许可协议: CC BY-NC-SA 4.0


The TIMIT corpus of read speech has been designed to provide the speech research community with a standardized corpus for the acquisition of acoustic-phonetic knowledge and for the development and evaluation of automatic speech recognition systems. The creation of any reasonably-sized speech corpus is very labor intensive. With this in mind, TIMIT was designed so as to balance utility and manageability, containing small amounts of speech from a relatively diverse speaker population and a range of phonetic environments. This section provides more detailed information on the contents of TIMIT and on the division of the TIMIT speech material into subsets for training and testing purposes.
TIMIT contains a total of 6300 utterances, 10 sentences spoken by each of 630 speakers from 8 major dialect divisions of the United States. The 10 sentences represent roughly 30 seconds of speech material per speaker. In total, the corpus contains approximately 5 hours of speech. All speakers are native speakers of American English and were judged by a professional speech pathologist to have no clinical speech pathologies.

Data Collection

The speakers were primarily TI personnel, many of whom were new to TI and the Dallas area. They were selected to be representative of different geographical dialect regions of the U.S.2 A speaker's dialect region was defined as the geographical area of the U.S. where he or she lived during their childhood years (age 2 to 10). The geographical areas correspond with recognized dialect regions of the U.S. (Language Files, Ohio State University Linguistics Dept., 1982), with the exception of the Western dialect region (dr7) in which dialect boundaries are not known with any confidence and "dialect region" 8 where the speakers moved around a lot during their childhood. The locale of each speaker's childhood is indicated by a color-coded marker on the map.
Recordings were made in a noise-isolated recording booth at TI, using a semi-automatic computer system (STEROIDS) to control the presentation of prompts to the speaker and the recording. Two-channel recordings were made using a Sennheiser HMD 414 headset-mounted microphone and a Breul & Kjaer 1/2" far-field pressure microphone (#4165).
The speech was directly digitized at a sample rate of 20 kHz using a Digital Sound Corporation DSC 200 with the anti-aliasing filter at 10 kHz. The speech was then digitally filtered, debiased, and downsampled to 16 kHz.

Subjects were seated in the recording booth and prompts were presented on a monitor. The subjects wore earphones through which a low-level (approximately 53 dB SPL) of background noise was played to eliminate the unusual voice quality produced by the "dead room" effect. TI attempted to keep both the recording gain and the level of noise in the subject's earphones constant during the collection. At the beginning of each recording day, a standard calibration tone was recorded from each microphone and the voltage at the subject's earphones was checked and adjusted as necessary.
The speakers were given minimal instructions and asked to read the prompts in a "natural" voice. The recordings were monitored, and any suspected mispronunciations were flagged for verification. Verification consisted of listening to the utterance by both the monitor and the speaker. When a pronunciation error was detected, the sentence was re-recorded. Variant pronunciations were not counted as mistakes.


Please use the following citation when referencing the dataset:

author = {Garofolo, J. and Lamel, Lori and Fisher, W. and Fiscus, Jonathan and Pallett, D.
and Dahlgren, N. and Zue, V.},
year = {1992},
month = {11},
pages = {},
title = {TIMIT Acoustic-phonetic Continuous Speech Corpus},
journal = {Linguistic Data Consortium}



Linguistic Data Consortium, University of Pennsylvania
The Linguistic Data Consortium (LDC) is an open consortium of universities, libraries, corporations and government research laboratories.