Datasets

This page provides a number of datasets grouped by application, as well as research results (papers, numerical results, output transcriptions, intermediary data, etc) corresponding to each dataset.

Automatic speech recognition

1st CHiME Challenge (2011)

Artificially distorted version of the small vocabulary GRID audio-visual corpus (audio only). Binaural reverberated speech with speaker situated in front of the microphones. Additive household noises impinging from different directions. Clean-training, noisy-training, development and evaluation sets available, see

Jon Barker, E. Vincent, N. Ma, H. Christensen, P. Green, "The PASCAL CHiME speech separation and recognition challenge", Computer Speech & Language, Volume 27, Issue 3, May 2013, Pages 621-633.

Available from Computer Speech and Language here

Corpus available here (no cost)

Resources

• Training recipe of the challenge for HTK here.

Baselines

• See the paper above for results for a wide range of techniques.

AURORA 5 (2007)

Artificially distorted version of the digits TI-DIGITS corpus. Additive noise and additive noise plus reverberant speech sets. Variable SNR range. Various mixed training sets, no evaluation set, see

G. Hirsch "Aurora-5 Experimental Framework for the Performance Evaluation of Speech Recognition in Case of a Hands-free Speech Input in Noisy Environments", Niederrhein University of Applied Sciences, 2007.

Paper available online here (no cost)

Corpus available from LDC here

Resources

• Training recipe for HTK is provided with the corpora.

Baselines

• Reproducible baseline: The above cited paper includes a baseline for the ETSI Advanced Front-End.

AURORA 4 (2002)

Artificially distorted version of the 5K word Wall Street Journal corpus (WSJ0). Stationary and non-stationary noises added. Second recordings with distant mismatched microphone. Clean-training, mixed-training, noisy training and test sets available. No evaluation set, see

G. Hirsch "Experimental Framework for the Performance Evaluation of Speech Recognition Front-ends on a Large Vocabulary Task", ETSI STQ Aurora DSR Working Group, 2002.

Paper available with the corpus.

Corpora available from ELRA here and here

Resources

• Training recipe for HTK available here. Note that this recipe is for Wall-Street Journal (WSJ0), which is the clean speech version of AURORA4. Small changes are needed in the feature extraction scripts to account for different file terminations.

Baselines

Contribute a dataset

To contribute a new dataset, please

• create an account and login
• go to the wiki page above corresponding to your application; if it does not exist yet, you may create it
• click on the "Edit" link at the top of the page and add a new section for your dataset (the datasets are ordered by year of collection)
• click on the "Save page" link at the bottom of the page to save your modifications

Please make sure to provide the following information:

• name of the dataset and year of collection
• authors, institution, contact information
• link to the dataset and to side resources (lexicon, language model, etc)
• short description (nature of the data, license, etc) and link to a paper/report describing the dataset, if any
• at least 1 research result obtained for this dataset (see below)

We currently cannot provide storage space for large datasets. Please upload the dataset at a stable URL on the website of your institution or elsewhere and provide its URL only. If this is not possible, please contact the resources sharing working group.

Contribute a research result

To contribute a new research result, please

• create an account and login
• go to the wiki page and the section corresponding to the dataset for which this result was obtained
• click on the "Edit" link on the right of the section header and add a new item for your result
• click on the "Save page" link at the bottom of the page to save your modifications

Please make sure to provide the following information:

• authors, paper/report title, means of publication
• link to the pdf of the paper
• link to derived data (output transcriptions, intermediary data, etc)
• Code and instructions to reproduce experiments (if available)

In order to save storage space, please do not upload the paper on this wiki, but link it as much as possible from your institutional archive, from another public archive (e.g., arxiv) or from the publisher website (e.g., ieexplore).

We currently cannot provide storage space for large datasets. Please upload the derived data at a stable URL on the website of your institution or elsewhere and provide its URL only. If this is not possible, please contact the resources sharing working group.