NOISE-ROBUST OPEN-SET SPEAKER RECOGNITION USING NOISE-DEPENDENT
Gaussian Mixture Classifier · 2002
Speaker recognition makes a dec ision to either ac cept or reject a recognized speaker candidate, based on some score (e.g. likeli hood) associated to t he item. Model-based classification can be used to make the decisi on. In mobile device applications, the back ground noise level may affect the score distributions and cause a decision failure. We describe a new decisio n p rocedure. which treats the scores as the outcome of Gaussian mixture distributi ons, where mean and covari an ce parameters are modeled as polynomial functio ns of noise level. We evaluate the procedure on a speaker recogni tion task in a mobile and noisy environment, using a hands -free microphone. Experiments show that the system delivers an equal error rate of o.�o.%, O.80� and 3.53% for parked, stop-and-go and highway drivmg conditions. The method maintains a balance between false acceptance and false rejection under all driving condit ions, making any empirical threshold adjustment unnecessary.