Neural Network Models for Speech Recognition in Mobile Environment

Neural networks (NN) are well known for capturing the complex distributions present in data. The ability of the model depends on the structure of the network and the nature of the data used for training [1]. In this paper, we are exploring neural network models for digit recognition in mobile environment. Due to recent advancements and applications in mobile communication, there is a need to develop the speech interface for the mobile devices. The sophisticated speech interface cannot be developed on mobile devices due to the limited availability of computation power, data storage, and power supply (battery). Hence, in mobile environment the speech interface will be realized by distributed computing: that is the speech signal captured by mobile devices is routed to the server at remote location, where the actual speech recognition or synthesis will takes place. Due to limited bandwidth available to mobile channels the speech signal captured by the mobile device is compressed by appropriate speech coder before sending to the remote location [2]. Here, speech recognition is performed on decoded speech. In this paper, we are examining the recognition performance for different coding rates. It is known that the decoded speech is perceptually similar to original speech, but at the signal level it is different from the original speech [2]. In this work, we are exploring neural network models for speech recognition under different coding rates with the hypothesis that these models will capture the perceptual and cognitive information from speech.