Dataset · disco-eth
WorldSpeech: 65k hours of multilingual speech
disco-eth/WorldSpeech
A multilingual ASR corpus with 65k hours of human transcripts for 127 language variants.
What's inside
WorldSpeech contains over 65,000 hours of speech with human transcripts for 127 language variants. Sources include national parliaments, public broadcasters, public domain audiobooks, and international organizations. Each row features a 24 kHz utterance with its transcript, an aligned ASR transcript, the CER between them, and WADA-SNR. The dataset was accepted at NeurIPS 2026.
What you can do with it
- Train a multilingual ASR model
- Evaluate ASR on low-resource languages and variants
- Fine-tune a TTS or audio classifier
- Tasks
- Automatic speech recognition, Text → speech, Audio classification
- Languages
- Afrikaans, Amharic, Arabic, Azerbaijani, Belarusian, Bangla, Catalan, Central Kurdish and 4 more
- Size
- 10M–100M rows
- Downloads this month
- 34.7K
- Likes
- 77
Models trained on it
- ArdaOkan/parakeet-tdt-0.6b-v3-tr-experimental-loraAutomatic speech recognition 155
- ArdaOkan/parakeet-tdt-0.6b-v3-tr-experimentalAutomatic speech recognition 121
- deepdml/whisper-base-es-mix-normAutomatic speech recognition 107
- deepdml/whisper-tiny-es-mix-normAutomatic speech recognition 83
- realhidden/whistle-finetuned-huAutomatic speech recognition 17
- turkish-sa-asr/kashgari-large-v1.0Automatic speech recognition 16
- duxx/turkish-stt-zipformerAutomatic speech recognition 0
How to load it
from datasets import load_dataset
ds = load_dataset("disco-eth/WorldSpeech")