Speaker Recognition Using Phoneme Specific Sentences.
Final rept. Oct 84-Sep 85,
NAVAL RESEARCH LAB WASHINGTON DC
Pagination or Media Count:
Listener tests involving the ability to distinguish between previously unfamiliar voices were conducted with phoneme specific sentences each sentence contains only certain classes of consonant phonemes using a familiarization-test procedure. There were three test conditions comparing unprocessed speech Linear Predictive Coding LPC processed speech, and speech at 800 bitssec. The results suggest that for unprocessed speech, speakers are better recognized when speaking sentences that contain voice fricatives or voiced or unvoiced stops and are not as well recognized if the sentences contains only glides or only nasals. On the other hand, sentences with voice fricatives and nasals were best for LPC speech. The results were also highly dependent on the grouping of the speakers. The recognition of more distinctive male voices and of female voices went down with LPC and 800 bitsec processing as expected, but the recognition of less distinctive males was no worse after processing than before. The fact that the effects of voice processing vary with the composition of the speaker set is discouraging to the development of a standardized test of speaker recognition. Speaker recognition with the 800 bits algorithm was very poor but performance was still better than pure guessing. Keywords Voice communications. Author
- Voice Communications