Research On Keyword Spotting Using Neural Architecture Search

With the development of the Internet of Things and artificial intelligence,keyword spotting is gradually being brought to people's lives in many scenarios such as smart homes,car systems,and voice assistants.As a special speech recognition,keyword spotting not only needs to achieve accurate recognition,but also has many restrictions on the memory,computational complexity,power consumption and latency.Compared with traditional methods,the deep learning-based keyword spotting model has a greater advantage in recognition accuracy.The existing deep model is difficult to achieve a balance among multiple performance indicators,and still has a large room for improvement in recognition accuracy;the artificially designed neural network architecture depends on the researcher's experience and professional knowledge,and often takes a lot of resources,meanwhile the depth of designed model cannot be flexibly adjusted according to different application environments of keyword spotting.In order to solve the above problems,this paper studies the keyword spotting method based on neural architecture search,the main research work is as follows:1.A keyword spotting method based on differentiable architecture search has been proposed.This method uses a differentiable architecture search to automatically construct a convolutional neural network for keyword spotting,and uses a hardwareaware cost function to better balance the recognition accuracy and computation cost of the model.Finally,the searched convolution cell can be used to flexibly construct keyword spotting models of different depths.The experimental results on the Speech Commands dataset show that this method reduces the word error rate by 19.4% ~ 68.8% compared with the existing keyword spotting model based on convolutional neural network,and the computation cost required is relatively low.2.A keyword spotting method based on attention mechanism is proposed for continuous audio stream scene.This method uses the differentiable architecture search method to search and construct a more effective recurrent neural network,and uses the attention mechanism to pay different attention to different regions in the audio stream, further improving the model's ability to recognize keywords.Finally,the experimental results on the continuous audio stream show that,compared with the state-of-the-art model based on recurrent neural network,the false rejection rate at 1 FA / hour is reduced by 8.8%,and the parameters is reduced by 11.3%.This article also explores cross-layer attention mechanism,which can further reduce the rate of false rejection rate for keyword spotting when the model parameters are basically unchanged.