McGill-NLP/african_celtic_dataset

Multilingual Speech Dataset (Speech-to-Speech / Speech-to-Text Ready)

An update is planned on January 25th, 2026, to remove and replace some speakers in the Igbo and English train sets. Currently, some speakers (users) might have lower quality audio for some utterances.

The test set will be published in April, in accordance with IWSLT 2026 shared task protocol.

Dataset Summary

This dataset is a large-scale multilingual speech corpus curated for speech-to-speech translation, speech-to-text, and multilingual speech processing research. The data is organized by language, speaker (user_id), and dataset split (train, dev), and includes rich acoustic and metadata annotations.

The dataset is published and maintained by Marie Maltais at mcgill-NLP.

---

Languages

The dataset currently includes speech data in the following languages:

• Yoruba (Y)
• Igbo (I)
• Hausa (H)
• English (E)

Additional languages may be added in future releases.

---

Dataset Organization

Hugging Face Representation

On Hugging Face, the dataset is exposed as a flattened DatasetDict with the following splits:

• train
• dev

Each row corresponds to a single utterance, with speaker, language, text, and acoustic metadata preserved.

---

Data Fields

Each example contains the following fields (aligned with dataset_info):

Field	Type	Description
audio	Audio (48 kHz)	Speech waveform
user_id	string	Speaker identifier
language	string	Language of the utterance
text_id	string	Identifier for the spoken text
text	string	Reference transcription
duration	float	Audio duration (seconds)
recorded_at	string	Recording timestamp or session info
original_sample_rate	int	Original sampling rate before resampling
silence_ratio	float	Ratio of silence frames
snr_db	float	Signal-to-noise ratio (dB)
speech_rate	float	Estimated speech rate
volume_db	float	Average volume (dB)
split	string	Dataset split (train, dev)

---

Audio Specifications

• Sampling rate (served): 48,000 Hz
• Format: WAV
• Channels: Mono

Original sampling rates are preserved in the original_sample_rate field.

---

Intended Uses

Primary Use Cases

• Speech-to-speech translation
• Speech-to-text and ASR
• Multilingual and low-resource speech modeling

---

Data Processing

• Audio segmented at the utterance level
• Speaker boundaries preserved via user_id
• Acoustic metrics computed automatically
• No forced alignment or phonetic labeling applied
• Text is provided as reference only and may contain natural speech variation

---

Licensing

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license unless otherwise stated.

Users must provide appropriate attribution when using this dataset.

---

Citation

If you use this dataset in academic or industrial research, please cite:

@dataset{mcgillnlp_multilingual_speech,
  title     = {IWSLT Shared Task - African-Celtic Speech Datset},
  author    = {Marie Maltais and David Adelani and Min Ma and Elizabeth Salesky and Shamsuddeen Hassan Muhammad},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/mcgill-NLP}
}

---

Contact and Maintenance

For issues, questions, or contributions, please use the Hugging Face dataset repository under mcgill-NLP.