| By analyzing the distinctive characteristics of an individual's voice, it is possible to uniquely identify one person from another. The field of speaker recognition, or voice biometrics, continues to advance as security threats in business, government, and law enforcement persist. The development of a high-level, robust, software-based speaker identification system was undertaken successfully in this work. This next generation speaker identification system focuses on two primary aspects: (1) incorporating text-independence and (2) using short sample speech utterances for performance evaluations. Using high-quality clean speech samples and degraded telephone quality speech, the identification accuracy remained high even with short sample speech utterance lengths. By employing both noise reduction and channel compensation techniques with noisy telephone grade speech, system accuracy was significantly improved.;The system used Gaussian Mixture Model (GMM) classification and performance evaluations were completed using commercially available speech databases. Two probability based evaluation techniques were used in conjunction with the GMM to determine identification accuracy, both with distinct strengths and weaknesses depending on the dataset. For conversational clean speech utterances, an identification accuracy of 97.6% was attainable with the GMM based approach for an utterance length of 5 seconds. Performance decreased slightly to 90.4% and 74.9% for degraded and severely degraded telephone speech, respectively, with 5 second utterance lengths. In comparison to existing implementations, substantial improvements in identification accuracy are attained with the system developed in this work. Overall, the software prototype designed in this project simulates a real-time implementation which will spark future development and implementation of a commercial grade product. |
