基于嵌入式平台与DNN-HMM的中文儿童语音能力评估研究

Research on the Evaluation of Children's Chinese Speech Ability Based on Embedded Platform and DNN-HMM

儿童语音能力评估对提高其语言发展规律的认识,促进儿童语言、认知和社交能力的全面发展有重要意义。利用嵌入式硬件平台及深度神经网络隐马尔可夫模型(DNN-HMM)开展中文儿童语音能力评估研究。首先,使用LD3320语音芯片设计嵌入式硬件平台,然后利用线性校正单元构建深度神经网络,利用构建的DNN-HMM模型对中文儿童语音数据进行训练与测试,并结合一般内容概率潜在语义分析(GC-PLSA)模型开展语音评分,最后,将模型移植到嵌入式平台进行语音能力评估实验。实验结果表明:与传统高斯混合、隐马尔可夫模型(GMM-HMM)相比,基于DNN-HMM模型在儿童长、短句连续语音词错率(WER)方面均降低约5.4%、5.6%,且DNN-HMM模型获得的平均得分也要高于GMM-HMM模型。

The assessment of children's speech ability is of the great significance in improving their understanding oflanguage development patterns and promoting the comprehensive development of children's language,cognitive,andsocial abilities. Using embedded hardware platforms and Deep Neural Network Hidden Markov Models(DNN-HMM)to conduct research on the evaluation of children's Chinese speech ability. Firstly,an embedded hardware platform isdesigned using the LD3320 speech chip. Then,a deep neural network is constructed using linear correction units. Theconstructed DNN-HMM is used to train and test children's Chinese speech data,and combined with the General Content Probabilistic Latent Semantic Analysis(GC-PLSA)model for speech scoring. Finally,the model is transplantedto the embedded platform for speech ability evaluation experiments. The experimental results show that comparedwith the traditional Gaussian mixture hidden Markov Model (GMM-HMM),the DNN-HMM model reduces theword error rate(WER)of children's long and short sentence continuous speech by about 5.4% and 5.6%,and the average score obtained by the DNN-HMM model is also higher than that of the GMM-HMM model.