语音识别解码.ppt

6.5 多遍融合解码系统 ROVER convert outputs into Word Transition Networks (WTNs) align and combine (WTNs) in a pre-specified order using voting to decide between aligned WTNs A simple example output: BUT IT DIDN’T ELABORATE 6.5 多遍融合解码系统 CNC In contrast to ROVER, align and combine CN(WTNs) word posteriors rather than voting-style approaches using voting to decide between aligned WTNs combined “posterior” found by CNC generally works slightly better than ROVER combining 2 systems ROVER very dependent on confidence scores system word posteriors, rather than 1-best helps but alignment more complex - not normally used with different segmentations 6.5 多遍融合解码系统 示例 练习： 动态规划对齐算法实现： 输入：两段文字（汉字）大于100个字 输出：对齐的结果，以及插入，删除，替代错情况 例子： Thank you for taking the class! 黎塔 lita@hccl.ioa.ac.cn 语言声学与内容理解重点实验室 6.3.1 解码算法 Token Passing Strategy Decoding is to find those paths through the search space which have the max log Prob. A token represents a partial path through the search space. At each time ,tokens are propagated along transitions in the network. At the same time, its log probability is accumulated by the corresponding transition probs and emission probs. As each token passing through the network, it must maintain a history recording its route. When there are multiple exits for a node, the token is copied so that all possible paths are explored in parallel. 6.3.1 解码算法 Token数据结构 6.3.1 解码算法 Search Alogorithm 6.3.2 词图 Nbest and Word Graph 和one best 结果相比 更高识别准确性 更丰富的信息 便于二次搜索 6.3.2 词图 Nbest many of the different word sequences are just one-word variations of each other the number of nbest hypotheses might grow exponentially with the length of the utterance. 6.3.2 词图 词图（Word Lattice） 由解码结果构成，一种压缩的多候选结构 每条边包括，时间，概率得分，word id 可以作为后端语义理解的输入 6.3.2 词图 Example 6.3.2 词图 典型的数据结构 Node=A B word score ts te other A: begin node B: end node Word: word id ts: start frame te: end frame Other: additional information 6.3.2 词图 词图密度 平均每个word对应的边的个数 词图