Dataset · google
Waxal: multilingual African speech corpus
google/WaxalNLP
Audio and text for ASR and TTS across dozens of African languages—a large open speech corpus from Google.
What's inside
Multilingual speech corpus for African languages; metadata specifies Acholi, Akan, Amharic, Baoule, Dagari, and Jula, among others. Contains audio recordings with text transcriptions intended for both ASR and TTS. Released alongside the WAXAL paper to improve accuracy and naturalness of speech recognition and synthesis for these languages.
What you can do with it
- Training ASR models for African languages
- Training TTS systems on low-resource languages
- Evaluating multilingual speech models on rare languages
- Tasks
- Automatic speech recognition, Text → speech
- Languages
- Acoli, Akan, Amharic, bau, dag, dga, Ewe, Akan and 4 more
- Size
- 1M–10M rows
- Downloads this month
- 27.8K
- Likes
- 313
Models trained on it
- badrex/Ethio-ASR-multilingual-600MAutomatic speech recognition 84.5K
- badrex/Ethio-ASR-amharicAutomatic speech recognition 731
- ZFTurbo/facebook_wav2vec2_xls_r_300m_lingala_shona 675
- Lab-et/omnivoice-amharicText → speech 421
- Tree-AI-lab/kouman-tts-bau-waxal-fullText → speech 390
- mradermacher/ndizi-gemma4-e2b-african-asr-merged-GGUFAutomatic speech recognition 241
- Shinzmann/sorottsText → speech 191
- Shinzmann/soro-tts-iboText → speech 186
How to load it
from datasets import load_dataset
ds = load_dataset("google/WaxalNLP")