Speaker Recognition on Lossy Compressed Speech using the Speex Codec

摘要

This paper examines the impact of lossy speech coding withSpeex on GMM-UBM speaker recognition (SR). Audio from120 speakers was compressed with Speex into twelve datasets, each with a different level of compression quality from 0(most compressed) to 10 (least), plus uncompressed.Experiments looked at performance under matched andmismatched compression conditions, using modelsconditioned for the coded environment, and Speex codingapplied to improving SR performance on other coders. Resultsshow that Speex is effective for compression of data used inSR and that Speex coding can improve performance on datacompressed by the GSM codec.