offiongbassey/efik_speech_dataset

Efik Text-to-Speech and ASR Dataset

This dataset contains Efik speech recordings aligned with textual transcriptions, designed for Text-to-Speech (TTS) and Automatic Speech Recognition (ASR) research and development.

It represents a significant expansion of the initial corpus introduced in the Interspeech 2026 paper "Towards Digital Preservation of Efik: TTS for a Low-Resource African Language." While the original paper documented a single-speaker, 3-hour corpus, this updated release contains two speakers and 8.79 hours of speech, recorded in a quiet environment.

All recordings were manually transcribed and reviewed by linguists and native speakers to ensure high linguistic fidelity and tonal accuracy. For more detail on the language, the baseline corpus, and the TTS experiments, please refer to the paper.

What's Inside

Each example includes:

Field	Description
audio	Speech waveform, 16 kHz, mono
text	Manual transcription
speaker_id	Speaker identifier (1 or 2)

What It's For

This dataset is intended for:

• Training and fine-tuning TTS models for Efik
• Training and fine-tuning ASR models for Efik
• Benchmarking speech models on low-resource tonal languages
• Linguistic research on Efik phonology and tonology

It is not intended for voice cloning, commercial use without permission, or redistribution.

How It Was Built

The recordings were collected from native Efik speakers in a quiet environment. Transcription was done manually, then reviewed by linguists and native speakers to verify tonal markings and orthography. Every audio-text pair was checked for alignment.

The text sources include encyclopaedic and informational content, as well as everyday conversational phrases.

Known Limitations

• The dataset is small by TTS/ASR standards. It is a foundation, not a finished product.
• Efik is tonal. Current models struggle with tonal accuracy in both synthesis and recognition.
• Two speakers only. It does not capture the full dialectal diversity of Efik.
• ASR models trained on this data may not generalize to noisy environments, conversational speech, or unseen speakers.

Citation

If you use this dataset, please cite:

@inproceedings{edet2026efik,
  title={Towards Digital Preservation of Efik: TTS for a Low-Resource African Language},
  author={Edet, Offiong Bassey and Oyo-Ita, Emmanuel and Archibong, Archibong and Effanga, David and Awak, Mbuotidem},
  booktitle={Interspeech},
  year={2026},
  url={https://arxiv.org/abs/2607.04515}
}