multilingual-asr

CoVoST 2 and Common Voice 17.0 Swahili and Hausa, re-packaged under one feature schema so the configs can be concatenated into a single multi-task training mix. Four configs, ~206 hours of distinct audio, 8.44 GB of Parquet. No Bambara.

Load

from datasets import load_dataset

asr = load_dataset("djelia/multilingual-asr", "covost2-transcription", split="train")
sw_test = load_dataset("djelia/multilingual-asr", "swahili", split="test")
print(asr[0]["language"], asr[0]["text"])

Stream rather than downloading 8.44 GB:

ds = load_dataset("djelia/multilingual-asr", "swahili", split="train", streaming=True)
for row in ds.take(5):
    print(row["duration"], row["text"])

Configs

Config	Splits	Rows	Audio	Task
covost2-transcription	train	86,062	16 kHz	transcribe X → X, 21 languages
translation	train	86,062	16 kHz	translate X → English
swahili	train, test	46,494 / 12,253	48 kHz	transcribe sw → sw
hausa	train, test	1,925 / 661	48 kHz	transcribe ha → ha

Fields

Field	Description
audio	16 kHz in the CoVoST 2 configs, 48 kHz in swahili and hausa
text	Transcript, or English translation in translation
duration	Seconds
language	Source language of the speech
task_type	"transcription" or "translation"; absent from swahili and hausa
source_dataset	Upstream corpus

Notes

covost2-transcription and translation are the same 86,062 clips with different targets — load one, or treat them as two tasks over shared audio, but do not concatenate them as separate corpora.

Sampling rates and column sets differ between the CoVoST 2 and Common Voice configs, so resample and align columns before mixing them in one dataloader.

covost2-transcription has no held-out split; swahili and hausa each carry a test split.