Speech recognition system consisting of auditory feature extracting cells and velocity-controlled delay-lines part II. Recognition module

• Main Authors: Yamauchi, Kouichirou, Fukuda, Megumu, Fukushima, Kunihiko
• Format: Conference Proceeding
• Language: English

A neural network model for speech recognition is proposed, based on neurophysiological findings of the auditory system. The first stage of the system is a feature extracting module, which extracts Constant-Frequency, FM-Ascending and FM-Descending components. The extracted features are sent to the next stage, the recognition module. The recognition module consists of three blocks of multi-layered networks (Fig 1): the upper, middle, and lower blocks. Each block is a neocognitron-like network whose first layer consists of velocity-controlled delay-lines. The propagation velocities of the delay-lines of the upper and lower blocks are faster and slower, respectively, than those of the middle one. The propagation velocities of these three delay-lines are variable but the ratio of the velocities is always kept constant. In order to realize time-distortion invariance, this system controls the velocity of all of the delay-lines in such a way that the durations of features on the delay-line of the middle network are the same as the duration of features of a training pattern. This is accomplished by comparing the outputs of the upper and lower blocks. In the simulation, the system was trained using some Japanese words. After the training was completed, speech signals which were spoken at various speeds were presented to the system. The system recognized each of the words correctly without being affected by their speeds.