NIST Speaker Recognition Challenge Shows Improvement
Results of the 2014 Speaker Recognition i-vector Machine Learning Challenge, held by the National Institute of Standards and Technology (NIST), indicated improved accuracy levels since the last NIST Speaker Recognition Evaluation in 2012.
The NIST challenge focused on the development of new methods using i-vectors for speaker detection in the context of conversational telephone speech. Results were analyzed by comparing more than 12 million pairs of voice tracks to validate the identity of the speakers biometrics.
Initial results indicate that the leading system achieved a relative improvement of approximately 38 percent over the baseline system. Meanwhile, approximately 75 percent of challenge participants submitted a system that outperformed the baseline.
Encouragingly, there were more than twice the number of participants and nearly twice as many systems submitted for evaluation this year.
The challenge was intended to foster interest in biometric speaker recognition by the machine learning community. By basing the tests on i-vectors (widely used by state-of-the-art speaker recognition systems) rather than audio data, the evaluation was intended to be readily accessible to participants from outside the audio processing field.
The goals of the challenge included the following:
- exploring new ideas in machine learning for use in speaker recognition;
- making the speaker recognition field accessible to more participants from the machine learning community;
- improving the performance of speaker recognition technology.
One biometric supplier that took part was SpeechPro. Its VoiceKey technology was tested along with more than 100 submissions from leading companies, universities, and laboratories around the globe, including Raytheon, IBM and 3M.
"There's been an increased interest in the use of voice biometrics and speaker recognition for law enforcement agencies as well as for commercial applications in banking or call-center industries in recent years," said SpeechPro President Alexey Khitrov, in a statement. "Discovering the true identity of a voice offers an unprecedented opportunity to protect access to information as well as secure it."