A hidden Markov model (HMM) is a statistical model of a sequence in which an observable signal is produced by hidden states that transition over time. For roughly three decades it was the backbone of speech recognition, modeling speech as a chain of hidden phone states, each emitting a slice of audio that the system observes without seeing the state directly.
A brief history of the hidden Markov model
The hidden Markov model reached speech research in the 1970s, with foundational work at IBM and CMU, and matured through the 1980s as compute and training data grew.
By the 1990s HMM systems trained on corpora like Switchboard defined the state of the art, and toolkits such as HTK and later Kaldi made the approach the default for a generation of engineers. Almost every commercial dictation and telephony recognizer of that era ran on HMMs.
How a hidden Markov model works in ASR
Each phoneme is represented by a few HMM states, often three, to capture its beginning, middle, and end. The model carries two kinds of probability: transition probabilities, the chance of moving from one state to the next, and emission probabilities, the chance a given state produced the observed audio frame.
Decoding searches for the most likely path through the states, typically with the Viterbi algorithm, and that path maps back to a sequence of phones and words. The same alignment machinery drives forced alignment, which lines a known transcript up with audio timestamps.
From HMM-GMM to hybrid HMM-DNN
Early systems computed emission probabilities with a Gaussian mixture model, the HMM-GMM pairing that dominated for years. The big upgrade came around 2012, when a deep neural network replaced the Gaussian mixture model for emission scoring.
This hybrid HMM-DNN kept the HMM's sequence structure but gained the accuracy of neural acoustic modeling, and it ran production speech recognition until end-to-end models arrived.
Why hidden Markov models still matter
HMMs are largely historical now, superseded by end-to-end networks that learn alignment on their own. But the ideas endure and explain terms you still meet, because the notion of hidden states aligning to audio underpins forced alignment and lattice generation.
Some embedded and low-resource systems still run HMM hybrids because they are compact, predictable, and cheap to decode. Knowing the HMM story makes the leap to modern acoustic models far easier to follow.
Frequently asked questions
What is a hidden Markov model? It is a model of a sequence where hidden states you cannot observe generate the signal you can. In speech it treats phones as hidden states that each emit audio frames.
How does a hidden Markov model work in speech recognition? It represents each phoneme as a few states with transition and emission probabilities, then uses the Viterbi algorithm to find the most likely state path through the audio and map it back to words.
When are HMMs still used today? Mostly in embedded, offline, or low-resource systems that need small, predictable models, and in alignment tasks. Mainstream cloud recognition has moved to end-to-end neural networks.