Professor/malagasy-speech-data

Malagasy Speech Data (Pooled)

A ~179.8-hour Malagasy speech corpus, drawn from a single source (Afrivoice) and filtered to only genuinely transcribed audio. Part of the AfroNet multi-language TTS data effort.

Source

DigitalUmuganda/Afrivoice (the general, pan-African Afrivoice release — not Afrivoice_Ethiopia, which we've separately ingested for 5 Ethiopian languages) — Malagasy portion: 34,293 clips, 179.8h, source dataset_id/source = afrivoice.

Why not WAXAL's mlg_asr config? WAXAL (google/WaxalNLP) does have a Malagasy ASR config, but it's the same underlying data as this source: WAXAL's mlg_asr speaker IDs are 28-character Firebase-Auth-UID-style strings that directly match Afrivoice's speaker_id format, its own source_datasets metadata references DigitalUmuganda/AfriVoice explicitly, and its transcripts follow the same image-description methodology. Rather than pool both (which would just mean training on the same recordings twice), we use Afrivoice directly as Malagasy's sole source.

A note on what "transcribed" means here

Afrivoice pairs each audio clip with an image the speaker was prompted to describe; transcription is the sentence the speaker was recorded saying. Only a portion of recordings are transcribed (~35% of total duration for Malagasy — 179.8h out of ~516h total) — the rest was recorded but never transcribed. No auto-transcription was used to unlock the untranscribed majority; only clips with a real, human-provided transcript are included here.

All audio is standardized to 16 kHz mono FLAC (lossless), 1–30 second clips. Source audio is real WAV — no WebM-mislabeling bug here, and it decodes directly via soundfile with no ffmpeg step 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 script), from Afrivoice's transcription field
duration	seconds
source	always afrivoice
dataset_id	always 0
split	train / val (250 clips held out for evaluation)
gender	speaker metadata where available
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/malagasy-speech-data", "manifest.parquet", repo_type="dataset")
df = pd.read_parquet(mp)

row = df.iloc[0]
shard_path = hf_hub_download("Professor/malagasy-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 Malagasy TTS/ASR research, in particular as finetuning data for a multilingual TTS model that doesn't natively support Malagasy. Speech is prompted by an image-description task, a narrower register than natural conversation. This is a research aggregation; usage should respect Afrivoice's own terms.

License

CC BY 4.0, per the upstream Afrivoice release.

Acknowledgments

Deep thanks to Digital Umuganda for the Afrivoice corpus.

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