Statistical speech recognition

Jelinek set out the full statistical scheme: an acoustic model, a language model, and a search for the most likely sentence given the signal.

Why it matters

Recognition stopped being a problem of phonetics and became one of estimating probabilities, the frame still in use.

The scheme separates the probability of the acoustics given the words from the prior probability of the words themselves, the language model. The IBM group hired no phoneticians and wrote no pronunciation rules: everything was estimated from data. Twelve years later the same group carried the apparatus over to machine translation. Modern systems changed the components but kept the search for the most likely sequence.

Event record

Event date

April 1976

Timeline date

Event date

Verification

Sources gathered automatically · September 17, 2026

Lines

ID

evt-0187

The April 1976 issue of the Proceedings of the IEEE.

Sources

primary Continuous speech recognition by statistical methods
Proceedings of the IEEE · Published April 1976
primary A Statistical Approach to Machine Translation (Brown, Cocke, Della Pietra, Della Pietra, Jelinek, Lafferty, Mercer, Roossin)
Computational Linguistics 16(2), pp. 79-85 (ACL Anthology) · Published June 1990

Related events

Earlier

Builds on The Baum-Welch algorithm 1970
The method rests on estimating hidden Markov models.
- Continuous speech recognition by statistical methods · Proceedings of the IEEE
Builds on A Mathematical Theory of Communication 1948
Decoding is posed as a noisy channel problem from communication theory.
- Continuous speech recognition by statistical methods · Proceedings of the IEEE

Later

Builds on Sphinx 1988links to this record
The system implements the statistical scheme on hidden Markov models.
- Recent Progress in the SPHINX Speech Recognition System · Speech and Natural Language Workshop
Related Statistical part-of-speech tagging 1988links to this record
The same statistical school: Church's paper calls the IBM speech group the strongest advocate of n-gram methods.
Enables Statistical machine translation 1988
Jelinek's group applied the same noisy channel to translation twelve years later.
- Continuous speech recognition by statistical methods · Proceedings of the IEEE
Related Kneser-Ney smoothing 1995links to this record
The smoothing concerns the language model, the part of the statistical speech recognition scheme that estimates the probability of a word sequence.
Related A neural probabilistic language model 2003links to this record
Instead of counting n-grams, a network with distributed representations estimates word probability.