Pattern recognition system
AssigneeKabushiki Kaisha Toshiba
InventorHideo Segawa
Life of this patent
Estimating the term…
Abstract
Multiple similarities between phoneme feature data extracted from a word input signal and reference pattern data read out from a reference pattern memory are calculated by a similarity calculation section. The multiple similarities are converted by a conversion filter into posterior probabilities using parameter sets which are read out from a filter parameter bank in accordance with the obtained multiple similarities. The posterior probabilities are subjected to matching with data stored in word dictionary by a word matching section, thus discriminating the input word.
Claims6
A pattern recognition system comprising:
- a feature extraction means for extracting feature data of an input pattern;
- a memory for storing a reference pattern of data each category;
- a multiple similarity calculation section for calculating multiple similarities between the feature data extracted by said feature extraction means and the reference patterns stored in said memory;
- posterior probability conversion means for converting the multiple similarities obtained by said similarity calculation section into posterior probabilities and
- means for performing matching between the posterior probabilities and data stored in a dictionary so as to obtain a pattern recognition output;
- wherein said posterior probability conversion means includes:
- a parameter table storing parameter sets each composed of a plurality of parameter pairs of .alpha. and .beta., said parameter .alpha. indicating a number of components which are not involved in a subspace spanned by a reference pattern, the parameter .beta. indicating a variance in a direction of non-reference subspace, and both parameters .alpha. and .beta. being estimated from multiple similarity values between a reference pattern and training data, for each category of learning patterns, and for every category of reference patterns;
- and means for calculating the posterior probabilities based on the multiple similarities calculated by said multiple similarity calculation section and the parameter sets stored in said parameter table.
A pattern recognition system comprising:
- means for extracting feature data of an input pattern;
- means for calculating mltiple similarities Si (i=1 to n) between reference patterns and the input pattern;
- means for obtaining predetermined parameter sets (.alpha.,.beta.) in accordance with the resultant similarities Si;
- means for calculating a posterior probability represented by:
- ##EQU10## where Ci is an input category, and P(Ci) and P(Cj) are respectively probabilities that the input category is categories Ci and Cj;
- and
- means for performing matching between the posterior probabilities and data stored in a dictionary so as to obtain a pattern recognition output.
Dependencies are inferred from the claim text.