Professor/swahili-speech-data

Swahili Speech Data (Pooled)

A ~1,004.1-hour pooled Swahili speech corpus, combining two independently-sourced datasets into one consistently-formatted corpus for speech modeling (TTS / ASR). Part of the AfroNet multi-language TTS data effort — sibling release to Yoruba/Hausa/Igbo/Kinyarwanda, sourced entirely differently: DSN African Voices and NaijaVoices don't cover Swahili at all.

Locale is Kenyan Swahili (sw_KE) specifically, per the Afrivoice source's own metadata — not necessarily representative of Tanzanian or other regional varieties.

Sources

Source	Clips	Hours	Style
Afrivoice Swahili (Digital Umuganda)	180,527	999.9 h	crowdsourced spoken image descriptions, 5 domains (agriculture/education/financial/government/health), 200h each
WAXAL (swa_tts, Google)	1,230	4.2 h	spontaneous
Total	181,757	1,004.1 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.

On domain selection and text handling

All 5 Afrivoice domains were sampled evenly (200h each) — unlike the Kinyarwanda sibling release, no domain was excluded or hit a natural hour ceiling; all 5 checked out cleanly on inspection (consistently ~10–12 characters/second across every domain, no sign of the duration/transcript mismatch found in Kinyarwanda's Scripted Education domain). Afrivoice's Education-domain transcripts sometimes wrap code-switched phrases in (sc)...(cs) markers (e.g. "(sc)water is life(cs)"); these markers are stripped during ingestion, keeping the enclosed text.

Per WAXAL's own swa_tts release, all three original partitions (train+ validation+test) are pooled together here — same policy as the other AfroNet languages; see the Yoruba/Hausa/Igbo cards' "note on WAXAL's original partitions" for the reasoning, and the ASR-specific caveat if you need clean splits for benchmarking instead.

Format

The dataset ships as WebDataset-style tar shards (shards/shard-00000.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). For Afrivoice this is the source's transcription field (natural capitalization/punctuation), not its lowercased/stripped normalized_transcription
duration	seconds
source	afrivoice | waxal
dataset_id	integer id per source (0=afrivoice, 1=waxal)
split	train / val (250 clips held out per source for evaluation) — this pool's own holdout, unrelated to WAXAL's original train/validation/test labels (see above)
speaker_id, gender	speaker metadata where available (Afrivoice doesn't expose per-speaker IDs)
domain	Afrivoice's source domain (agriculture/education/financial/government/health); null for waxal
dbfs, clip_ratio, sil_ratio	cheap DSP quality proxies computed for every clip: loudness, fraction of clipped samples, fraction of near-silent frames. Unlike the Kinyarwanda release, Afrivoice_Swahili's manifest doesn't include a LUFS field, so these proxies are the only quality signal available here
has_disfluency	always false — neither source flags disfluencies

Usage

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

mp = hf_hub_download("Professor/swahili-speech-data", "manifest.parquet", repo_type="dataset")
df = pd.read_parquet(mp)

row = df.iloc[0]
shard_path = hf_hub_download("Professor/swahili-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 Swahili TTS/ASR research, in particular as pooled finetuning data for a multilingual TTS model that doesn't natively support Swahili. The bulk of this corpus (Afrivoice) is speech describing photographs across five institutional domains — a fairly narrow register (descriptive, matter-of-fact) compared to natural conversation or narrative speech; there's no small studio-quality anchor source here the way YECS (Yoruba) or the mbazaNLP corpus (Kinyarwanda) provide for their languages. Locale is specifically Kenyan Swahili. This is a research aggregation; usage should respect the terms of each constituent source below.

License

Both constituent sources are CC BY 4.0. Consult each source's own page for full terms: Afrivoice Swahili · WAXAL.

Citations

If you use this pooled dataset, please cite the original sources it draws from — consult each source's own HuggingFace page for their preferred citation.

@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 Digital Umuganda for the large-scale Afrivoice Swahili image-description corpus across five domains, and to Google for the WaxalNLP TTS data.

This dataset was pooled by Victor Olufemi and LyngualLabs as part of the AfroNet multi-language TTS data effort.