用于语音识别拒识的隐马尔可夫模型状态及状态驻留相关的声学置信量度

HMM STATE AND STATE DURATION DEPENDENT CONFIDENCE MEASURES FOR UTTERANCE REJECTION

随着语音识别系统继续从实验室转向实际应用，语音拒识就变得愈来愈重要．为解决语音识别系统对识别候选的接受／拒识判决问题，文中提出了基于隐马尔可夫模型（ＨＭＭ）的语音识别系统中状态和状态驻留相关的声学置信量度准则．给定状态下特征矢量的平均观测先验概率和给定特征矢量状态的后验概率均比较容易设定统一的拒识门限，且不需专门的训练．而状态驻留分布相关法则是基于驻留分布概率和置信区间理论，不仅可设定一个拒识门限，同时可给出语音识别候选的状态驻留可信度．实验表明上述拒识准则能很好地拒识误识别候选和词表外语音（ＯＯＶ或非关键词）。

Utterance rejection is becoming increasingly important as speech recognition systems continuously migrate from the laboratory to actual applications. Proposed in this paper are state and state duration dependent acoustic confidence measures for acceptance/rejection of recognition hypothesis in speech recognition systems based on hidden Markov model (HMM). The state dependent confidence measure is computed for each frame of speech as the feature vectors output probability or posteriori state probability given the observation features. It is easy to be implemented by using one single global threshold and no extra training is needed. The state duration dependent one is based on the duration distribution probability and confidence interval theory. Although it is required that the state duration distribution be trained, the data can be easily obtained during the traditional HMM training. Experiment results show that the methods can reject incorrect candidates and OOV (out of vocabulary) words effectively, thus significantly increasing the recognition accuracy with low rejection rate.