Real time speech processing to eliminate slamdowns in digital voice systems

Digital voice messaging systems are being used extensively by mobile phone users to store and pass on messages to other users. However with the customer base increasing at an alarming rate, the number of stored calls is increasing dramatically and the number of empty messages, or slamdowns as they are known in the trade, is becoming problematic. A method of detecting these empty messages so that they can be deleted before they are stored was required. Algorithms were developed by using real data recorded on a `live´ system and assessing their performance. Having extracted suitable coefficients from segments of the data using standard digital processing techniques, a multiperceptron artificial neural network was used to produce a logical speech/nospeech result. The network was optimised by investigating aspects such as the training data, the test data, number of hidden nodes, the number of passes and number of segment features being used. The results were biased in favour of speech and the measured success rate for unseen data is in excess of 95%. The software based solution is suitable for the current digital voice messaging system but it was decided to develop a stand alone ASIC based solution for future products. It has been necessary to further optimise the parameter extraction routines and the neural network to simplify the hardware functions required. The final solution uses integer arithmetic and an EPROM to store the neural network coefficients and a sigmoid function lookup table