Abstract: Most modern people ignore the importance of reading.However,for children aged 5~12,reading aloud is not only an essential skill in the learning process,but also an effective means of cultivating sentiment.Since there is a nonlinear relationship between the characteristics of the spoken speech signal and the evaluation criteria,the recurrent neural network is suitable for time series prediction,but its prediction effect is limited for long-term span.According to the characteristics of children’s spoken speech and its evaluation system,a new model combining Deep Speech and three-la-yer LSTM (Long Short-Term Memory) neural network was designed.Firstly,on the basis of adding attention mechanism,the accuracy and fluency measure of speech evaluation are put forward,and the spectrum map is used as the input of feature extraction.Among them,the accuracy of reading uses the new version of Deep Speech to improve the accuracy of phoneme recognition.For fluency evaluation,the spectrogram is sent to the three-layer LSTM model to present the effects of the time series.Then,the results are sent to the attention mechanism for weight adjustment,and finally the total evaluation results are used for the evaluation of children's spoken speech.The experiment uses the children’s reading corpus,which is provided by the “export chapter” software,and the experimental environment uses the TensorFlow platform.The experimental results show that compared with the traditional model,this model can accurately judge the correctness of spoken speech and the fluency of reading aloud,and the scoring results obtained by its evaluation model are more accurate.

Key words: Attention mechanism, DeepSpeech, Evaluation of spoken speech models, Long Short-Term Memory, Spectrogram