burkimbia/gomdeesa-videos

gom-dɛɛsa videos

Moore speech from Burkinabè films, series, shows, field reports and testimonies on YouTube and Facebook, each clip with the French subtitle burned into its picture, read by OCR, or the English one for a few NGO films and documentaries. The columns of burkimbia/gomdeesa-paires, with the subtitles as shown and french as the target. Private: the videos belong to their channels.

Training data. No test set is held out of these videos (D28). Split by recording, never by row.

Columns

column	meaning
id	videos_locales:<channel>:<video>:<rank>; without its first part, the window's id in burkimbia/gomdeesa-ocr-brut
audio	the clip, 16 kHz mono FLAC
moore	empty: the videos have no written Moore
subtitles	the subtitles shown during the clip, end to end
subtitle_language	french or english, the language they are written in (D30)
context_before, context_after	the subtitles of the windows just before and just after in the same video, kept in this set or not (D34)
french	the target: subtitles when they are French, their French translation when they are English (D34)
french_translated	true when french translates English subtitles, done by Claude with the two neighbouring lines (D34)
is_music	true for a song sung in Moore with its lyrics translated on screen (D33)
source	videos_locales
recording	URL of the video, YouTube or Facebook
start, end	span of the clip in the video (s)
duration	length of the clip (s), 5 to 30

Clips of one video share recording.

How the rows were made

Each video was read by EasyOCR at two frames a second, its subtitles reread by Qwen3-VL-2B, logos, credits and scrolling text taken out, and the subtitles less than 2 s apart packed into windows of 5 to 30 s (D16, D18, D19). A window is kept when its text is written in French or English, when MMS-LID does not hear it as Dioula or Bambara above 0.5, and when it is clean (D23): not flagged by the checks without a model, not a scene already said in another video (S61). Then out go the windows whose text is a credit line, a news ticker or has a letter French does not write, those MMS-LID hears as French as a whole, and those with a piece of about 3 s that mms-lid-256 hears as French above 0.5 (S81). A video spoken in Dioula is left out whole.

Known limits (S57): on 39 windows a Moore speaker listened to, 29 had text matching the sound, 3 were cut at one end and 7 did not match; sketch titles of Ah! Ouaga! still pass as text.

french: 91.98 h ; english: 8.54 h ; songs among them: 0.62 h.