Speech Parameter Extraction by a Cooperative Feature & Pattern Selector Approach.
Oscar Mayora, Francesco Curatelli, L. Motta · IIA/SOCO · 1999
This paper proposes an hybrid method for efficient feature and pattern selection for speech recognition. First a signal processing tool, segments and extracts a large number of speech parameters (LPC, cepstral, PLP, formants, derivatives) from a multiple speaker digit-database in Italian language. Then, the resulting codified phones are input to a data reduction algorithm for reducing in both, feature and pattern dimensions. The feature selection stage is based on a statistical approach that eliminates the lower covariance feature fields. The pattern reduction takes place with a multidimensional geometric approach that selects the most relevant patterns of each class. The stop condition for both types of reductions, is controlled by a preset tolerance error present for speech recognition evaluation. Experimental results have shown reductions up to 65% of original data set dimensions maintaining word recognition rates above a 92% level.