AfriSpeech/afrispeeches-aligned-v2

AfriSpeeches Aligned v2 (transcript-verified)

A cleaned version of AfriSpeech/afrispeeches-aligned - sentence-aligned, Resemble-enhanced speeches by African heads of state and leaders (English, French, Portuguese, Arabic) - with clips whose transcript does not match the audio removed.

What changed from v1

Every clip was transcribed a second time with omniASR-CTC-1B-v2 (Meta Omnilingual ASR) and compared with its v1 transcript. A clip is kept when all of these hold:

• Transcript agrees with the audio: CER(v1 transcript vs omniASR) <= 50% (<= 75% when the transcript contains digits, since omniASR spells numbers out). Removes wrong or garbled transcripts and unusable audio.
• Plausible speaking rate: transcript characters per second within its language's robust band (median +/- 3 MAD, log space). Removes clips whose transcript covers only part of the audio (alignment cut-offs).
• Duration <= 20 s.
• Correct language: the two Equatorial Guinea speakers labelled Portuguese in v1 (Teodoro Obiang Nguema Mbasogo, Francisco Macías Nguema) speak Spanish and are removed.

The threshold is deliberately lenient: strongly accented speech is where ASR is weakest, and this corpus is valued for its accents, so only clearly broken pairs are dropped. The v1 text is kept unchanged; the omniASR transcript and the CER are added as columns so you can filter more strictly.

Split	v1 clips	v2 clips	kept	v2 hours
train	128,464	113,691	88.5%	156.7
validation	22,225	19,624	88.3%	27.1

Language	Clips	Hours
ar	7,103	9.6
en	47,451	63.9
fr	63,313	88.4
pt	15,448	21.9

Loading

Same layout and configs as v1 (default, one config per leader, Country_person_lang for bilingual leaders); configs left empty by the filter are removed.

from datasets import load_dataset
ds = load_dataset("AfriSpeech/afrispeeches-aligned-v2", split="train", streaming=True)
strict = ds.filter(lambda r: r["omni_cer"] <= 25)   # stricter subset

Fields

All v1 fields (audio, text, speaker, country, iso3, language, split, duration, start_s, end_s, source_chunk_id, title, url) plus:

Field	Description
omni_transcript	omniASR-CTC-1B-v2 transcript of the clip (lowercase, unpunctuated, numbers spelled out)
omni_cer	CER (%) between the v1 transcript and omni_transcript, letters/digits only

Note: omni_cer measures disagreement, not ground truth - for heavily accented or code-switched speech omniASR can be the one that is wrong.

License

CC BY-NC 4.0, as v1.