Datasets:

Professor
/

yoruba-speech-data

language:
  - yo
license: unknown
multilinguality: monolingual
task_categories:
  - text-to-speech
  - automatic-speech-recognition
tags:
  - yoruba
  - speech
  - tts
  - asr
  - code-switching
  - african-languages
  - low-resource
pretty_name: Yoruba Speech Data (Pooled)
size_categories:
  - 100K<n<1M

Yoruba Speech Data (Pooled)

A ~1,039-hour pooled Yoruba speech corpus, combining four independent sources into one consistently-formatted dataset for speech modeling (TTS / ASR). Built to give a Yoruba TTS finetune enough scale and diversity to move past what a single ~100-hour source can teach a model — more speakers, more domains, more of the language.

Sources

Source	Clips	Hours	Style
DSN African Voices (Data Science Nigeria)	144,628	309.0 h	spontaneous, multi-speaker, includes SNR metadata
NaijaVoices	610,635	614.0 h	read + spontaneous, multi-speaker, culturally-rich prompts
YECS (LyngualLabs)	89,981	107.5 h	Yoruba-English code-switching, 140 speakers, 16 semantic domains
WAXAL (`yor_tts`, Google)	1,993	8.1 h	spontaneous
Total	847,237	1,038.7 h

All audio is standardized to 16 kHz mono FLAC (lossless). Clips are 1–30 seconds; empty/garbage transcripts and undecodable audio were dropped at ingestion. Text is lightly normalized (Unicode NFC, whitespace/punctuation cleanup) but tones, digits, and code-switching are preserved — nothing is stripped or transliterated.

Format

The dataset ships as WebDataset-style tar shards (shards/shard-00000.tar … shard-00064.tar, ~1 GB each, one {key}.flac file per clip) plus a single manifest (manifest.parquet / manifest.jsonl) that indexes every clip:

Column	Description
`key`, `shard`	which tar file + entry holds this clip's audio
`text`	transcript (native script, tones/code-switch preserved)
`duration`	seconds
`source`	`dsn` \| `naijavoices` \| `yecs` \| `waxal` — which corpus this clip came from
`dataset_id`	integer id per source (0=dsn, 1=naijavoices, 2=yecs, 3=waxal)
`split`	`train` / `val` (300 clips held out per source for evaluation)
`speaker_id`, `gender`	speaker metadata where available
`snr`	signal-to-noise ratio in dB (DSN only; `null` elsewhere)
`dbfs`, `clip_ratio`, `sil_ratio`	cheap DSP quality proxies computed for every clip: loudness, fraction of clipped samples, fraction of near-silent frames — useful for filtering a "clean subset" without running a neural MOS model over the whole corpus
`has_disfluency`	transcript contains a bracketed disfluency marker (e.g. `[um]`)

Usage

from huggingface_hub import hf_hub_download
import pandas as pd, tarfile, io, soundfile as sf

# 1. grab the manifest (small, ~260 MB) to browse/filter before downloading audio
mp = hf_hub_download("Professor/yoruba-speech-data", "manifest.parquet", repo_type="dataset")
df = pd.read_parquet(mp)

# e.g. a "clean" subset for a polish/finetune pass: quiet, unclipped, non-silent
clean = df[(df.dbfs > -25) & (df.clip_ratio < 0.001) & (df.sil_ratio < 0.5)]

# 2. pull one shard on demand and decode a clip
row = df.iloc[0]
shard_path = hf_hub_download("Professor/yoruba-speech-data", f"shards/{row.shard}", repo_type="dataset")
with tarfile.open(shard_path) as tar:
    audio_bytes = tar.extractfile(f"{row.key}.flac").read()
arr, sr = sf.read(io.BytesIO(audio_bytes))

The tar shards are also directly readable by the webdataset library for streaming training pipelines.

Intended use & limitations

Built for Yoruba TTS/ASR research, in particular as pooled finetuning data for a multilingual TTS model that doesn't natively support Yoruba. It mixes spontaneous, read, and code-switched speech across many speakers and recording conditions — expect variable audio quality (see the DSP quality columns to filter). This is a research aggregation; usage should respect the terms of each constituent source below.

License

This pooled release does not impose an additional license beyond what each source provides. Consult each source's own page/terms before commercial use: DSN African Voices · NaijaVoices · YECS · WAXAL.

