Research On Keyword Recognition Based On Query By Example

Keyword recognition based on query by example is a significant branch of keyword recognition. By this technology, we can look for keywords in audio document and return speech duration corresponding to queried keyword without textual information. This approach is widely used in low-resource languages now. For the past few years, with the acceleration of internationalization, speech processing in small languages gains the wide interest in research community and becomes a hot issue. In this dissertation, we investigated two technologies on keyword recognition. These technologies can be used for improving the discrimination of feature and making the keyword models more robust and accurate.With the continuous reform of learning methods. Deep Neural Network (DNN) is successfully applied in pattern recognition and receives a lot of attention. In speech recognition, compared with conventional Gaussian Mixture Model-Hidden Markov Model (GMM-HMM), DNN-HMM based on DNN posterior significantly reduces word error rate. The BN feature extracted by DNN with narrow middle layer can improves the performance of GMM-HMM to the level close to that of DNN-HMM. We compared Perceptual Linear Prediction with BN feature to find the more discriminatory feature.In the task of query by example, there is a terrible shortage of examples of keywords. It is vital to efficiently exploit underlying information in examples. In this dissertation, we used HMM framework and investigated three different technologies on model training and validated those performances. In low-resources language, the Maximum a Posterior (MAP) technology is used to improve the performance.It is proved by experiments that these proposed methods used for improving keyword recognition system are viable and effective. And the experiments are all carried on the TIMIT database and a Tibetan database.