Neuro-fuzzy filtering techniques for automatic speech recognition enhancement

In order to enhance the percentage of recognised words for ASR (automatic speech recognition) systems, an adaptive neuro-fuzzy filtering technique in the time domain is proposed. Architecture of neural networks is shown having the purpose of exploiting spatial information, with the advantage of little computational load with respect to other classical beamforming techniques. Several results are described for complex acoustic scenarios with VOX ASR system by DIBE (University of Genova).