Study On Methods Of Text-dependent Speaker Recognition

As a special form of speech recognition, Speaker recognition is one of the current researching hotspots in speech signal processing technology. It is of important practical value for identity authentication in the realm of confidential ground and safe-precaution field, etc.Two kinds of system models was introduced according to the physical generating process of speech signal as well as human's hearing characteristics. Furthermore, several prevalent speaker feature parameters were analysed and extracted. Secondly, taking separately Continuous Hidden Markov Model (CHMM) and BP net for example, two different styles of model-building theory and recognition strategy about Hidden Markov Models(HMM) and Artificial Neural Network(ANN) were also expounded here, some certain applied problems of theirs in speaker recognition were discussed in succession. Two different sorts of speaker models were constructed and the corresponding arithmetics were accomplished via programming. Thirdly, a new combined form of feature parameters was attempted in the thesis, and reasonably compounding Characteristic parameters has been proved helpful in improving the correct recognition rate. Finally, without increasing difficulty and complexity on ANN's design and structure, a novel feature parameter processing scheme was proposed and applied, which resolved the problem effectively on how to pass large scale's multi-dimensionalmulti-frames characteristic parameters to ANN for training and ensure its high recognition performance at the same time. The scheme reduced the complexity of ANN's structure and solved some practical problems on net design by adopting varieties of each dimension of the feature parameter to reflect the change of voice characteristics of the speaker.In the simulation test using Mel cepstrum as well as its difference as combined feature parameter, a BP network has been established which consumed 8s during the course of training by applying the scheme mentioned above. And in recognition process, ten untrained samples extracted randomly were all distinguished correctly within about 10.3 seconds. The result has proved that the processing scheme for feature parameter can availably reduce the scale and complexity of net structure and provide a solution for long-time sound samples to input its characteristic parameters to ANN for effectual training.