The DARPA TIMIT Acoustic-Phonetic Continuous Speech Corpus

Authors: 

John S. Garofolo, Lori F. Lamel, William M. Fisher, Jonathan G. Fiscus, David S. Pallett, Nancy L. Dahlgren

Description of the Dataset:

The TIMIT corpus is designed to provide speech data for the acquisition of acoustic-phonetic knowledge and the development and evaluation of automatic speech recognition systems.

It includes read speech from 630 speakers from 8 major dialect regions of the United States.

The corpus consists of 2 dialect sentences, 450 phonetically compact sentences, and 1890 phonetically diverse sentences.

Data Creation Method:

• The TIMIT corpus was created through a joint effort involving the Massachusetts Institute of Technology (MIT), Stanford Research Institute (SRI), and Texas Instruments (TI).
• Text corpus design was a collaboration among MIT, SRI, and TI.
• Speech was recorded at TI.
• Transcription was done at MIT.
• Data was maintained, verified, and prepared for CD-ROM production by the National Institute of Standards and Technology (NIST).

Number of Speakers:

• The dataset contains speech from 630 speakers.

Total Size of Data:

• The corpus consists of 6300 sentences.

Real Samples:

• All samples in the dataset are real speech recordings from 630 speakers.

Fake Samples:

• There are no fake samples in the TIMIT corpus. All data is real.

Extra Details about the Data:

• The data includes orthographic transcriptions, word-level transcriptions, and phonetic transcriptions.

Data Types:

• .wav: Speech waveform files.
• .txt: Orthographic transcriptions.
• .wrd: Time-aligned word transcriptions.
• .phn: Time-aligned phonetic transcriptions.

Average Length of Data:

• Each speaker reads 10 sentences, leading to a total of 6300 sentences. The exact average length in terms of duration isn’t specified, but the dataset ensures a broad coverage of phonetic contexts.

Data Collection Period:

•  Not specified.

Annotation Process:

• Speech was transcribed at MIT, and phonetic transcriptions were created by trained linguists.

Usage Scenarios:

• The TIMIT corpus is widely used for:
  • Research in acoustic-phonetic studies.
  • Dialect and linguistic analysis

Technical Specifications:

• Sampling Rate: The waveform files (.wav) are typically sampled at 16 kHz.
• Bit Depth: Usually 16-bit audio files.

Data Accessibility:

• The dataset was originally distributed on CD-ROM and might now be available through specific research institutions or online repositories.

Challenges and Limitations:

• Limited dialect representation for certain regions, particularly the Western dialect where boundaries are less clearly defined.