ghananlpcommunity/ghana-named-entities-tts-twi

Ghana Named Entities TTS — Twi

A Twi-language speech dataset built from descriptions of Ghana named entities (people, places, organisations, and concepts). Each audio clip is a synthesised reading of a passage that describes several named entities.

Dataset description

Field	Value
Language	Twi (ISO 639-1: tw)
Domain	Named entities — Ghanaian proper nouns and institutions
Audio format	WAV, 24 000 Hz, mono
Voice	Female, young adult, moderate pitch (synthetic)
Examples	6 704
Split	Single train split

Columns

Column	Type	Description
id	string	Unique identifier (row_0, row_1, …)
audio	Audio	WAV file decoded at native 24 kHz
text	string	Twi transcription aligned to the audio

Generation pipeline

The dataset was created through a four-step pipeline:

1. Source collection — ghana_named_entities.csv contains ~253 000 Ghana named entities with English descriptions scraped and curated by GhanaNLP Community.

2. Chunking — English descriptions were grouped into passages of 10 entities each using combine-descriptions.py, producing combined_descriptions.csv.

3. Translation — Passages were translated into Twi with Gemini (gemini-3.1-flash-lite-preview, thinking level HIGH) via translate.py, yielding translated_twi.csv.

4. Speech synthesis — Twi passages were converted to speech with OmniVoice (k2-fsa/OmniVoice) running on a Modal T4 GPU via generate-speech.py:

   • Voice instruct: female, young adult, moderate pitch
   • Language code: abr (Akan/Twi dialect code accepted by OmniVoice)
   • Diffusion steps: 16, guidance scale: 2.0
   • Output sample rate: 24 000 Hz

Usage

from datasets import load_dataset

ds = load_dataset("ghananlpcommunity/ghana-named-entities-tts-twi", split="train")
sample = ds[0]
print(sample["text"])
# sample["audio"]["array"] → numpy array at 24 000 Hz

Intended uses

• Training and evaluating Twi TTS systems
• Training and evaluating Twi ASR systems
• Low-resource African-language speech research
• Named-entity pronunciation modelling for Ghanaian proper nouns

Limitations

• Audio is synthetic (not human-recorded); prosody and pronunciation may contain TTS artefacts.
• Each clip covers multiple entities concatenated into a passage rather than isolated utterances.
• Translation was automated; some Twi renderings may not be idiomatic.

Citation

If you use this dataset please cite:

@dataset{ghananlp2024twi_tts_named_entities,
  author    = {GhanaNLP Community},
  title     = {Ghana Named Entities TTS -- Twi},
  year      = {2024},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/ghananlpcommunity/ghana-named-entities-tts-twi}
}

License

CC BY 4.0