In order to separate the audio source from the monaural audio signal the neural network discrimination circuit

摘要

As for the neural network discrimination circuit, before the many option which it knocks down is mixed to the single monaural audio signal the audio source which is unknown is separated, to category the ability to convert is offered. This is piled (depending upon when) to divide the monaural audio signal to the base line frame, the window it indicates the frame, it extracts the many describe feature in each frame, it is achieved the non-alignment which was trained beforehand neural network as the discrimination circuit by adopting. Each as for neural network output, existence of specified type of audio source in each base line frame of the monaural audio signal is shown. The neural network discrimination circuit is suitable although it copes with the reverberation and blockade of time of parameter and the source where the signal and the source change widely and the pile, and the actual signal of frequency domain satisfactorily. For example as for output of the discrimination circuit, (in order like ICA is) to form the many audio channel for source separated algorithm, it is possible to use as front end and, or, it is possible to use as the parameter in after-treatment algorithm for example (, music to category is converted, the source the track/truck is done, way navigation, rimikishingu, security and audit, intended for telephoning and radio communication, and video conference, audio index is formed,).