典型车载场景下语音识别算法效果评价与分析  

Evaluation and Analysis of the Effect of Speech Recognition Algorithms in Typical In-car Scenarios

作　　者：李臻 朱哲慧 张立军[2] 何志祝 宋正河[1] 朱忠祥[1] LI Zhen;ZHU Zhehui;ZHANG Lijun;HE Zhizhu;SONG Zhenghe;ZHU Zhongxiang(National Experimental Teaching Demonstration Center for Mechanical and Agricultural Engineering China Agricultural University,Beijing 100083,China;School of Automobile Studies,Tongji University,Shanghai 201804,China)

机构地区：[1]中国农业大学机械与农业工程国家级实验教学示范中心,北京100083 [2]同济大学汽车学院,上海201804

出　　处：《实验室研究与探索》2022年第1期67-73,共7页Research and Exploration In Laboratory

摘　　要：从语音识别系统前端信号处理出发研究了语音信号特征的提取方法。对4种语音识别主流算法模型进行部署,构建了不同车载应用场景,并选择字错误率和实时率两种主流评价指标对算法模型进行了效果测试,开展了算法评价与分析,同时运用开源二维网格搜索法优化了百度DeepSpeech2模型,并对其进行硬件适配,使得模型识别精度、识别速率得到小幅度提升。此外,基于以上研究结果提出了合理的算法优化建议,为模型算法的稳健性研究提供理论参考。This paper conducts a study on the front-end and back-end of the speech recognition system,studies the method of speech signal feature extraction.Then this paper selects four mainstream algorithm models to realize the deployment and builds different on-board application scenarios.Two mainstream evaluation indexes:word error rate and real-time factor,are chosen to test the algorithm model.The algorithm evaluation and analysis are given based on the test results.The model of Baidu DeepSpeech2 is optimized by using a two-dimensional grid search method with the hardware properly configured.The test results show that the model recognition accuracy and recognition rate are slightly improved.Based on the above results,reasonable algorithm optimization suggestions are put forward to provide practical test results for future research work,in order to enhance the robustness of the model.

关 键 词：语音识别 特征提取 语言模型 信号处理 

分 类 号：TN912.34[电子电信—通信与信息系统]