Hidden Markov Models
In this chapter theory of Hidden Markov Models is described. Specially the gradient based ML and MMI training is treated mathematically in detail. Finally it is shown how the idea of HMM can be used in isolated and continuous recognition.
Introduction
As mentioned earlier, ASR problem can be attacked from two sides; namely
The Hidden Markov MOdel (HMM) is a result of the attempt to model the speech generation statistically, and thus belongs to the first category above. During the pas several years it has become the most successful speeh model used in ASR. The main reason for this success is it's wonderful ability to characterize te speech signal in a mathemathically tractable way.
In a typical HMM based ASR system, the HMM stage ids proceeded by the preprocessing (parameter extraction) stages. Thus the input to the HMM is a discrete time sequence of parameter vectors, such as those described in the previous chapter. The parameter vectors can be supplied to the HMM, either in vektor quantized form or in raw continuous form. It can be desiganed HMM is to handle any of the cases, but importand point is how
In this chapter theory of Hidden Markov Models is described. Specially the gradient based ML and MMI training is treated mathematically in detail. Finally it is shown how the idea of HMM can be used in isolated and continuous recognition.
- Introduction
- Definition of Hidden Markov Model
- Assumptions in the theory of HMMs
- There basic problems of HMMs
- The Evaluation Problems and the Forward Algorithm
- The Decoding Problem nd the Viterbi Algoritma
- The Learning Problem
- Maximum Lielihood (ML) eriterion
- Baum-Weleh Algorithm
- Gradien based method
- Maximum Mutual Information (MMI) criterion
- Gradien wrt transition probabilities
- Gradient wrt observation probabilities
- Use of HMMs in isolated recognition
- Training
- Recognition
- Statistical Language models
- training of a HMM based continuous recognizer
- ML training
- MMI training
- recognition using a HMM continuous recognizer
- Viterbi based recognition
- Level Building
- N-best search
- Calculatin of the recognizer performance
Introduction
As mentioned earlier, ASR problem can be attacked from two sides; namely
- From the side of speech generation
- From the side of speech perception
The Hidden Markov MOdel (HMM) is a result of the attempt to model the speech generation statistically, and thus belongs to the first category above. During the pas several years it has become the most successful speeh model used in ASR. The main reason for this success is it's wonderful ability to characterize te speech signal in a mathemathically tractable way.
In a typical HMM based ASR system, the HMM stage ids proceeded by the preprocessing (parameter extraction) stages. Thus the input to the HMM is a discrete time sequence of parameter vectors, such as those described in the previous chapter. The parameter vectors can be supplied to the HMM, either in vektor quantized form or in raw continuous form. It can be desiganed HMM is to handle any of the cases, but importand point is how
.jpg)
.jpg)
