Professor/maasai-speech-data

Maasai Speech Data (Pooled)

A ~83.7-hour Maasai speech corpus, drawn from a single source (African Next Voices) and filtered to only genuinely transcribed audio. Part of the AfroNet multi-language TTS data effort.

Source

Anv-ke/Maasai — African Next Voices, a pilot data-collection effort in Kenya led by the KenCorpus Consortium (a coalition of Kenyan universities and research centers), funded by the Gates Foundation. 31,762 clips, 83.7h, source dataset_id/source = anv_ke.

A naming note: WAXAL (google/WaxalNLP) also has a config called mas, but that config is actually Masaaba/Lugisu (a Bantu language of Uganda/Kenya), not Maasai (a Nilotic language), per WAXAL's own documentation — so there is no overlap or naming collision with this dataset despite the shared 3-letter code in casual use.

Note: this dataset is gated on HuggingFace — an account must click through the access request on the dataset page before the API can download it.

Structure and what "transcribed" means here

The source splits data into train/dev/dev_test; only train is used here (dev/dev_test are held-out evaluation partitions, same policy we apply to DSN's splits for the Nigerian-language releases).

Within train, two categories are pooled together:

• Scripted — read from a prepared script, 100% transcribed by construction. Each row also ships an English translatedText alongside the native transcript (not included in this release's text field, which is native-language only).
• Unscripted — topic-prompted natural speech. The large majority carries a real, reviewed transcript (Anv-ke's own workflow marks each as approved/rejected after review); only transcribed rows are included here.

A text-encoding bug in the source, fixed during ingestion: unscripted transcripts in the raw parquet files are mojibake — UTF-8 bytes that got decoded as Latin-1 somewhere upstream. This is fixed via a encode('latin-1').decode('utf-8') round-trip that's a safe no-op on already-correct text (a genuine non-Latin-1 character can't itself be Latin-1-encoded, so the fix only fires on rows that actually need it) — scripted transcripts, which are correct as shipped, pass through unchanged.

All audio is standardized to 16 kHz mono FLAC (lossless), 1–30 second clips. Source audio is real WAV, embedded directly in the source's parquet files (no WebM-mislabeling issue like some other sources in this collection).

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):

Column	Description
key, shard	which tar file + entry holds this clip's audio
text	transcript (native Maasai script)
duration	seconds
source	always anv_ke
dataset_id	always 0
split	train / val (250 clips held out for evaluation)
speaker_id	Anv-ke's recorder_uuid
gender	speaker metadata, joined from the source's meta.csv
domain	e.g. scripted/Agriculture and Food, unscripted/Education and Technology
dbfs, clip_ratio, sil_ratio	cheap DSP quality proxies: loudness, fraction of clipped samples, fraction of near-silent frames
has_disfluency	always false — this source doesn't flag disfluencies

Usage

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

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

row = df.iloc[0]
shard_path = hf_hub_download("Professor/maasai-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 Maasai TTS/ASR research, in particular as finetuning data for a multilingual TTS model that doesn't natively support Maasai. Source recordings cover 2 dialects pooled together (per the upstream dataset card); dialect is not preserved as a separate field in this release. This is a research aggregation; usage should respect African Next Voices' own terms.

License

CC BY 4.0, per the upstream Anv-ke/Maasai release.

Acknowledgments

Deep thanks to the KenCorpus Consortium / African Next Voices for the source corpus, and the Gates Foundation for funding its collection.

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