Study of Effective Speech Recognition based on the Bi-directional Search Algorithm

基于双向搜索算法的有效语音识别研究

• 批准号: 07680401
• 负责人: KITA Kenji
• 金额: $ 1.28万
• 依托单位: The University of Tokushima
• 依托单位国家: 日本
• 项目类别: Grant-in-Aid for Scientific Research (C)
• 财政年份: 1995
• 资助国家: 日本
• 起止时间: 1995 至 1997
• 项目状态: 已结题
• 来源: https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-07680401/
• 关键词: speech recognition; bi-directional search; hidden Markov model; one-pass algorithm; acoustic model; language model; finite-state automaton; context-free grammar

The most widely used search strategy in continuous speech recognition is the left-to-right time-synchronous search. However, the left-to-right strategy often produces erroneous results because it does not utilize a full path score. In this research project, we studied a bi-directional search algorithm consisting of a forward time-synchronous search and a backward time-asynchronous search. In the first pass, the HMM-based one-pass algorithm, guided by a finite-state automaton (FSA), is used time-asynchronously for preparing a partial path map. Here, the partial path map keeps recognition likelihood of all partial paths that lead to any grammar state at every time instant. After the first pass, a new, forward time-synchronous search, guided by a finite-state or context-free grammar (CFG), is performed for finding the best recognition hypothesis. The second pass uses the partial path map to compute the full path likelihood.We implemented an experimental speech recognition system based on the above-mentioned bi-directional search algorithm. In the system, discrete HMMs were used as acoustic models. We also conducted experiments to compare the accuracy of several kinds of search strategies using the ATR speech database. FSA-based one-directional search attained an accuracy of 77.1%-87.3%, while FSA-backward CGF-forward search attained an accuracy of 88.6%. These results shows the effectiveness of the bi-directional search algorithm.

连续语音识别中使用最广泛的搜索策略是从左到右的时间同步搜索。但是，从左到右的策略经常会产生错误的结果，因为它没有使用完整的路径分数。在本研究项目中，我们研究了一种由前向时间同步搜索和后向时间异步搜索组成的双向搜索算法。在第一遍中，在有限状态自动机(FSA)的指导下，时间异步地使用基于HMM的一遍算法来准备部分路径映射。这里，部分路径映射保持在每个时刻导致任何语法状态的所有部分路径的识别可能性。在第一遍之后，在有限状态或上下文无关语法(CFG)的指导下执行新的前向时间同步搜索以找到最佳识别假设。第二遍利用部分路径映射计算全路径似然，实现了一个基于上述双向搜索算法的实验语音识别系统。该系统采用离散隐马尔可夫模型作为声学模型。我们还利用ATR语音库进行了实验，比较了几种搜索策略的准确率。基于FSA的单向搜索的准确率为77.1%~87.3%，而FSA反向CGF正向搜索的准确率为88.6%。这些结果表明了双向搜索算法的有效性。

项目成果

小田 裕樹, 北 研二: "単語の位置情報に基づくコーパスからのコロケーションの自動抽出" 自然言語処理. 5・1. 79-99 (1998)

Hiroki Oda、Kenji Kita：“根据单词位置信息从语料库中自动提取搭配”自然语言处理 5・1（1998）。

Kenji Kita: "Dialogue Knowledge Acquisition from Annotated Corpora" Proceedings of IEEE SMC96. 556-561 (1996)

Kenji Kita：“从注释语料库中获取对话知识”IEEE SMC96 论文集。

Tatsuya Iwasa: "Error Correction of Speech Recognition Outputs Using Generalized LR Parsing and Confusion Matrix" Proceedings of ROCLING VIII. 101-110 (1995)

Tatsuya Iwasa：“使用广义 LR 解析和混淆矩阵对语音识别输出进行纠错”ROCLING VIII 论文集。

Kenji Kita et al.: "One-Pass Scarch Algorithm for Continuous Speech Recognition Using Generalized LR Parsing:A CFG-Driven.Frame-Synchronous HMM-Based Approach" Transactions of IPSJ. 36・5. 1252-1259 (1995)

Kenji Kita 等人：“使用广义 LR 解析进行连续语音识别的单通道 Scarch 算法：基于 CFG 的帧同步 HMM 方法” IPSJ 36・5（1995 年）。

kenji Kita et al.: "Automatic Acquisition of Probabilistic Dialogue Models" Proceedings of ICSLP96. 196-199 (1996)

kenji Kita 等人：“概率对话模型的自动获取”ICSLP96 论文集。