Rikard Neiderud 

                                    
Speaker Recognition in /k/


Abstract:

Speaker recognition is the process of  recognizing people by their
voices. Automatic systems often use some form of spectral-based
technique to represent each speaker, one of the more successful being
the Mel-cepstrum. In this thesis, speaker recognition in the phoneme
/k/ is studied by means of Mel-cepstral coefficients based on several
non-parametric spectral estimators. Two studies are made.

In the first, recordings of 8 speakers uttering /aka/ are used in a
speaker identification comparing the performance of the traditional
spectral estimators using a single window, such as the Hamming window,
and multiple window methods. The results indicate a significant
increase in performance using the latter.

The objective of the second study was to see if the imitations made by
professional imitators of their target speakers were mistakenly
identified as the targets, using speaker recognition based on /k/. The
results here suggest that /k/ may be quite robust against voice
disguise in contrast to many other phonemes.