Citations

If you use this pooled dataset, please cite the original sources it draws from:

@misc{datasciencenigeria_african_voices_2025,
  title        = {African Voices: Multilingual Speech Dataset for Low-Resource African Languages},
  author       = {DataScience Nigeria},
  year         = {2025},
  note         = {Latest release, November 2025},
  howpublished = {\url{https://www.africanvoices.ai}},
  institution  = {Data Science Nigeria},
  keywords     = {speech recognition, multilingual datasets, African languages, low-resource ASR}
}

@article{emezue2025naijavoices,
  title   = {The NaijaVoices Dataset: Cultivating Large-Scale, High-Quality, Culturally-Rich Speech Data for African Languages},
  author  = {Emezue, Chris and Community, NaijaVoices and Awobade, Busayo and Owodunni, Abraham and Emezue, Handel and Emezue, Gloria Monica Tobechukwu and Emezue, Nefertiti Nneoma and Ogun, Sewade and Akinremi, Bunmi and Adelani, David Ifeoluwa and others},
  journal = {arXiv preprint arXiv:2505.20564},
  year    = {2025}
}

@misc{lynguallabs_yecs_2026,
  title        = {{YECS}: A 120-Hour Community-Curated Yoruba-English Code-Switching Corpus},
  author       = {{LyngualLabs}},
  year         = {2026},
  note         = {140 speakers; 16 semantic domains; word-level language tags},
  howpublished = {\url{https://lynguallabs.org/yecs}}
}

@article{waxal2026,
  title   = {WAXAL: A Large-Scale Multilingual African Language Speech Corpus},
  author  = {Anonymous},
  journal = {arXiv preprint arXiv:2602.02734},
  year    = {2026}
}

If you use models or benchmarks built on this data as part of the WAXAL edge-TTS/ASR effort, please also consider citing the collective's own benchmark work:

@article{waxalnet2026,
  title  = {The WAXAL ASR Benchmark: Fine-Tuned Edge Models Across 19 African Languages},
  author = {Olufemi, Victor Tolulope and Babatunde, Oreoluwa and Njema, Ramsey and
             Gbotemi, Bolarinwa and Yen, Wanchi Lucia and Uzodinma, John and
             Ajayi, Sunday and Williams, Oluwademilade and Moshood, Kausar and
             Anyaele, Innocent Elendu and Arefaine, Akebert Tesfahunegn and
             Hunzwi, Candace and Daniel, Wongel Dawit and Namuganga, Emmilly Immaculate and
             Kadima, Cleophas and Bahizire, Athanase Biluge and Ranaivoson, Onitsiky and
             Aaron, Emmanuel and Ladislaus, Nicholaus Dismas and Muhammed, Idris and
             Simenya, Jonathan Enoch and Koome, Martin and Endaylalu, Matewos Tegete and
             Adeyemo, Peter Ifeoluwa and Birindwa, Hondi Prisca and Eze-Mbey, Ukachi Agnes and
             Oduro-Yeboah, Yacoba and Aremu, Toluwani and Adjovi, Pericles and
             Ngueajio, Mikel K and Mitra, Prasenjit},
  year   = {2026},
  note   = {arXiv preprint arXiv:2606.02375}
}

Acknowledgments

Deep thanks to the teams and communities behind every source dataset that made this pooled corpus possible:

Data Science Nigeria — for the African Voices corpus and its careful per-clip metadata (speaker demographics, domain, SNR).
The NaijaVoices team and community (Chris Emezue, Busayo Awobade, Abraham Owodunni, Handel Emezue, Gloria Monica Tobechukwu Emezue, Nefertiti Nneoma Emezue, Sewade Ogun, Bunmi Akinremi, David Ifeoluwa Adelani, and the wider NaijaVoices community) — for building one of the largest, culturally-grounded Nigerian speech datasets to date.
LyngualLabs and the 140 YECS speakers — for the YECS Yoruba-English code-switching corpus.
Google — for the WaxalNLP TTS data.
And the many, many speakers across all four sources whose voices actually make this dataset what it is.

This dataset was pooled by Victor Olufemi and LyngualLabs as part of an independent Yoruba TTS finetuning effort. It also draws on and gratefully acknowledges the work of the WAXAL Research Collective, whose edge-ASR benchmark (Olufemi, Babatunde, Njema, Gbotemi, and the full collective, 2026) this effort sits alongside.