Professor/sidama-speech-data

Sidama Speech Data (Pooled)

A ~188.2-hour Sidama speech corpus, drawn from a single source (Afrivoice Ethiopia) and filtered to only genuinely transcribed audio. Part of the AfroNet multi-language TTS data effort — sibling release to Yoruba/Hausa/Igbo/Kinyarwanda/Swahili, but Sidama (and its four sibling Ethiopian-language releases, Amharic/Oromo/Tigrinya/Wolaytta) are each published independently, not bundled into one combined "Ethiopia" dataset, even though they share a single upstream source.

Source

Afrivoice Ethiopia (Digital Umuganda) — 38,303 clips, 188.2h, source dataset_id/source = afrivoice.

Why not also use WAXAL's sid_asr config? Strong circumstantial evidence points to WAXAL's Ethiopian ASR configs (amh_asr/orm_asr/sid_asr/tir_asr/wal_asr) being derived from the same underlying Digital Umuganda collection, not an independent source: matching Firebase-style speaker-ID format between the two, matching "describe this image" content register, and an exact match on the set of five languages covered. Pooling both would very plausibly mean training on the same speakers/clips twice, not doubling real coverage — see the project README for the full reasoning.

A note on what "transcribed" means here

Afrivoice Ethiopia collects audio across three registers — Scripted (read from a pre-written script, 100% transcribed by construction), and two Unscripted categories, Expert and general (further broken into "image prompt", "spontaneous", "undefined" sub-categories in the raw data) — where the large majority of recorded audio has no transcript at all. Only a minority of Unscripted/Expert clips were manually transcribed by Digital Umuganda.

This dataset pools only clips with a real, human-provided transcript — no auto-transcription was used to unlock the much larger untranscribed portion (several hundred additional hours exist per language, recorded but untranscribed). That was a deliberate choice: unverified ASR output on audio nobody has checked isn't something we want to train a TTS model on. If that untranscribed audio is auto-transcribed and verified in the future, it could meaningfully scale this release up.

All audio is standardized to 16 kHz mono FLAC (lossless), 1–30 second clips. Source audio ships as .wav-named files in the raw dataset, but is actually WebM/Opus (confirmed via ffprobe) — decoded via ffmpeg during ingestion, not a plain WAV read.

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 script), from Afrivoice's transcription or script field
duration	seconds
source	always afrivoice
dataset_id	always 0
split	train / val (250 clips held out for evaluation)
gender	speaker metadata where available
domain	which register the clip came from: scripted, unscripted/expert, unscripted/image prompt, unscripted/spontaneous, or unscripted/undefined
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/sidama-speech-data", "manifest.parquet", repo_type="dataset")
df = pd.read_parquet(mp)

row = df.iloc[0]
shard_path = hf_hub_download("Professor/sidama-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 Sidama TTS/ASR research, in particular as finetuning data for a multilingual TTS model that doesn't natively support Sidama. Roughly half this corpus is "Scripted" register (closer to natural read speech); the rest is a mix of spontaneous/expert speech describing images, a narrower register than natural conversation. This is a research aggregation; usage should respect Afrivoice Ethiopia's own terms.

License

CC BY 4.0, per the upstream Afrivoice Ethiopia release.

Acknowledgments

Deep thanks to Digital Umuganda for the Afrivoice Ethiopia corpus.

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