Datasetpublic

vnahata/OmnilingualASR-retrieval

Omnilingual ASR speech-text retrieval (MTEB) Read speech paired with its human transcription, for languages that no existing MTEB audio task covers. Source: facebook/omnilingual-asr-corpus at revision 8648ba8, cc-by-4.0, official test split. Recordings are re-encoded from FLAC to Opus at 16 kHz. Repeated transcripts are dropped, since one would otherwise be relevant to several recordings while only one is marked correct. Built by… See the full description on the dataset page: https://huggingface.co/datasets/vnahata/OmnilingualASR-retrieval.

sourceHugging Facecc-by-4.0updated 21d agoView on Hugging Face

0likes1.1kdownloads

Dataset Card

Omnilingual ASR speech-text retrieval (MTEB)

Read speech paired with its human transcription, for languages that no existing MTEB audio task covers.

Source: facebook/omnilingual-asr-corpus at revision 8648ba8, cc-by-4.0, official test split. Recordings are re-encoded from FLAC to Opus at 16 kHz. Repeated transcripts are dropped, since one would otherwise be relevant to several recordings while only one is marked correct.

Built by scripts/data/omnilingual_asr_retrieval/create_data.py in the MTEB repo.