The TIMIT corpus

A joint effort by MIT, SRI and Texas Instruments produced a corpus of 630 speakers across eight dialects, annotated at the phoneme level.

Why it matters

Speech recognition got a shared ruler five years before text processing got the Penn Treebank.

Each speaker read ten phonetically rich sentences: 6300 sentences in all, more than five hours of speech with time-aligned transcription. The recording was done at Texas Instruments, transcription at MIT, verification and publication by NIST. The corpus is still used as a teaching example, though its size is long since trivial. What it established matters more: results became comparable between laboratories.

Event record

Event date

October 1990

Timeline date

Event date

Verification

Sources gathered automatically · September 17, 2026

Lines

ID

evt-0195

The full CD-ROM release was October 1990; a prototype disc appeared in December 1988 and the documentation in February 1993.

Sources

primary NIST-DARPA Interagency Agreement: SLS Program (David S. Pallett, NIST), in Speech and Natural Language: Proceedings of a Workshop Held at Pacific Grove, California, February 19-22, 1991 (ACL Anthology H91-1092)
National Institute of Standards and Technology; DARPA Speech and Natural Language Workshop proceedings, Morgan Kaufmann (ACL Anthology copy) · Published 1991
primary DARPA TIMIT acoustic-phonetic continuous speech corpus CD-ROM, NIST speech disc 1-1.1
NIST · Published February 1993
primary TIMIT Acoustic-Phonetic Continuous Speech Corpus
Linguistic Data Consortium · Published 1993
primary The DARPA TIMIT Acoustic-Phonetic Continuous Speech Corpus CD-ROM (NISTIR 4930)
National Institute of Standards and Technology (Garofolo, Lamel, Fisher, Fiscus, Pallett, Dahlgren) · Published February 1993

Related events

Earlier

Related Resource Management: a shared speech recognition test 1989links to this record
The same NIST group that ran the tests is named in its 1989 report as distributing TIMIT on the first CD-ROM of speech data in the United States.
- National Institute of Standards and Technology (NIST) (formerly National Bureau of Standards), principal investigator David S. Pallett, in Speech and Natural Language: Proceedings of a Workshop Held at Philadelphia, Pennsylvania, February 21-23, 1989, p. 191 (H89-1028) · Association for Computational Linguistics (ACL Anthology); DARPA workshop proceedings
Related Sphinx 1988
It supplies the shared data that systems of that generation began to be measured on.
Related The UCI Machine Learning Repository 1987links to this record
The same idea in another field: shared data makes results from different laboratories comparable.

Later

Related The Switchboard corpus 1992links to this record
After read speech came spontaneous speech, a far harder case.
Related The Penn Treebank 1993
The same move as the Penn Treebank: a shared annotated set as the condition for comparison.
Builds on Deep networks in speech recognition 2009links to this record
The results were measured on the TIMIT corpus.
- Deep Belief Networks for phone recognition · NIPS 2009 workshop on deep learning for speech recognition