User-Customized Password Speaker Verification Using Multiple Reference and Background Models
In this paper, we discuss and optimize a HMM-based User-Customized Password Speaker Verification (UCP-SV) system, where users can have their own passwords (with no lexical constraints) after a short enrollment phase involving a few repetitions of the chosen password. In this case, a customer is accepted only if it boththe password it and the voice characteristics are verified ascorrect. Although potentially more ``user-friendly'' and more robust, the development of such a system raises several practical issues, including automatic HMM inference, speaker adaptation, and efficient score normalization, all using a limited (3-5) enrollment utterances. In our case, HMM inference (HMM topology) is based on hybrid HMM/ANN, while the parameters of the model, as well as their adaptation, will use GMM. However, results from our UCP-SV baseline system showed that the main problem lies in the model used for score normalization. Therefore, to circumvent this problem, the main contribution of the paper is the investigation of the use of multiple reference models for customer acoustic modeling and multiple background models for score normalization. In this framework, we have compared several approaches, such as dynamic model selection (DMS) and fusion techniques. Results show that an appropriate selection criteria for customer and score normalization models can improve significantly the UCP-SV performance, making the UCP-SV system significantly more competitive compared to text-dependent SV system.