低资源连续语音识别中的集外词处理技术

Out-of-Vocabulary(OOV) words processing is an important part of continuous speech recognition system. State-of-the-art methods face many problems due to the little understanding of formation mechanism of out-of-vocabulary(OOV) words as well as non-jointing modeling in acoustic and semantic units. It is acknowledged that each OOV word occurring in the speech signal translates to multiple errors in the.recognition output which can raise the confusability among the in-vocabulary (IV) words, and OOV detection and discovery accuracies are still low. In low amount of speech and text data or in highly agglutinative languages which exhibit large amounts of new words from inflection and compounding, the OOV problem is more serious. So the project is focused on the OOV processing for low resource and multilingual speech recognition. On the basis of understanding of pronunciation week modeling and heterogeneity for OOV, the denotation mechanism of.acoustic pronunciation unit and language knowledge is studied in depth. The main research contents are as follows, multiple-unit denotation model of OOV and pronunciation lexicon with new word leaning, hybrid language modeling and its robust parameter estimation, OOV detection and recovery, external data utilizing strategy and semi-supervised learning. Through the above research, the performance.of OOV detection and OOV recovery under low resource condition will be improved Greatly. This research is promising in overcoming the bottleneck problems of low resource continuous speech recognition, enriching the speech recognition theories so as to provide an effective means in deep-seated, multi-level cognition of speech signals in low-resource and multilingual speech recognition.

集外词处理技术是连续语音识别系统的重要组成部分，但当前主流的方法缺乏集外词形成机理的认知，缺少声学、语义单元的联合建模，因此集内词错误扩散效应明显、集外词检测与恢复能力有限。尤其在低资源语言的语音与文本数据资源严重缺乏、词缀变化产生大量新词时，集外词问题更为严峻。本项目针对实际应用数据缺乏条件下的多语言语音识别的集外词问题，从集外词的发音不确定性、弱模型性和异质性出发，深入研究发音单元的声学特征空间和语言单元有效表征机理，构建适用于低资源条件下的新型集外词处理方法。研究内容包括：集外词的多单元表示模型与字典学习算法；联合语言模型及其稳健性参数估计算法；基于集成学习的集外词的检测与恢复方法；集外语音本文数据利用策略及训练方法。通过以上研究成果，大大提升低资源条件下语音识别系统集外词检测和恢复性能，克服现有低资源连续语音识别面临的瓶颈问题，为深层次、多角度感知语音信号提供新的有效手段。

集外词处理技术是连续语音识别系统的重要组成部分，但当前主流的方法缺乏集外词形成机理的认知，缺少声学、语义单元的联合建模，因此集内词错误扩散效应明显、集外词检测与恢复能力有限。尤其在低资源语言的语音与文本数据资源严重缺乏、词缀变化产生大量新词时，集外词问题更为严峻。本项目针对实际应用数据缺乏条件下的多语言语音识别的集外词问题，从集外词的发音不确定性、弱模型性和异质性出发，深入研究发音单元的声学特征空间和语言单元有效表征机理，构建适用于低资源条件下的新型集外词处理方法。研究内容包括：（1）集外词的多单元表示模型与字典学习算法。将集外词表示为音节、词素和字母音素对，然后构建混合字典；（2）联合语言模型及其稳健性参数估计算法。构建包含词和子词的混合语言模型，然后利用融合策略或者混合语料进行精确参数估计；（3）基于集成学习的集外词的检测与恢复方法。利用前面的不同单元的模型对集外词的位置进行精确估计，并采用子词单元组合或转换策略策略进行集外词恢复。（4）集外语音本文数据利用策略及训练方法。通过以上研究成果，大大提升低资源条件下语音识别系统集外词检测和恢复性能，克服现有低资源连续语音识别面临的瓶颈问题，为深层次、多角度感知语音信号提供新的有效手段。