9. IPSJ SIG-SLP Corpora and Environments for Noisy Speech Recognition

9-c. Audio-Visual Speech Recognition Evaluation Environment (CENSREC-1-AV)

Data DOI

https://doi.org/10.32130/src.CENSREC-1-AV

Producer, Project

Noisy Speech Recognition Evaluation Working Group,

Special Interest Group on Spoken Language Information Processing,

Information Processing Society of Japan (IPSJ)

Contents

Common platform for evaluating independently speech recognition accuracy and speech interval detection under noisy environment.

An evaluation corpus for audio-visual speech recognition of continuously spoken single digits in Japanese.

The digit sequence of each utterance and the pronunciation of Japanese digits are the same as the CENSREC-1 (AURORA-2J) database.

Including color and infrared mouth images, which were recorded simultaneously with speech.

• Training data
  • Clean audio and visual data in an office environment
• Test data
  • Clean audio and visual data in the office environment
  • Noisy data generated by baseline scripts

    Audio:

    added two types of noise, driving on city streets and driving on expressway, at six different SNR levels

    Image:

    simulated driving conditions using Gamma transform

Speaker

Training data:

42 speakers (22 males, 20 females) 3234 utterances in total

Test data:

51 speakers (25 males, 26 females) 1963 utterances in total

Speech and image file format

Audio: 

WAV format (16kHz, 16bit, Mono)

Image: 

Windows BMP format; 24-bit RGB (color), 8-bit grayscale (infrared)

Distribution media

2 DVDs

Licensing

For research and development purposes only

Price

No fee

Image sample

Digit strings same as CENSREC-1

Color and infrared mouth images recorded simultaneously with speech

Examples of color and infrared pictures (3 frame/sec, upper: color, lower: infrared)