ghananlpcommunity/twi-health-asr-gemini-500hrs

This dataset is shared under CC BY-NC 4.0, which means you are free to use, share, and adapt it for non-commercial research and educational purposes with attribution. You can read the full license at https://creativecommons.org/licenses/by-nc/4.0/.

Twi Health Speech Dataset Gemini (500 hours)

A domain-specific speech recognition dataset for Twi, one of Ghana's most widely spoken languages, sourced from publicly available video content on health and wellness. Created by Mich-Seth Owusu and published under the GhanaNLP Community to support the development of open-source ASR tools for Ghanaian languages.

This is a higher-quality re-transcription of the audio in `ghananlpcommunity/twi-health-asr`: the same health-domain Twi audio, but transcribed with Google Gemini instead of the Google Speech Recognition API, which produces noticeably better Twi transcriptions.

Dataset Details

Data Source

Audio was extracted from publicly available videos on health and wellness — covering topics such as general health education, herbal and traditional medicine, nutrition, fitness, and health talk shows delivered in Twi. The videos were chunked into segments of up to 30 seconds and transcribed to form this parallel audio-text dataset.

Speech is natural and conversational, and includes the code-switching between Twi and English that is characteristic of real-world Ghanaian speech.

Dataset Structure

Each example contains:

There is a single train split.

Usage

from datasets import load_dataset

# Stream to avoid downloading everything at once
ds = load_dataset("ghananlpcommunity/twi-health-asr-gemini", split="train", streaming=True)

for example in ds.take(3):
    print(example["transcription"])

Transcription Method & Limitations

Transcriptions were generated automatically using Google Gemini (`gemini-3.5-flash`), prompted to transcribe the Twi (Akan) speech with correct orthography and punctuation. Gemini produces substantially more accurate and natural Twi transcriptions than the Google Speech Recognition API used for the original twi-health-asr dataset.

Cleaning applied

• —Empty transcriptions removed (silence, music, or no recognizable speech).
• —Runaway-repetition outputs removed. Large language models occasionally derail into a repetition loop, emitting a single word or short phrase many times in a row. Chunks exhibiting such loops (a word repeated ≥ 10× consecutively, or a long output with a very low ratio of unique words) were dropped, since the audio continues past the point where the model stopped tracking and the pair can no longer be trusted.

⚠️ Despite Gemini's strong performance, these are still machine-generated transcriptions, not human-verified ground truth. Treat them as high-quality silver-standard labels. Manual verification is recommended before using this data in production ASR systems.

Related Datasets

• —`ghananlpcommunity/twi-health-asr` — the same audio transcribed with the Google Speech Recognition API.
• —`ghananlpcommunity/twi-agriculture-asr` — companion agriculture-domain Twi ASR dataset.

Citation

If you use this dataset, please cite:

@misc{twi-health-asr-gemini,
  author       = {Owusu, Mich-Seth},
  title        = {Twi Health Speech Dataset (Gemini transcriptions)},
  year         = {2026},
  publisher    = {GhanaNLP Community},
  howpublished = {\url{https://huggingface.co/datasets/ghananlpcommunity/twi-health-asr-gemini}}
}

Acknowledgements

Created as part of the GhanaNLP Community effort to build open speech resources for Ghanaian languages.