对经典隐马尔可夫模型的经验性改进

文章分析了经典隐马尔可夫模型（Hidden Markov Model，HMM）齐次假设的理论缺陷，以及两种非齐次HMM。语音识别对比实验表明，经验性的惩罚概率法是稳健的、且更有效的补偿方法。实验结果还指出在最优惩罚概率下，经典HMM达到了与非齐次的基于段长分布的HMM（Duration Distribution Based HMM，DDBHMM）几乎相同的识别率，证明了齐次假设并不影响经典HMM在实用中的重要性。文章提出了一种改进Baum－Welch重估算法的初值的经验方法，用于HMM参数的估计，在汉语连续语音识别实验中一致性地降低了音节误识率。

Empirical Improvements of Classical Hidden Markov Model

The theoretical weakness of classical Hidden Markov Model（HMM），i．e．its homogeneous assumption and two inhomogeneous HMMs are studied in this paper．Control experiment for speech recognition shows that the empirical penalty approach is robust and more effective in compensating this unreasonable assumption．As a byproduct，the significance of classical HMM in practice is verified to be not affected by the homogeneous assumption，since classical HMM attains almost the same accuracy as the inhomogeneous Duration Distribution Based HMM（DDBHMM）under optimal probability penalties．This paper proposes an empirical method to improve parameter initialization for Baum－Welch re－estimation algorithm，which is applied to parameter estimation of HMM．In mandarin continuous speech recognition experiment the method results in consistent reduction of syllable error rate．