Paper Review #11: A Tutorial on Hidden Markov Models and Selected Application in Speech Recognition

1. Paper Bibliography

• Title: iscrete Markov Processes A Tutorial on Hidden Markov Models and Selected Application in Speech Recognition

• Authors: Rabiner, Lawrence

• Publication: Proceedings of the IEEE 77.2 (1989): 257-286.

2. Summary

• Statistical methods of Markov source / Hidden Markov modeling
• rich in mathematical structure; the theoretical basis for use in a wide range of applications
• work well in practice (applied properly)
• Review paper
• Review paper

3. Details

• Discrete Markov Processes
• Extension to Hidden Markov Models
• Elements of an HMM
• The Three Basic Problems for HMMs
• Solutions to the Three Basic Problems of HMMs
• Solution to Problem 1
• Solution to Problem 2
• Solution to Problem 3
• Types of HMMs
• Continuous Observation Densities in HMMs
• Autoregressive HMMs
• Variants on HMM Structures - Null Transitions and Tied States
• Inclusion of Explicit State Duration Density in HMMs
• Optimization Criterion
• Comparison of HMMs
• Implementation Issues for HMMs
• Scaling
• Multiple Observation Sequences
• Initial Estimates of HMM Parameters
• Effects of Insufficient Training Data
• Choice of Model
• Implementation of Speech Recognizers Using HMMs
• Overall Recognition System
• Isolated Word Recognition
• LPC Feature Analysis
• Vector Quantization
• Choice of Model Parameters
• Segmental k-Means Segmentation into States
• Incorporation of State Duration into the HMM
• HMM Performance on Isolated Word Recognition
• Connected Word Recognition Using HMMs
• Connected Digit Recognition from Word HMMs Using Level Building
• Level Building on HMMs
• Training the Word Models
• Duration Modeling for Connected Digits
• Performance of the Connected Digit HMM Recognizer
• HMMs for Large Vocabulary Speech Recognition
• Limitations of HMMs

With HMM, an advanced process
Foward chaining + Backward chaining = Viterbi
1. Evaluation problem: calculating probability (F + B)
2. Decoding Problem: given a sequence observation what is most likely hidden data
3. Training Problem: Baum Welch