To allow experiments on multimodal automatic speech recognition, audio/visual synchronization or speech-driven lip animation systems, an audio-visual speech corpus, the BL-Database, has been collected by the Metiss team (Panama team since 2013) at INRIA. BL-Database stands for Blue-Lipstick Database because, during recording, speakers were wearing blue lipstick to facilitate the extraction of lips in the video (cf. Figure below).
The recordings have been performed during two sessions. The data of the first session can be used to analyze the 2D movement of the mouth while the data collected by the second session have been performed for 3D analysis.
Database content
Data and information collected and generated during the 2 recording sessions are:
- Session 1 (dedicated to the 2D analysis): 238 utterances enunciated by 4 men and 4 women
- Audio data from 2 sources (each audio file corresponds to one sentence uttered by one speaker),
- Video data recorded by 1 front-view camera (each video file corresponds to one sentence uttered by one speaker),
- Time-aligned phonetic transcription associated to the audio and video data,
- Non-nominative speaker information (age, gender).
- Session 2 (dedicated to the 3D analysis): 238 utterances enunciated by 4 men and 5 women
- Audio data from 2 sources (each audio file corresponds to one sentence uttered by one speaker),
- Video data recorded by 2 spatially calibrated cameras and 1 depth camera (each video file corresponds to one sentence uttered by one speaker and recorded by one camera),
- Time-aligned phonetic transcription associated to the audio and video data,
- Calibration data (23 images of a planar checkerboard),
- Non-nominative speaker information (age, gender).
Sample speakers from the database (front and side view)
More details about the content of the database can be found in :
Y. Benezeth, G. Bachman, G. Lejan, N. Souviraa-Labastie, F. Bimbot, "BL-Database: A French audiovisual database for speech-driven lip animation systems", INRIA research report n RR-7711, 2011.
Requesting the data
The database can be obtained upon request, for research purposes only. To obtain the corpus, please contact Yannick Benezeth.
Data are available here.