High performance speech and gesture recognition based on the stochastic model with mutual state-observation-dependencies

针对语音和手势识别中时间变化较为复杂的随机现象，提出了部分隐马尔可夫模型（PHMM）。它可以处理观察和状态转换中的观察依赖行为。一些模拟实验显示了PHMM的巨大潜力。此外，在手势识别和孤立口语词识别实验中，PHMM的性能都超过了HMM。在原始PHMM的表述中，我们使用公共的隐状态和可观测状态对来确定观测和状态转移的随机现象。在这里修改的公式中，我们使用共同的隐藏状态，但不同的可观察状态的观察和状态转移分开。本文还提出了平滑部分隐马尔可夫模型（Smoothed Partly Hidden Markov Model，SPHMM），该模型中的观测概率和状态转移概率是由基于PHMM和基于HMM的概率的几何平均值定义的。连续语音识别实验表明，当平滑权值设置适当时，SPHMM的识别性能优于HMM和PHMM。

Aiming at treating more complicated temporal changes of stochastic phenomena, Partly-Hidden Markov Model (PHMM), is proposed and applied to speech and gesture recognition. It can treat the observation dependent behaviors in both observations and state transitions. Some simulation experiments showed the high potential of PHMM. In addition, from the gesture recognition and the isolated spoken word recognition experiments, PHMM showed the performance to exceed HMM.In the formulation of original PHMM, we used common pair of hidden state and observable state to determine the stochastic phenomena of the observation and the state transition. In the formulation modified here, we use common hidden state but different observable state for the observation and for the state transition separately. This slight modification brought the big flexibility in the modeling of phenomena and reduced the word errors compared with HMM and traditional PHMM using continuous speech.We also proposed Smoothed Partly-Hidden Markov Model (SPHMM), in which the observation and state transition probabilities are defined by the geometric means of PHMM-based ones and HMM-based ones. From continuous speech recognition experiments, it was found that SPHMM gave the best performance compared with HMM and PHMM when the weight of smoothing was set adequately.