MooreSpeechYouTube is a raw internal corpus of GO AI Corp, kept until it is rebuilt. It is not open for sharing or public use, and access requests are not approved.

MooreSpeechYouTube (raw)

Mooré speech from YouTube news and talk shows, with automatic transcripts. Internal, raw, to be rebuilt.

Internal use only: not for sharing, and access requests are not approved. The audio comes from public YouTube videos and belongs to their owners, and the transcripts are not reliable.

En bref. Corpus interne de parole en mooré tirée de YouTube (journaux et émissions), avec des transcriptions automatiques peu fiables. Il sera refait avec FasoSpeechCorpora. Non destiné au partage.

What it is

17,764 clips, 41.90 hours, 44.1 kHz.
Collected in 2024 with the first MooreSpeechCorpora toolkit: videos cut at pauses and transcribed by a speech model.
The transcripts are machine-made and often wrong. Some are hallucinated (repeated syllables), some clips hold several speakers, and some transcripts are too long or too short for their clip.
The channels and video ids were not recorded.

It will be rebuilt with FasoSpeechCorpora (better segmentation, two-model transcript checks, speaker labels), or dropped.

Columns

id, audio, text (automatic transcript), duration, sample_rate, origin (earlier dataset and row), audio_IDs where available, and qc_flags:

Flag	Clips	Meaning
`possible_hallucination`	839	Repeated syllables or n-grams in the transcript; listen first
`transcript_too_short` / `transcript_too_long`	463 / 410	Characters per second outside the normal range
`empty_text`	82	No transcript
`same_text_as_other_clip`	236	Another clip has the same transcript (often a hallucination)

16,383 clips (37.5 h) carry no flag.

License

The audio belongs to the owners of the YouTube videos. This copy is kept for internal research only and is not shared.

Contacts

Wendyellé A. Alban NYANTUDRE: nyantudre@gmail.com
GO AI Corp

Downloads last month: -

Total file size:

13.3 GB