HMM算法在语音识别中的应用——算法学习

总体框架

输入

Treat acoustic(听觉的) input O as sequence of individual observations

$$O=o_1,o_2,...,o_t$$

结果

Define sentence as a sequence of words

$$W=w_1,w_2,...,w_n$$

判断模型

• 最大概率：$W=\mathop{arg \ max}_{W \in L}{\ P(W|O)}$
• 贝叶斯：$W=\mathop{arg \ max}_{W \in L}{\ \frac {P(O|W)P(W)} {P(O)}}$
• 化简：由于$P(O)$对所有$W$一样，$W=\mathop{arg \ max}_{W \in L}{\ {P(O|W)P(W)} }$

模型

• Feature Extraction: 39 “MFCC” features
• Acoustic Model: Gaussians for computing p(o|q)

• Lexicon(词典)/Pronunciation(发音) Model: HMM, what phones can follow each other
  

  ---

  

• Language Model: N-grams for computing $p(w_i|w_{i-1})$

markov chian

• states: $Q = q_1,q_2,...,q_N$,$q_t$ is the state at time t.
• transition probability: $A=[a_{11},a_{12},...,a_{NN}]$
  
  • $a_{ij}$ is the probability of trasition from i to j.
  • $a_{ij}=P(q_{t-1}=i|q_t=j)$
  • $\sum_{j=1}^N {a_{ij}} = 1$, for $i \in [1,N]$
• markov assumption:
  
  • $P(q_i|q_1q_2...q_{i-1})=P(q_i|q_{i-1})$
• initial status
  
  • $\pi _i =P(q_1=i)$
  • $\sum_{j=1}^N {\pi _j}=1$

hidden markov chian

three problem

thrid problem

以上所有内容从以下附件中抽取

• lec3.pdf
• 斯坦福大学关于语音识别的全部课件