面向机器人的自然人机对话技术研究是指根据智能机器人语音交互的要求，研究人机口语对话系统关键技术，使用户和机器人能比较自然地对话。自然对话是指在用户说话方式、说话内容受限较小的情况下，机器人对话系统能够正确识别用户语音、理解用户的意图、并做出合理应答，使整个对话过程流畅自然。 根据自然人机对话的特点，论文以限定领域连续语音识别和关键词检测技术为基础，重点研究了语言模型训练技术，以提高人机对话系统的识别率。同时总结了儿童语音识别技术的现状，开发了基于ARM平台的命令词识别系统。 论文的主要工作包括以下几个方面： 1.改进语料整理方法，提出了新的语料平衡方法。整理语料库中的大规模宽领域语料和限定领域实际场景语料，用有限状态网络(finite state network, FSN)句法规则扩展限定领域语料。按最终训练语料中关键词概率接近实际场景语料中关键词概率的原则，将已有语料进行平衡混合，提出了实际场景语料和FSN语料的平衡方法。用平衡语料训练的语言模型，使识别器的识别率有了较大提高。 2.总结了儿童语音识别技术的现状。儿童语音识别技术的研究有利于从特征提取和声学模型的角度改进识别器的性能，提高口语对话系统对不同年龄段用户语音的适应能力。论文介绍了儿童语音语料库的建设方法和国内外有代表性的语音语料库、儿童语音的声学特征分析及影响儿童语音识别的其它因素、以及改善儿童语音识别的四种自适应技术。 3.开发了基于ARM平台的命令词识别系统。开发性价比高的语音识别系统是语音识别技术推广的重要课题，嵌入式系统是有效的实现手段，论文设计了ARM开发板启动程序，完成了程序移植工作，并提出了提高系统运行速度的方法。
英文摘要
The research on natural human-machine dialogue technology for intelligent robot is to explore spoken dialogue system’s key technology according to the requirements of the intelligent robot’s speech interaction, and fulfill the natural dialogue between the users and robot. Natural dialogue means the condition that users ask their questions with less limited tongue, and then robot can recognize users’ speech, understand users’ intent and make a reasonable response, thus a natural and fluent dialog progress can be achieved. According to the characteristics of the natural human-machine dialogue, this paper emphasizes the language model training and dialogue management technology to improve the recognition rate and flexibility of man-machine dialogue system based on the limited field continuous speech recognition and keyword spotting. Additionally, the current status of Children’s speech recognition technology and ARM-based command recognition system are investigated in this paper. The main task of this paper include the following: 1.Improving corpus processing and dialogue management methods. The large-scale, wide-ranging field corpus and limited-field real scene corpus are collected from the corpus database. And then the limited-field corpus is expanded by Finite State Network (FSN) grammars. This paper proposes an effective method to retrieve the training corpus from real scene corpus and FSN corpus by comparing of the probabilities of the key words in both limited-field FSN corpus and limited-field real scene corpus. This method balances the two kind of corpus to well interpret the content of limited application field. Experiments prove the high performance of the recognizer with the use of language model trained by the balanced corpus. 2.Summarizing the current Children’s speech recognition techniques. Research on Children’s speech recognition is helpful to improve the performance of speech recognizer from the aspects of feature extraction module and acoustic module, which promotes the adaptation of the dialog system to users with different ages. This approach introduces the Children’s speech and corpus database; the acoustic characteristics of Children’s speech and the factors affecting the Children’s speech recognition; and four main methods of Children’s speech recognition. 3.Developing the ARM platform based command word recognition system. The cheap speech recognition systems for simple applications are an important means to promote the speech recognition technology. This approach develops ARM board initializing program, implants the recognition program, and gives some method to speed up the system operation efficiency.
修改评论