摘要
定量地分析与评价经典隐马尔可夫模型（Ｈｉｄｄｅｎ Ｍａｒｋｏｖ Ｍｏｄｅｌ，ＨＭＭ）的性能，是汉语连续语音识别研究中尚未解决并且亟需解决的问题。文章构造了基于经典ＨＭＭ模型的汉语连续语音识别系统。针对语音单元和输出概率这两个自由度上的各种组合，研究了经典ＨＭＭ模型的复杂度、稳健性、精确性与训练集合的数据量、训练时间、解码效率等特性之间的关系；并且通过实验分析了多候选的构造和剪枝的意义。该文构造的系统与具有国内最高水平的 ＴＨＥＥＳＰ系统的识别率相当，所得实验结果和结论为汉语语音识别的深入研究提供了必要的参考和依据。
In the area of mandarin continuous speech recognition,there has always been an urgent but unaddressed task to quantitatively analyze and evaluate the performance of classical Hidden Markov Model(HMM).A baseline system based on classical HMM is implemented in this paper.The relations between complexity, reliability, ac curacy of HMMs and training data sufficiency,training time,decoding efficiency are studied under various combinations of the two axes, choice of acoustic units and output pdfs form of HMMs.N-Best syllable candidates construction and pruning are analyzed experimentally.The system in this paper shows that the state-of-the-art performance is almost the same as the THEESP system.The experimental results and conclusions in the paper are expected to provide useful reference for fur- ther study of mandarin speech recognition.
出处
《计算机工程与应用》
CSCD
北大核心
2001年第13期1-4,101,共5页
Computer Engineering and Applications
基金
国家863计划基金
国家杰出青年科学基金(编号:69625103)资助
