gom-dɛɛsa pairs
Moore speech clips of 30 s at most, each with the French it says, to train gom-dɛɛsa: French
subtitles learned directly from Moore speech. One table for every source: source names where a
row comes from, and each source is also a configuration, so one can be loaded alone with
load_dataset("burkimbia/gomdeesa-paires", "biblecom_at").
Training data only. The test set is built apart, from local spontaneous speech (D04 in the
project's FINDINGS.md), and no row here belongs to it.
Columns
| column | meaning |
|---|---|
id |
source:reference, unique across the table |
audio |
the Moore clip, 16 kHz mono FLAC |
moore |
the Moore text of the clip, empty for sources that have none |
french |
the French |
source |
where the row comes from, see below |
recording |
URL of the recording the clip is cut from |
start, end |
span of the clip in that recording (s) |
duration |
length of the clip (s), 30 at most |
Clips cut from one recording share recording. Split by it, never by row, so that a chapter or a
video does not end up on both sides.
Sources
| source | content | rows | hours | rights |
|---|---|---|---|---|
biblecom_at |
Old Testament, WẼNNAAM SEBRE 2014 (bible.com version 3058), audio by Davar Partners 2022, timed verse by verse by bible.com. French: Louis Segond 1910 | 22 765 | 64.43 | Moore Bible © Bible Society of Burkina Faso, 2014 |
biblecom_nt |
New Testament, SEBR SÕNGO (bible.com version 1217, catholic, 1998), dramatized reading. French: Louis Segond 1910 | 7 917 | 19.62 | SEBR SÕNGO © Alliance Biblique du Burkina Faso, 1998 |
mooreburkina_nt |
New Testament of the Moore Bible app of mooreburkina.com (media.ipsapps.org/mos2/osa), audio by Faith Comes By Hearing (MHMBSBN2DA), timed verse by verse by the app. French: Louis Segond 1910 | 7 671 | 20.14 | as published by mooreburkina.com and Faith Comes By Hearing |
mooreburkina_proverbes |
Moore proverbs of mooreburkina.com with their French: read by Thimote, French in brackets (apps p1 to p8), and by Rachida (prv-v10, prv-v11), kept when MMS forced alignment puts all the words of the proverb inside the clip and none of the next phrase's. Three Thimote cues that gloss in French are left out | 1 105 | 2.17 | as published by mooreburkina.com |
bible_jw |
New World Translation read in Moore on jw.org: chapter audio cut verse by verse at the jw.org markers, or by MMS forced alignment for the chapters without them, verses whose cut falls in speech left out. Moore and French verse texts from sawadogosalif/MooreFRCollections_BibleOnlyText | 30 462 | 88.11 | © Watch Tower Bible and Tract Society of Pennsylvania (jw.org) |
revues_livres |
Magazines and books of jw.org read in Moore, cut paragraph by paragraph by MMS forced alignment, each paired with the French paragraph of the same document and paragraph number | 9 772 | 31.41 | © Watch Tower Bible and Tract Society of Pennsylvania (jw.org) |
videos_jw |
JW videos dubbed in Moore (jw.org), kept when the Moore and French durations agree within 0.2 s; windows of whole French sentences of about 15 s, 30 s at most, cut at the French subtitle times on the Moore audio; windows whose every sentence was in an earlier video left out. No Moore text | 53 401 | 168.63 | © Watch Tower Bible and Tract Society of Pennsylvania (jw.org) |
spg |
Sid Pa Gilmde, a bilingual TV show, episodes 1 to 5: Moore sentences force-aligned with MMS, whole ones only (keep, not trimmed), with their French, from burkimbia/spg_series | 654 | 1.16 | CC BY-NC 4.0, as burkimbia/spg_series |
| total | 133 747 | 395.67 |
The Louis Segond of 1910 is in the public domain. A Bible pair is a verse; where the two Bibles cut a passage differently, a pair joins neighbouring verses.
How it is made
scripts/03-corpus-jw-et-bible/paires_audio.py in BurkimbIA/bia-speech-translation downloads
each recording once, decodes it to 16 kHz mono with ffmpeg and cuts each pair at its timings; pairs
longer than 30 s are left out. The pairs come from biblecom.py, nt_mooreburkina.py, mooreburkina.py and bible_jw.py paires, videos_jw.py paires and spg.py gomdeesa in the same folder, and from the paragraph corpus of passage_coupes.py corpus.
Counts and checks are in the project's FINDINGS.md (S19, S28, S30, S31, S34, S35, S36, S37, S38, S39, S40, S41, D14).
- Downloads last month
- 21