Voice Identity Recognition Based on the Parameters of the Spectral Voice Source Model

The information content of the parameters of a spectral voice source model in an automatic voice identity recognition problem is studied. For the voice parameters, the identity recognition error was 20.8%; using these parameters together with the pitch period reduced the error to 13.8%. Lastly, the combined use of the spectral model parameters with the pitch period and mel-frequency cepstral coefficients provided the highest accuracy (the recognition error was 1.2%).