Milestone · February 1989
The tutorial account of hidden Markov models
Rabiner wrote an exhaustive guide: the three HMM problems, the algorithm for each, and the engineering detail of implementation.
Why it matters
A method scattered across dozens of papers became accessible to any engineer, and that is what carried it out of the laboratories.
The paper works through evaluating likelihood, finding the best state sequence and training the parameters, along with the practical traps: scaling, multiple observations, initialisation. It is among the most cited texts in the history of signal processing. Its role is telling: a field is moved not only by a new idea but by whoever makes an existing one usable.
Event record
- Event date
- February 1989
- Timeline date
- Event date
- Verification
- Sources gathered automatically · September 17, 2026
- Lines
- ID
- evt-0194
The February 1989 issue of the Proceedings of the IEEE.
Records that link to this one
- Related Deep networks in speech recognition
The Gaussian mixtures standard since Rabiner's tutorial are replaced by a network.
- Builds on Conditional random fields
CRFs are framed as an answer to a limit of hidden Markov models, which cannot represent many overlapping features of the observation; on the same Penn treebank data the paper compares CRFs with HMMs.
Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data - Builds on Sign language sentences from video
The abstract carries hidden Markov models over to gesture from speech recognition, where they had been used "prominently and successfully".
Real-time American Sign Language recognition from video using hidden Markov models (T. Starner, A. Pentland), Proceedings of the International Symposium on Computer Vision, 265-270 - Related Unit selection: a voice from pieces of recording
The paper itself reduces the unit database to an ergodic HMM and cites Rabiner's tutorial; the difference is that its costs are not probabilities.
Unit selection in a concatenative speech synthesis system using a large speech database (Andrew J. Hunt, Alan W. Black), Proc. ICASSP-96, vol. 1, pp. 373-376, Atlanta, 7-10 May 1996 - Related A voice from a statistical model: HTS
Here the hidden Markov model no longer recognises speech but produces it: spectrum, pitch and duration come out of the same model.