Mobile phone speaker recognition application

Abstract

Today, smartphone are able to handle confidential matters such as online banking, credit/debit card purchases which has cause security breaches. Privacy has become a challenging issue to uphold. The most common protection design of a smartphone are password protection and swiping pattern protection. One possible solution to the problem is to implement biometric system using speaker recognition system which allow the system to identify user’s identify by using his/her voice biometric. The objective of this project is to develop a speaker verification system on android platform. The android application is used to verify speaker’s utterance against a trained user model, which was adapted from the GMM-UBM. The application consists of several functions such as recording of speech, replaying of speech, adapting user model from UBM and performing likelihood calculation to accept or reject the identity. The system first extract features of the speech in the front end processing. These features are used to determine the likelihood of the user. There are two approaches to estimate the result - The first approach is to use other speaker models to cover all alternative hypotheses and the second approach is to train a single model using pool speech from several speakers. This is also known as UBM. This approach allows us to train UBM once and use it for every hypotheses speaker. Experimental results showed that the basic requirement to make the system a more reliable and accurate system, the system required 3 aspect of principal – there are a good mixture size, more speaker and utterance per speaker and lastly more adaptation. Adaptation help in making the system more accurate, as files are used to adapt the model, the model has more information on the parameters of the user thus yielding better result in performance.