing the corresponding cumulative similarities with the at least one threshold value;
• a threshold computing unit for computing a threshold of pruning which gives a predetermined number of partial reference patterns according to a correspondence relation between the predetermined number of partial reference patterns and the threshold of pruning;
• a similarity computing unit for computing acoustic feature similarities with respect to the extracted feature quantities of the input speech for each of the predetermined number of partial reference patterns;
• and
• a determining unit for determining, as a result of recognition, a word corresponding to one of the predetermined number of partial reference patterns having a highest value of the computed acoustic feature similarities.