A hybrid neural network based speech recognition system for pervasive environments

One of the major drawbacks to using speech as the input to any pervasive environment is the requirement to balance accuracy with the high processing overheads involved. This paper presents an Arabic speech recognition system (called UbiqRec), which address this issue by providing a natural and intuitive way of communicating within ubiquitous environments, while balancing processing time, memory and recognition accuracy. A hybrid approach has been used which incorporates spectrographic information, singular value decomposition, concurrent self-organizing maps (CSOM) and pitch contours for Arabic phoneme recognition. The approach employs separate self-organizing maps (SOM) for each Arabic phoneme joined in parallel to form a CSOM. The performance results confirm that with suitable preprocessing of data, including extraction of distinct power spectral densities (PSD) and singular value decomposition, the training time for CSOM was reduced by 89%. The empirical results also proved that overall recognition accuracy did not fall below 91%.