Written and maintained by CASRAI Editorial Board
Published
On this page
TIMIT is a speech corpus: a set of recordings of 630 American English speakers, each reading ten sentences, with time-aligned word and phone transcriptions for every recording. The Linguistic Data Consortium (LDC) distributes it as catalog item LDC93S1, and it was designed for acoustic-phonetic research and for developing and evaluating automatic speech recognition systems. This guide decodes how its 6,300 sentences and its train and test subsets are built.
The corpus is a long-standing reference dataset in speech processing, which is why it appears in electrical engineering and signal-processing coursework and in the methods sections of papers on speech and language technology. Reproducing a result on it depends on knowing exactly which speakers and sentences sit in which subset, and that is the part most summaries skip.
Short answer: what TIMIT contains
According to the LDC catalog record, TIMIT contains about five hours of English read speech from 630 speakers of eight major dialects of American English, each reading ten phonetically rich sentences. The transcriptions are time-aligned and hand verified, and the test and training subsets are balanced for phonetic and dialectal coverage. The recordings are one-channel PCM at 16,000 Hz, recorded with a microphone. 6,300 sentences in total (10 sentences x 630 speakers) make up the corpus.
Definition and scope
TIMIT is the TIMIT Acoustic-Phonetic Continuous Speech Corpus. The LDC record describes it as a joint effort among the Massachusetts Institute of Technology (MIT), SRI International and Texas Instruments (TI). The speech was recorded at TI, transcribed at MIT, and verified and prepared for CD-ROM production by the National Institute of Standards and Technology (NIST). The corpus readme states that its purpose is the acquisition of acoustic-phonetic knowledge and the development and evaluation of automatic speech recognition systems, and that the work was sponsored by DARPA-ISTO. The name is commonly read as a blend of Texas Instruments and MIT, the two sites that appear in its recording and transcription history.
The LDC record identifies the corpus with these catalog details: catalog number LDC93S1, ISBN 1-58563-019-5, ISLRN 664-033-662-630-6, DOI 10.35111/17gk-bn40 and membership year 1993. The record lists no updates. The authors named on the record are Garofolo, Lamel, Fisher, Fiscus, Pallett, Dahlgren and Zue. A persistent identifier such as a DOI matters for citation: the LDC record gives the form Garofolo et al., TIMIT Acoustic-Phonetic Continuous Speech Corpus LDC93S1, Philadelphia: Linguistic Data Consortium, 1993.
Scope matters as much as content. TIMIT is read speech, not conversation: every speaker reads prepared sentences. Speaker metadata on the LDC record covers gender, dialect, birth date, height, race and education level, and about 70% of the speakers are men and 30% women. Dialect region in the readme means the region where the speaker lived as a child, which is a coarse label rather than a measured accent.
How TIMIT is organised: sentence types and speakers
The three sentence types
The readme divides the text material into three kinds of sentence, and the totals reconcile exactly to 6,300 utterances. The table below reconstructs that arithmetic from the counts the readme gives.
| Type | Designed by | Distinct texts | Read per speaker | Speakers per text | Utterances |
|---|---|---|---|---|---|
| SA (dialect sentences) | SRI | 2 | 2 | 630 | 1,260 |
| SX (phonetically compact) | MIT | 450 | 5 | 7 | 3,150 |
| SI (phonetically diverse) | TI | 1,890 | 3 | 1 | 1,890 |
| Total | 2,342 | 10 | 6,300 |
The two SA sentences are read by all 630 speakers, which is what makes them useful for comparing dialects: the words are identical and only the speakers differ. The SX sentences were designed for phonetic coverage, and each of the 450 texts is spoken by seven speakers. The SI sentences were selected from the Brown Corpus and from a playwrights’ dialog collection; each is read by one speaker only. In the readme’s file naming, these appear as sa, sx and si files.
Speakers by dialect region
The speakers are divided into eight dialect regions, coded dr1 to dr8. Counts from the readme are in the data table below; they total 438 men and 192 women, or 630 speakers. The regions are New England (dr1), Northern (dr2), North Midland (dr3), South Midland (dr4), Southern (dr5), New York City (dr6), Western (dr7) and Army Brat (dr8), the last being speakers who moved around during childhood rather than a geographic region. The readme notes that the boundaries of the Western region are not known with any confidence, so dr7 should not be read as a precisely defined dialect area.
