Are emission probabilities learned or manually specified?

Either is possible. A teaching model may specify them directly, while a trained HMM estimates them from data under its assumptions.

Does the largest emission identify the most likely state?

Not necessarily. Priors, transitions, and other observations can change the posterior or best sequence.

Must emissions be words?

No. Observations can be other discrete symbols or continuous feature vectors, depending on the model.

AI Glossary / Natural language and speech / Intermediate

Emission probabilities

Definition

Emission probabilities describe how likely particular observations are given a hidden state in a probabilistic model such as a hidden Markov model, or HMM. Written P(observation | state), they connect the model's unseen states to visible data. They are different from transition probabilities, which describe movement between states.

What it means

An HMM represents a sequence using hidden states and observed outputs. In a simplified part-of-speech model, the states might be grammatical tags and the observations might be words. An emission probability then describes how likely a word is under a particular tag.
The direction of conditioning matters. P(word | tag) is not the same as P(tag | word). To infer a tag from a word sequence, the system also needs information about possible states and their transitions.
For discrete observations, the emission probabilities across all possible outputs of a given state sum to one. Continuous-observation models instead use probability densities, whose interpretation differs from the probability of an exact value.

How it works

Specify the possible hidden states and the observation representation. Each state receives an emission distribution over the observations.
If aligned states and observations are available during training, discrete emission estimates can be based on counts, with suitable estimation choices for sparse data. If states are hidden, HMM training methods can estimate parameters from observation sequences under the model's assumptions.
During inference, emission likelihoods are combined with initial-state and transition probabilities. A single locally attractive observation-state pairing may not belong to the highest-probability complete state sequence. The basic HMM also assumes that an observation depends on its current state rather than directly on all other observations.

Three practical examples

A teaching HMM assigns different word probabilities to noun and verb states in a small part-of-speech example.
A speech-model explanation distinguishes an acoustic observation model from the probabilities of moving between modeled speech states.
A sequence-analysis exercise checks whether each discrete emission distribution is normalized and uses the same observation vocabulary.

Think of it this way

Think of two hidden boxes containing differently mixed colored tokens. Observing a blue token gives evidence about which box produced it, but the box's token proportions are not themselves the probability that you chose that box. The analogy captures emission likelihoods and conditioning; a real HMM also models how hidden states change over time.

Worked example

Consider an invented one-step model with hidden states A and B and observations blue or red. State A emits blue with probability 0.8 and red with probability 0.2. State B emits blue with probability 0.3 and red with probability 0.7. Each state's probabilities sum to one.
Suppose the prior state probabilities are P(A)=0.2 and P(B)=0.8. After observing blue, the unnormalized weights are 0.2 × 0.8 = 0.16 for A and 0.8 × 0.3 = 0.24 for B.
The total probability of blue is 0.40, giving posterior probabilities P(A | blue)=0.16/0.40=0.4 and P(B | blue)=0.24/0.40=0.6.
Although A has the larger blue emission probability, B remains more probable after the observation because its prior probability is much larger. In a full HMM sequence, the surrounding observations and transition structure add further information. This small calculation demonstrates why emission likelihood and posterior state probability must not be confused.

Illustrative case study

Fictional case: TagTrail builds a classroom HMM for labeling a tiny vocabulary. A display initially labels P(word | tag) as “confidence in the tag.” Reviewers notice that it ignores the initial-state and transition probabilities used by the actual decoder.
The team changes the display to name the emission probability explicitly and adds a separate explanation of sequence-level inference. It uses a toy example where the strongest local emission does not produce the preferred state after other model information is considered.
This fictional case illustrates a reporting correction, not a measured tagging improvement. The interface now describes the quantity it actually shows instead of presenting a conditional likelihood as a complete decision confidence.

Common misconceptions and mistakes

Do not reverse the conditioning: P(observation | state) generally differs from P(state | observation).
Do not confuse emissions with transitions. One connects states to observations; the other connects states across time.
Avoid assuming that all emission models use discrete probability tables. Continuous densities have different units and may exceed one at particular points without violating normalization over the observation space. Also remember that modeled emissions need not imply a literal physical production mechanism.

Business use

The concept helps teams interpret sequence models and their diagnostics. It explains how observed data supports hidden-state hypotheses while keeping local observation evidence separate from the complete inference.
A practical review should check the observation representation, vocabulary or density family, normalization, and training assumptions. When a dashboard displays a score, confirm whether it is an emission likelihood, posterior state probability, or sequence score. These quantities answer different questions and should not share an ambiguous “confidence” label.

Check your understanding

State A emits a symbol more frequently than state B. Does observing that symbol guarantee that A is the more probable hidden state?

Answer: No. The posterior also depends on prior state information and, in a sequence model, transition and surrounding-observation evidence. The worked example shows B remaining more probable.

Related terms

Learn next

Build a two-state, two-observation HMM and verify each probability distribution. Then compare a local emission lookup with a posterior calculation and a complete sequence-decoding result, noting which information each calculation uses.

Probability distribution

Sources

Sources: Jurafsky and Martin, Speech and Language Processing, Hidden Markov Models, model definition and training sections: https://web.stanford.edu/~jurafsky/slp3/A.pdf Independently checked the HMM appendix’s emission definition, state-conditioning direction and Baum-Welch training discussion.

Previous: Emergent Bias Next: Emphatic temporal-difference learning