Applications of generalized radial basis functions in speaker normalization and identification

The paper describes two applications of radial basis function networks to automatic speech recognition. We used local basis networks of elliptical kernels of different functional form, with recursive allocation of units and on-line optimization of parameters (GRAN model). In the first application, the neural network is used as a front end of a continuous speech speaker-dependent recognition system to normalize the input data from new speakers. With a limited amount of new acoustic data, the recognition error of phone units from the Italian speech corpus APASCI is decreased with an adaptability ratio of 25%. The same model has also been applied in a speaker identification task on a database collected at IRST consisting of isolated digits. An identification error rate of 17% has been obtained on the whole database (50 speakers).