Voice controlled Urdu Interface using isolated and continuous speech recognizer

摘要

We propose multilayer feed forward Neural Network for Urdu isolated words and Urdu continuous speech recognition. Our application not only recognizes the voice commands but also converts continuous speech into Urdu Fonts. Our work is based on converting analog signal to digital signals, removing silence and noise through filters, extracting 39 Mel Frequency Cepstral Coefficients(MFCCs) from the speech signal, applying k-mean algorithm and using them as an input to multi-layer perceptron neural networks. Urdu text stores in “Jameel Noori Nastaleeq” fonts. Isolated words gain accuracy of 95.6% in speaker dependent mode while 30–40% in speaker independent mode. Continuous speech also shows a good accuracy of 96.5% in speaker dependent mode and 30–40% for speaker independent mode. Current work is being extended to improve accuracy for speaker independent system by improving dataset and feature set.