Hidden Markov models (HMMs) are a type of probabilistic algorithm used in many applications in computer science, including speech recognition and bioinformatics. HMMs involve using a series of probability distributions over a set of possible outcome states of a process to determine the likelihood of each possible sequence of events or observations.
HMMs are typically used in machine learning and artificial intelligence tasks. For example, they are used to identify patterns in sequence data, such as strings of words in speech, musical notes in a score, or biological markers in DNA and proteins. An HMM is composed of a set of states, each of which correspond to a certain probability distribution. Transition probabilities between the states govern how one state is transitioned to another. Observations are taken and the probability of observing each value is, in turn, related to the current state of the model. The transitions and observations are used to derive the most likely path through the model.
When applied to speech recognition, HMMs are used for feature extraction (or acoustic modelling). The HMMs are learnt from examples of audio (written or spoken words) and used to detect sounds in live speech. This is done by training an HMM with known audio sequences (utterances) and then using it to recognise unknown sequences.
HMMs are also used in bioinformatics. Here, they are used to analyse the probability of observing specific DNA or protein sequences. This is often done in order to identify and compare gene sequences, or to detect regulatory elements such as promoters and enhancers. HMMs are also used to analyse medical data, such as heart rate.
HMMs are a versatile and effective tool for a wide range of computational tasks. By analysing sequences and learning the probability of transitioning from one state to another, HMMs can identify patterns in data and can provide useful insights for a variety of disciplines.
