Research Of User Customized Password Voiceprint Recognition System And Embedded Implementation

The technology of voiceprint recognition(speaker recognition) has the advantages of convenience, benefit and accuracy. It is used more and more popular in daily life. Voiceprint recognition is a kind of authentication based on biological features. According to the text content, it can be divided to the types of text-dependent and text independent.In this work, we focus on User Customized Password (UCP) voiceprint recognition system. We have implemented it on PDA, which includes voice collection, model training, and voiceprint verification.In this work, we have proposed three novel methods:1. HMM recognition methods based on model distance. Model distance stands for the small difference between speakers. First, we train the speaker models in advance, and then carry out the verification process using model distance. This method has been adopted successfully in Forensic Authentication.2. Dynamic threshold for the verification. During the verification process, different people may have different thresholds. In this work, we adopt dynamic threshold and train different thresholds for different speakers.3. Two-level decision mechanism based on HMM-UBM. In this method, the first decision is made according to the UBM. After the first step, the speaker is located in the certain or uncertain space. If the speaker is in the uncertain space, then the second decision will be conducted based on model distance according to dynamic threshold.For the hardware platform, personal computer, Sunplus SPCE061A and PDA are all used. We have successfully developed embedded speech recognition system in SPCE061A and have applied it in practice. However, when developing speaker recognition on this platform, the system doesn't perform ideally for its limitation of storage and calculation. To have a better performance, we design and implement the speaker verification system based on PDA. Several new methods are applied to improve the verification performance, such as the transformation from floating-point operation to fixed-point operation, pretreatment of the calculation, fast and approximate mathematic calculation etc. These methods guarantee the real-time performance of embedded system.