Linear input network for neural network automata model adaptation

Abstract

The paper describes an experimental investigation of the applicability of linear input networks (LIN) as a channel and noise adaptation technique for an application of the Loquendo neural network based speech recognizer in a car environment. The considered application is an automated call center that provides traffic information through a voice dialogue system. The connection to the call center is achieved by means of a commercial device placed in the car and made up of a microphone which is placed in front of the driver and equipped with an echo canceller and built-in noise reduction. The connection with the call center is set up through a GSM link. By experiment, the LIN technique adapts the basic neural network speech recognizer to this new environment. Some variants devoted to reducing the number of estimated parameters are also introduced. The LIN technique, is also compared with some classical denoising techniques based on noise spectral subtraction. The obtained results confirm the validity of LIN for channel and noise adaptation, while the introduced variants are a valid alternative when a reduced model size is important. The best performances in our specific application were of 57.14% error reduction versus the performance obtained by general acoustic models and were achieved by joint use of a LIN and noise spectral subtraction.