The collection of telephone databases, for training speech recognizers, is a time consuming and costly work. In the paper we propose a method for producing simulated telephone data starting from clean wide band databases. The result of the simulation is the generation of a noisy database that can be used, in addition to other techniques, for compensating or adapting speech recognizer parameters with respect to different test environments. For the first of the two adopted test sets, performance improvements ranging from about 30% to about 9% have been measured, as a function of the quantity of real telephone data used. in addition to the simulated ones, for system training. For the second test set no significant improvements were obtained
Use of Simulated Data for Robust Telephone Speech Recognition
Falavigna, Giuseppe Daniele;Gretter, Roberto;Orlandi, Marco
1999-01-01
Abstract
The collection of telephone databases, for training speech recognizers, is a time consuming and costly work. In the paper we propose a method for producing simulated telephone data starting from clean wide band databases. The result of the simulation is the generation of a noisy database that can be used, in addition to other techniques, for compensating or adapting speech recognizer parameters with respect to different test environments. For the first of the two adopted test sets, performance improvements ranging from about 30% to about 9% have been measured, as a function of the quantity of real telephone data used. in addition to the simulated ones, for system training. For the second test set no significant improvements were obtainedI documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
