Hidden Markov Models intro

A Hidden Markov Model (HMM) assumes a hidden state that evolves over time, while you observe noisy signals. Classic tool for speech and sequence labeling before deep end-to-end models dominated many tasks.

HMMs teach state, observation, and decoding—ideas that still appear in tracking and some production pipelines.

You don’t see the “true weather state,” only someone with an umbrella or not. You infer likely hidden states over time using transition and emission probabilities.

Guessing if a friend is busy or free from sparse message timing patterns—imperfect observations.

Invent 2 hidden states and 2 observations; write what would make one observation more likely in each state.

⚠️ Myth: Deep learning deleted HMMs from history.

✓ Reality: They’re still taught and used in niches; conceptually foundational.

⚠️ Myth: Markov means the past never mattered in reality.

✓ Reality: It means your *model* assumes the present state summarizes what you need.

• Course 06 uncertainty; Course 13 ASR history
• Russell & Norvig AIMA (HMMs); Jurafsky & Martin: https://web.stanford.edu/~jurafsky/slp3/ ↗