ghanaopenai/twi-agriculture-speech

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 Agriculture Speech Dataset

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

Dataset Details

Data Source

Audio was extracted from publicly available videos on agriculture — covering topics such as farming practices, crop cultivation, and agri-extension content delivered in Twi. These videos were chunked into segments and transcribed to form this parallel audio-text dataset.

Transcription Method & Limitations

Transcriptions were generated automatically using the Google Speech Recognition API (free tier) via the `SpeechRecognition` Python library.

Google's Speech-to-Text service reports a Word Error Rate (WER) of approximately ~7–10% for high-resource languages such as English under clean acoustic conditions (Google, 2023). However, for low-resource languages like Twi, WER is expected to be significantly higher, as Twi is underrepresented in Google's training data and lacks a robust dedicated acoustic model. Domain-specific terminology (agricultural vocabulary) may further increase the error rate.

⚠️ Users should treat these transcriptions as noisy silver-standard labels, not gold-standard ground truth. Manual verification is strongly recommended before using this data in production ASR systems.

Why release it anyway?

This dataset is deliberately released in its current, imperfect state because:

• —There is a severe lack of any Twi ASR data — even noisy data is a valuable starting point.
• —It provides a baseline resource that the community can iteratively improve through correction, augmentation, and re-annotation.
• —It can be used to fine-tune models (e.g. Whisper, wav2vec2) that can themselves reduce WER further on Twi, creating a positive feedback loop for the community.
• —Agriculture is a critical domain for Ghana, and building speech tools in this space has real practical value for farmers and agri-extension workers.

Contributions to correct or extend the transcriptions are warmly welcomed.

Dataset Structure

Each example contains:

• —audio — the raw audio waveform (decoded by the datasets library)
• —transcription — the corresponding Twi text (auto-generated, see note above)

DatasetDict({
    train: Dataset({
        features: ['audio', 'transcription'],
        num_rows: ...
    })
})

Usage

from datasets import load_dataset

ds = load_dataset("ghananlpcommunity/twi-agriculture-asr")

sample = ds["train"][0]
print(sample["transcription"])
# play sample["audio"]["array"] at sample["audio"]["sampling_rate"]

Intended Use

This dataset is intended for:

• —Training and fine-tuning ASR models for Twi, especially in the agriculture domain
• —Bootstrapping better transcription pipelines for Ghanaian languages
• —Benchmarking low-resource speech recognition systems
• —Supporting agri-tech and language preservation efforts in Ghana

Community

This dataset is published under the GhanaNLP Community — a collective effort to build natural language processing resources for Ghanaian languages and make AI more accessible to Ghanaian language speakers.

• —🌍 GhanaNLP on Hugging Face
• —🐙 GhanaNLP on GitHub

Citation

If you use this dataset, please cite:

@dataset{owusu2026twi_agri,
  author       = {Mich-Seth Owusu},
  title        = {Twi Agriculture Speech Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  organization = {GhanaNLP Community},
  url          = {https://huggingface.co/datasets/ghananlpcommunity/twi-agriculture-asr},
  note         = {Transcriptions generated via Google Speech Recognition API (free tier).
                  Expected WER is higher than reported averages for low-resource languages.}
}

License

This dataset is released under CC BY-NC 4.0 — free to use, share, and adapt for non-commercial research and educational purposes with attribution.