Hidden Markov Model

HMM Full Form in English, letter by letter

HMM full form is Hidden Markov Model, a statistical model for sequences whose underlying states cannot be observed, used in speech and bioinformatics.

HMM Full Form in Indian Languages

Before deep learning took over, speech recognisers and gene finders ran on an HMM: Hidden Markov Model, a probabilistic way to work out what cannot be seen directly. Six-language names follow.

What a Hidden Markov Model is

HMM full form is Hidden Markov Model. It models a system that moves between states you cannot see, where each state produces an observable signal with some probability.

Three sets of numbers define it: initial state probabilities, a transition matrix for how likely one state is to follow another, and an emission matrix for what each state is likely to produce.

The Markov assumption is the simplifying trick: the next state depends only on the current one, not on the whole history.

Where you have met one already

Classic speech recognition ran on HMMs for decades, and some text-to-speech systems still use statistical models from the same family.

Handwriting recognition and OCR split words into character states. In bioinformatics, profile HMMs in tools such as HMMER spot protein families in genomes.

Part-of-speech tagging (a basic NLP task), gene finding, gesture and activity recognition and some fraud-detection models all have HMM versions. In chat, of course, HMM is just 'hmm', the sound of thinking.

Three problems, three algorithms

Likelihood

How probable is this observed sequence? The forward algorithm answers that, and it is used to score which model best fits a spoken word.

Decoding

Which hidden states are most likely? Andrew Viterbi's 1967 algorithm finds the best path efficiently, and it is what turns acoustic frames into words.

Learning

How do you fit the parameters? Baum-Welch, an expectation-maximisation method, trains the matrices from example data without any labelled states.

Assumptions that make an HMM go wrong

A first-order Markov model ignores long-range context. A decision five steps back cannot influence the present without extra states being added.

States must also be designed by hand and emissions assumed independent, and real data often violates both. Conditional random fields fixed part of this by allowing overlapping features without forcing a generative model.

Deep learning, the branch of AI built on recurrent networks and then transformers, now dominates sequence labelling, largely because it learns its own features instead of relying on yours.

When to still choose an HMM

Small datasets: a few hundred labelled sequences can fit a small HMM, where a large neural model would overfit.

Small footprint and fast inference: useful in embedded audio, on-device keyboards and real-time decoding where a heavy model will not fit.

Interpretability: you can print the transition and emission matrices and see which state follows which, which matters when a result has to be explained in a report or viva.

A worked example with weather and moods

Suppose you cannot see the weather, only your friend's mood. An HMM might say a sunny day is followed by another sunny day with 80% probability, and that your friend is cheerful on 90% of sunny days but only 30% of rainy ones.

Given a week of moods, you can infer the hidden weather. The single most likely chain of weather states is exactly what the Viterbi algorithm returns.

Swap the states and it becomes speech: hidden states are phonemes and observations are audio frames. Swap them again and it becomes gene finding: hidden states are coding and non-coding stretches, and observations are the letters A, C, G and T.

Put numbers on it. Say a day is sunny with 60% chance at the start, a sunny day is followed by sunny with 80% and by rainy with 20%, and a rainy day is followed by rainy with 60%. If your friend is cheerful on 90% of sunny days and 30% of rainy ones, the moods cheerful then grumpy fit the chain Sunny then Rainy with probability 0.6 x 0.9 x 0.2 x 0.7 = 0.0756. The other chains score lower (Rainy then Rainy 0.0504, Sunny then Sunny 0.0432, Rainy then Sunny 0.0048), so Viterbi picks the first without listing them all.

Where the name comes from

The chain is named for Andrey Markov, who around 1913 analysed the sequence of vowels and consonants in Pushkin's Eugene Onegin to show that a chain of dependent events can still be studied mathematically.

The 'hidden' models were developed in the late 1960s and 1970s, when Leonard Baum and colleagues worked out the forward-backward methods. They became standard in speech work at IBM and elsewhere.

Viterbi's decoding algorithm, first devised for error-correcting codes, became the usual way to read a sequence out of an HMM.

HMM, Hidden Markov Model, models what you cannot see by modelling what it produces. Even if you never write one, the idea explains how a lot of earlier speech and language technology worked.

HMM Full Form: Common Questions