Train, core test and complete test sets
The readme defines a training set and two test sets, and the difference between the two is a frequent source of confusion when results from different papers are compared. No sentence text appears in both training and test material, so a recogniser is never scored on a sentence it saw in training.
- Core test set: 24 speakers, two men and one woman from each of the eight dialect regions, reading 192 sentences in total (five SX and three SI sentences each), with distinct prompts. The readme lists core speaker IDs such as DAB0, WBT0 and ELC0 for dr1.
- Complete test set: 168 speakers (112 men and 56 women) and 1,344 utterances, about 27% of the material, covering 624 distinct texts.
The arithmetic is worth checking once: 24 speakers x 8 sentences gives 192, and 168 speakers x 8 sentences gives 1,344. Both figures are consistent with the test material containing only the SX and SI sentences, not the SA sentences that every speaker reads. Because the core set is a 24-speaker subset of the complete test material, a score reported on the core set and a score reported on the complete set are different measurements and should be labelled as such.
File layout and transcription format
The readme gives the directory structure as /timit/ followed by train or test, then a dialect-region folder (dr1 to dr8), then a speaker folder, then the sentence files. Speaker folders begin with m or f for the speaker’s sex, followed by the speaker ID, which is three initials plus a digit. A worked path from the readme is /timit/train/dr1/fcjf0/sa1.wav, meaning a training-set recording by female speaker FCJF0 from dr1 reading the first SA sentence.
Each sentence has four files, distinguished by extension:
- .wav: the speech waveform, with a SPHERE header.
- .txt: the orthographic text of the sentence.
- .wrd: the time-aligned word transcription. Words were aligned to the phonetic segments by dynamic string alignment, using the lexicon file timitdic.txt.
- .phn: the time-aligned phonetic transcription, following the protocols of Seneff and Zue (1988), with the phone codes listed in the file phoncode.doc.
The alignment files use one line per unit in the form BEGIN_SAMPLE END_SAMPLE TEXT, and the first begin sample is 0. The readme’s example is the sentence ‘She had your dark suit in greasy wash water all year.’ from speaker FNLP0, whose word file begins with lines such as ‘7470 11362 she’. Times are sample numbers, so at the stated 16,000 Hz sampling rate a sample index converts to seconds by dividing by 16,000. The readme also warns that the last END_SAMPLE value may be less than the length of the waveform file, so alignment code should not assume the labels cover every sample.
Tools for working with such alignments include Praat, which supports time-aligned annotation tiers, and ELAN, which is built for annotating media; a conversion step from the sample-index format is needed first.
Related corpora derived from TIMIT
Several later corpora are listed on the LDC record as related works. They reuse TIMIT material or its design under different recording or channel conditions, so a result on one cannot simply be carried to another.
| Corpus | LDC catalog number |
|---|---|
| NTIMIT | LDC93S2 |
| TIMIT MS-WAV version | LDC93S1W |
| FFMTIMIT | LDC96S32 |
| CTIMIT | LDC96S30 |
| HTIMIT | LDC98S67 |
| STC-TIMIT 1.0 | LDC2008S03 |
| WTIMIT 1.0 | LDC2010S02 |
| Noisy TIMIT Speech | LDC2017S04 |
| Global TIMIT (for example, Mandarin Chinese variant) | LDC2021S03 |
Where it matters for labs and research groups
For a lab, the practical questions are licensing, citation and reproducibility. The LDC record states that the corpus is available under the LDC User Agreement for Non-Members, with fees shown behind the LDC login, so a group should confirm its licence position with LDC before copying data between machines or sharing it with collaborators. The catalog identifiers (LDC93S1 and the DOI above) are what a methods section should cite so that a reader obtains the same edition: the record lists no updates, which means there is no later version to confuse with the 1993 release.
Reproducibility depends on stating which subset was used. A methods section that says only ‘evaluated on TIMIT’ leaves open whether the training portion, the 24-speaker core test set or the 168-speaker complete test set was used, whether SA sentences were excluded, and which phone label set the .phn files were mapped to. Saying so in one sentence removes the ambiguity. Because the readme defines the core and complete test sets by speaker and sentence, they can be rebuilt from the directory structure rather than from a third-party list.
