Research On Keyword Spotting Based On RNN Feature Extractor

Voice keyword spotting is one of the important branches in the field of speech recognition.It comprehensively uses the technology of various disciplines such as statistical learning,signal processing and multimedia.It has a very broad field in audio document retrieval,sensitive word monitoring,and smart device wake-up Application prospects.Query-by-example keyword spotting method is one of the main methods under the voice keyword spotting task.This method can re-register the keyword template for the changed keyword or the out-of-set query word without the need to renew the entire model Training has a unique advantage in the task of speech keyword spotting for variable keywords and out-of-set words,and it has gradually attracted more and more researchers' attention.However,how to easily register the new template and improve the accuracy of detection restricts the development of query-by-example keyword spotting.this article will start from these two perspectives to improve the performance of query-by-example keyword spotting.The main work and content of this article consists of the following aspects:1.Introduce the selection and division of data sets,explain in detail how to establish a GMM-HMM-based keyword spotting baseline model,and analyze and explain the experimental results of the baseline model.2.Implemented a keyword spotting model based on the RNN feature extractor,and compared it with the experimental results of the baseline model,and found that no matter whether it is a keyword-only voice sample or a keyword-containing voice sample test,when lacking of phonetic level sample labeling and constructing keyword units with phrases,the keyword spotting model based on the RNN feature extractor have been greatly improved,and the performance of the model under different number of registered templates is also explored.The result is that its performance is almost not affected by the number of template registrations,means it has better ready-to-use.3.Implemented adding the filler model to the RNN feature extractor.On the premise of ensuring that the experimental parameters of the two are the same,it is found that the RNN feature extractor after adding the filler model compares with the test performance of the speech sample containing keywords compared to the unadded The model has achieved a certain degree of improvement.and after adding the filler model,the keyword spotting model based on the RNN extractor is still very good ready-to-use.