Fusion approach for improving the performance in voice-biometrics

Voice biometrics, also called speaker recognition, is the process of determining who spoke in a recorded utterance. This technique is widely used in many areas e.g., access management, access control, and forensic detection. On the constraint of the sole feature as input pattern, either low level acoustic feature e.g., Mel Frequency Cepstral Coefficients, Linear Predictive Coefficients or high level feature, e.g., phonetic, voice biometrics have been researched over several decades in the community of speech recognition including many sophisticated approaches, e.g., Gaussian Mixture Model, Hidden Markov Model, Support Vector Machine etc. However, a bottleneck to improve performance came into the existence by only using one kind of features. In order to break through it, the fusion approach is introduced into voice biometrics. The objective of this paper is to show the rationale behind of using fusion methods. At the point of view of biometrics, it systematically classifies the existing approaches into three fusion levels, feature level, matching-score level, and decision-making level. After descriptions of the fundamental basis, each level fusion technique will be described. Then several experimental results will be presented to show the effectiveness of the performance of the fusion techniques.