Old Chinese versionIn the previous chapter, we have introduced DTW which is ideal for speaker-dependent ASR applications. This is suitable for the retrieval of utterances from the same speaker. A typical example of such applications is name dialing on mobile phones where you need to record your voices first. On the other hand, if we want to have speaker-independent ASR, we need to resort to HMM (Hidden Markov Models) which is a statistic model that requires a massive amount of training data for reliable recognition. Based on the used statistical models, HMM can be classified into two categories:

- Discrete HMM (DHMM for short): This type of HMM evaluates probabilities based on discrete data counting.
- Continuous HMM (CHMM for short): This type of HMM evaluates probabilities based on continuous probability density functions such as GMM.
We shall introduce these two types of HMMs for speech recognition in the following sections.

Data Clustering and Pattern Recognition (資料分群與樣式辨認)