Datasets:

Professor
/

luganda-speech-data

language:
  - lg
license: cc0-1.0
multilinguality: monolingual
task_categories:
  - text-to-speech
  - automatic-speech-recognition
tags:
  - luganda
  - ganda
  - speech
  - tts
  - asr
  - african-languages
  - common-voice
pretty_name: Luganda Speech Data (Pooled)
size_categories:
  - 10K<n<100K

Luganda Speech Data (Pooled)

A ~111.4-hour Luganda speech corpus, drawn from Mozilla Common Voice and filtered to the train split. Part of the AfroNet multi-language TTS data effort.

Source

Mozilla Common Voice 26.0, Luganda (lg) locale, accessed via the Mozilla Data Collective mirror. 67,269 clips, 111.4h, source dataset_id/source = common_voice.

Common Voice is crowdsourced, scripted read speech: contributors record themselves reading prompted sentences, and other contributors validate the recordings. The full corpus splits into train/dev/test/validated (a superset of all three plus extra validated-but-unsplit clips) -- only train is used here, same held-out-partition policy we apply to DSN African Voices' splits.

Why not WAXAL's lug_tts config too? WAXAL's Luganda TTS config (~39h) is a separate, independently-collected corpus with no evidence of overlap with Common Voice -- unlike the Ethiopia/Shona/Lingala/Fulani cases elsewhere in this collection, there's no matching-hours or matching-speaker-ID signal here. It remains a candidate for a future top-up if more Luganda hours are wanted.

All audio is standardized to 16 kHz mono FLAC (lossless), 1–30 second clips. Source audio is MP3, decoded directly via soundfile (no external decoder needed).

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 Luganda script), from Common Voice's `sentence` field
`duration`	seconds
`source`	always `common_voice`
`dataset_id`	always `0`
`split`	`train` / `val` (250 clips held out for evaluation)
`speaker_id`	Common Voice's `client_id`
`gender`	speaker-reported gender where available (Common Voice's `gender` field, e.g. `male_masculine`/`female_feminine`)
`dbfs`, `clip_ratio`, `sil_ratio`	cheap DSP quality proxies: loudness, fraction of clipped samples, fraction of near-silent frames
`has_disfluency`	always `false`

Usage

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

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

row = df.iloc[0]
shard_path = hf_hub_download("Professor/luganda-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 Luganda TTS/ASR research, in particular as finetuning data for a multilingual TTS model that doesn't natively support Luganda. Common Voice audio is crowdsourced (variable recording equipment/environment across contributors) but validated by community review, so quality is generally clean read speech. This is a research aggregation; usage should respect Common Voice's own terms.

License

CC0 1.0 Universal (public domain dedication), per the upstream Common Voice release.

Acknowledgments

Deep thanks to Mozilla Common Voice and its Luganda-speaking contributors for the source corpus, and to the Mozilla Data Collective for hosting it.

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