Hidden Markov Models in speech recognition

One historical method used for phoneme detection is a machine learning technique called Hidden Markov Models (HMMs). At their core, HMMs are statistical models that use hidden or unobserved states, unlike a regular Markov chain where the state is visible. In HMMs, the observation (e.g. the audio waveform of recorded speech) is visible, but the probability of state transitions is not. These hidden states are a common attribute of machine learning-based models. You can experiment with the visualization below to get a sense for how isolated word recognition works.

Someone has stolen my credit card.

The graphic below represents how the HMM is working under the hood. As the waveform is processed, the HMM constructs this probability lattice to find the best path.