Research On Aduio Keyword Spotting Technology Based On Neural Network

Keyword spotting(KWS)is a rising and popular technology which could detect specific speech content from the current environment.Recently,it'has been proved that neural network has been successfully applied in various fileds,which also promoted the development of KWS significantly.Although the state-of-the-art KWS technology could achieve an extremely high recognition accuracy on pre-defined keywords,they can't support users to define keywords,or the accuracy of user-defined keywords is extremely low.To solve the above issues,a novel two-stage training technique that combines pretraining and incremental training is proposed in the paper.Firstly,the traditional neural network-based KWS model is improved with the technique of Query-by-Example in the paper,which make the traditional neural network support user-defined keywords.Secondly,in order to make a significant improvement on the recognition accuracy of userdefined keywords,a fine-grained classification traning method is proposed,which improve the discrimination ability of the network.Thirdly,In order to make the model learn the information of user-defined keywords,an incremental training method based on end-to-end loss function is proposed in the paper,supplemented by data augmentation,which makes the model match with user-defined keywords better.In the paper,a classical model cnn-trad-fpool3 and the best state-of-the-art model res15 are selected as the baseline.Experiments show that the accuracy of user-defined keywords on cnn-trad-fpool3 and res15 is increased by 21.78% and 24.42% respectively with the proposed method.Finally,a KWS system is built on PC platform and Nvidia Jetson TX2 embedded platform in the paper.Besides,the proposed methods are also implemented in the KWS system and the whole calculation process can be implemented offline.The KWS system is not only able to detect the 10 pre-defined English keywords,but also support users to defined their own keywords,which satisfies the personalized requirements of users and avoids the disclouse of their privacy.