Dialect balance is also a design property to keep in mind when interpreting results. Region sizes range from 33 speakers (dr8) to 102 speakers (dr2 and dr3), so per-dialect error rates rest on different sample sizes. The core set takes two men and one woman from each region regardless of region size.
Common misunderstandings
- It is not free to download from the LDC. The record lists the LDC User Agreement for Non-Members, with fees behind login; that is a licence arrangement, not an open dataset.
- It is read speech, not conversational speech. Every utterance is a prepared sentence, so it does not represent spontaneous speech.
- The sampling rate is 16 kHz. The LDC record lists 1-channel PCM at 16,000 Hz.
- dr8 is not a place. The Army Brat label marks speakers who moved around during childhood.
- ‘The test set’ is ambiguous. The core set (24 speakers, 192 sentences) and the complete set (168 speakers, 1,344 utterances) are different subsets.
- The gender balance is not 50/50. The speaker counts are 438 men and 192 women.
Frequently asked questions
What does TIMIT stand for?
TIMIT is the name of the TIMIT Acoustic-Phonetic Continuous Speech Corpus, produced by MIT, SRI International and Texas Instruments. The name is commonly read as a blend of Texas Instruments and MIT.
How many speakers and sentences does TIMIT have?
It has 630 speakers, each reading ten sentences, for 6,300 sentences, according to the readme and the LDC catalog record. The speakers cover eight major dialects of American English.
Is TIMIT free?
The LDC record lists it under the LDC User Agreement for Non-Members, with fees behind the LDC login, so it is a licensed corpus rather than a freely downloadable one.
What is the TIMIT core test set?
It is the 24-speaker subset with two men and one woman per dialect region, reading 192 distinct sentences (five SX and three SI sentences each). The complete test set is larger, with 168 speakers and 1,344 utterances.
What sampling rate and format does TIMIT use?
The LDC record lists one-channel PCM at 16,000 Hz. The readme describes the waveform files as having a SPHERE header, with separate text, word-alignment and phone-alignment files for each sentence.
Sources
- Linguistic Data Consortium, TIMIT Acoustic-Phonetic Continuous Speech Corpus (LDC93S1) catalog record, https://catalog.ldc.upenn.edu/LDC93S1, checked 2026-11-14.
- LDC93S1 corpus readme, https://catalog.ldc.upenn.edu/docs/LDC93S1/readme.txt, checked 2026-11-14.
- UMBC CISAAD, The DARPA TIMIT Acoustic-Phonetic Continuous Speech Corpus, https://cisaad.umbc.edu/the-darpa-timit-acoustic-phonetic-continuous-speech-corpus, checked 2026-11-14 (secondary; agrees with the readme on sentence counts).
- Wikipedia, TIMIT, https://en.wikipedia.org/wiki/TIMIT, checked 2026-11-14 (secondary; used only for the name’s origin).
Data: TIMIT speakers by dialect region and sex
Counts of male and female speakers in each of the eight TIMIT dialect regions (dr1 to dr8), compiled from Table 1 of the LDC93S1 corpus readme, with region totals recomputed from the male and female counts and the overall total of 630 speakers.
Period covered: 1990 corpus release. Coverage: United States (American English dialect regions).
| Dialect region | Region code | Male speakers (speakers) | Female speakers (speakers) | Total speakers (speakers) |
|---|---|---|---|---|
| New England | dr1 | 31 | 18 | 49 |
| Northern | dr2 | 71 | 31 | 102 |
| North Midland | dr3 | 79 | 23 | 102 |
| South Midland | dr4 | 69 | 31 | 100 |
| Southern | dr5 | 62 | 36 | 98 |
| New York City | dr6 | 30 | 16 | 46 |
| Western | dr7 | 74 | 26 | 100 |
| Army Brat (moved around) | dr8 | 22 | 11 | 33 |
Sources
- LDC93S1 corpus readme (Table 1) (accessed 2026-11-14)
- LDC catalog record LDC93S1 (accessed 2026-11-14)
Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.







