ghananlpcommunity/twi-speech-sota-240hrs

🇬🇭 Twi ASR Dataset

A speech dataset of Twi (Akan) extracted from Ghanaian news media broadcasts, designed for training and fine-tuning Text-To-Speech (TTS) models.

---

📂 Dataset Structure

Column	Type	Description
audio	Audio	24 kHz mono WAV audio segment
text	string	Verbatim Twi transcription of the audio segment
duration	float	Duration of the audio segment in seconds

---

📊 Statistics

Metric	Value
Total clips	132,212
Total duration	237.71 hours
Mean clip duration	6.47 s
Min / Max clip duration	1.01 s / 15.0 s
Mean words per clip	16.0
Min / Max words	1 / 16
Vocabulary size	42,970 unique words
Sample rate	24,000 Hz (mono)

---

🚀 Usage

from datasets import load_dataset

dataset = load_dataset("ghananlpcommunity/twi-speech-sota-240hrs")
train = dataset["train"]

example = train[0]
print("Transcription:", example["text"])
print("Duration (s):", example["duration"])
print("Audio array shape:", example["audio"]["array"].shape)
print("Sample rate:", example["audio"]["sampling_rate"])

---

🎯 Intended Use Cases

• Building TTS models from scratch or finetuning for Twi (Akan)
• Linguistic research on Twi phonology and prosody
• Low-resource African language ASR benchmarking

---

📜 Citation

@dataset{twi_asr,
  author    = {Owusu, Mich-Seth},
  title     = {Twi ASR Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {[https://huggingface.co/datasets/](https://huggingface.co/datasets/)ghananlpcommunity/twi-speech-sota-200hrs}
}

---

🙏 Acknowledgments

Created by Mich-Seth Owusu for the Ghana NLP Community.