In search for the relevant parameters for speaker independent speech recognition

One of the problems with speaker-independent speech recognition is the huge amount of training data required, which implies a high cost. The performance of a discrete density hidden-Markov-model speaker-independent speech recognition system when using a small set of examples for training is investigated. By using LPC (linear prediction coding)-based analysis, an approximately 12% error rate was obtained on a highly confusable telephone-quality vocabulary. Using RASTA PLP analysis, a 4% error rate can be achieved. The reason for this improvement is that the RASTA filter filters out the convolutional noise of the different telephone lines and PLP analysis suppresses speaker-dependent details. RASTA filtering was also tried out on the LPC cepstra and gives, with a higher model order, the same results as RASTA PLP.<>