Comparison of different speech enhancement methods on recognition of noisy speech

الملخص EN

The performance of different distance measures, along with different noise reduction techniques, on the LPC based, speaker independent, recognition of noisy speech is investigated.

White as well as colored normally distributed additive noise of different levels was used to corrupt the test utterances.

The data base used to carry out the comparative study consists of the ten Arabic digits 0-9.

The training data were obtained from 800 utterances spoken by 20 different speakers, while for the test, 130 utterances were collected from six additional speakers.

Automatic end point detection based on energy threshold is used for all training and test utterances.

A fixed end point dynamic programming is used both in the clustering and recognition stages.

Results of the study show that significant improvement in recognition of noisy speech is possible by the appropriate choice of distance measure and noise reduction technique