语音识别的自适应束剪枝方法

在语音识别的应用中，如何提高识别的效率性是一个重要的方向。尤其在大词汇表的识别中，庞大的搜索空间带来相应的计算代价，而传统剪枝方法在减少计算量的同时牺牲了识别率。为此引入自适应控制理论，自动调整束宽限定搜索空间在预定的规模。在此基础上，又提出了利用基线系统的平均激活模型音子模型实例作为自适应系统动态参考值的方法，实现启发式的束宽调节。应用此方法的解码器在不损失识别率情况下，计算时间和搜索空间比采用传统剪枝算法下降了55％和71％，显著地提高了解码器的效率。

Adaptive Beam Pruning for Speech Recognition

In large vocabulary continuous speech recognition, huge spaces are searched during the recognition process, resulting in vast computational cost. While most pruning search strategies can reduce the computation, the recognition rate often decreases. Based on adaptive control theory,a novel pruning method that can automatically steer beam width to make search space attain a predefined size is presented. Average active phone-model-instance as the dynamic reference signal of the adaptive system is farther used. Compared with the base system which is integrated with fixed beam pruning and MAPMI pruning, the proposed method leads to a significant reduction in computing time and a slightly improvement in word accuracy. By measuring the RTF, this system is proved to have good real-time performances.