A study on speech feature extraction and application in mandarin LVCSR

Noise is a pivotal factor that reduces recognition accuracy of a speech recognition system. So how to extract effective speech characters is very important for a speech recognition system to increase accuracy. The paper analyses speech feature extraction and improves it. The experiments indicate that the algorithm combination LDA+MLLT+CMS has a better robustness than other combinations. Average syllable recognition rate reach 43.88% by using it in conditions with noise. The algorithm combination has also a good performance in Mandarin large vocabulary continuous speech recognition (LVCSR). Syllable recognition accuracy achieves 83.68%. Therefore the combination has a good effect on speech recognition system.