ghananlpcommunity/ghana-one-voice
Ghana One Voice Speech in 43 Ghanaian and West African languages, all converted into a single consistent voice. Every clip keeps its original transcript, so the dataset pairs one speaker's voice with the phonetic range of dozens of languages. Roughly 5 hours per language, about 215 hours in total. What this is Source audio comes from many different speakers, recording conditions and microphones. Each clip has been passed through ghana-vc, a voice-conversion model… See the full description on the dataset page: https://huggingface.co/datasets/ghananlpcommunity/ghana-one-voice.
Ghana One Voice
Speech in 43 Ghanaian and West African languages, all converted into a single consistent voice.
Every clip keeps its original transcript, so the dataset pairs one speaker's voice with the phonetic range of dozens of languages. Roughly 5 hours per language, about 215 hours in total.
What this is
Source audio comes from many different speakers, recording conditions and microphones. Each clip has been passed through ghana-vc, a voice-conversion model, so the speaker is the same throughout while the speech — the words, the language, the prosody — is unchanged.
Voice conversion operates on speaker timbre rather than on words, so the conversion is not tied to any particular language.
Hear the voice: ghana-vc demo
Languages
43 languages, 107,805 clips.
Usage
Each language is a separate config:
from datasets import load_dataset
ds = load_dataset("ghananlpcommunity/ghana-one-voice", "Ewe_ewe", split="train")
ds[0]["audio"] # converted audio, 22.05 kHz
ds[0]["text"] # transcript
ds[0]["universal_text"] # universal graphemes, or NoneList every config with get_dataset_config_names("ghananlpcommunity/ghana-one-voice").
Columns
universal_text
An extra column carrying the transcript mapped onto a shared cross-language grapheme inventory with africa-g2p, so the same sounds are written the same way across languages. Every language here has it. The column is present in every config, so subsets concatenate cleanly.
How it was made
- Source: ghananlpcommunity/ghana-speech for 42 languages, and ghanaopenai/ghana-english-tts-clean2 for Ghanaian English
- Model: ghanaopenai/ghana-vc, a Seed-VC fine-tune
- Settings: 50 diffusion steps, clips between 1 and 20 seconds
- Compute: one HF Job per language on an L4, about 1.7 GPU-hours each
Reproduce a single language with GhanaNLP/ghana-vc:
ghana-vc convert \
--dataset ghananlpcommunity/ghana-speech \
--config-name Ewe_ewe \
--output <your-org>/<your-dataset> \
--diffusion-steps 50Limitations
- The target voice is synthetic, from a TTS speaker, so converted audio carries that character.
- Transcripts are passed through unchanged from the source datasets and have not been re-verified here.
- Clips outside 1–20 seconds were skipped, so this is not a complete copy of the source corpora.
- Languages are capped at 5 hours, which is a sample rather than everything available upstream.
Licence
CC-BY-NC-4.0 — non-commercial.
The voice-conversion model derives from Seed-VC, which is GPL-3.0, and was fine-tuned on a CC-BY-NC-4.0 corpus. Source audio remains under the terms of the datasets it came from. Check all of these before any use beyond research.
Credits
- Source audio: Ghana NLP Community and Ghana OpenAI
- Voice conversion: Seed-VC by Plachtaa
- Graphemes: africa-g2p
