WazobiaVoice TTS (wazobia-tts-cc)

A transcribed speech dataset spanning five Nigerian language varieties, built for text-to-speech and speech-recognition work: Yoruba, Hausa, Igbo, Nigerian-accented English, and Nigerian Pidgin.

Dataset summary

• Total clips: 5866
• Total audio: ~18.0 hours
• Format: single-speaker short clips, transcribed, filtered for quality and language accuracy
• License: CC-BY-4.0

Language	Clips	Share
Hausa (ha)	2360	40.2%
Nigerian Pidgin (pcm)	1170	19.9%
Yoruba (yo)	1141	19.5%
Nigerian-accented English (en)	1131	19.3%
Igbo (ig)	64	1.1%

Dataset structure

Each row contains:

Field	Type	Description
audio	Audio	The speech clip (embedded, 24kHz)
text	string	Transcription
lang	string	Language code (yo, ha, ig, en, pcm)
duration	float	Clip duration in seconds
source_file	string	Originating source recording, for provenance/traceability

Collection & processing

Audio is sourced from radio and manually-provided recordings, then transcribed with Axiveri/NaijaVox-2.0, a Whisper large-v3 fine-tune covering Yoruba, Hausa, Igbo, Nigerian Pidgin, and Nigerian-accented English.

Filtering, applied per-language rather than uniformly:

• All languages: empty transcriptions and Whisper's repeated-token hallucination pattern are dropped.
• Yoruba / Hausa / Igbo: additionally filtered for English code-switching (a speaker dropping into full English mid-clip), since these are monolingual targets.
• Pidgin / English: the code-switch filter is not applied -- Nigerian Pidgin is English-lexified at its core, so the same filter would incorrectly discard most genuine Pidgin speech.

Known limitations

• Source material is unevenly distributed across languages; Pidgin in particular currently comes from a single manually-sourced recording rather than a dedicated radio stream. A better Pidgin radio source is an open item.
• Transcriptions are model-generated (NaijaVox-2.0), not human-verified; expect the normal error rate of an ASR system, higher on noisier source audio.
• source_file is retained for provenance, not for balanced sampling -- don't assume uniform distribution across source recordings.

Acknowledgements

Transcription for this dataset was produced using Axiveri/NaijaVox-2.0, a Whisper-large-v3 LoRA fine-tune (fully merged, Apache-2.0) covering Yoruba, Hausa, Igbo, Nigerian Pidgin, and Nigerian-accented English, with dedicated <|pcm|> and <|ig|> vocabulary tokens purpose-built for these languages. NaijaVox-2.0 reports the following WER on FLEURS test sets:

Language	WER
Nigerian Pidgin	14.7%
Nigerian English	19.6%
Yoruba	22.3%
Hausa	25.8%
Igbo	30.5%

Full credit to the NaijaVox-2.0 model and its authors for making this transcription pipeline possible.

Citation / acknowledgement

If you use this dataset, please credit Axiveri / Ememzyvisuals Digitals (Africlaude AI) and link back to this repository.