Hidden Markov Model

Created: 2022-04-11T23:46:09-05:00

Return to the Index

An adaption of Markov state machines where the machine and transitions are managed by example and not directly known. Using some algorithms the machine can be trained to immitate a set of state transitions to produce given example pairs, or consulted given an output to see what the most likely transitions to produce it were.

Markov models

Dragon Speak

These were also the secret to Dragon Naturally Speaking according to "How to Create a Mind."

• Sound is first run through band passes to find energy at a given set of frequencies (immitating the cochlea)
• These sound signatures are kept in a vector quantizing codebook to replace sound fragments with ex. 4,000 exemplars.
• Hidden Markov Models are used so that sound can be paired with a list of exemplars and the desired transcription information.
• Sound coming in is first filtered, matched to the nearest codebook, then the HMM is used to map the codebook sequence back to phonemes etc.