A new arXiv study finds that pretrained self-supervised speech models like Wav2Vec2 and HuBERT, despite training data skewed toward high-resource languages, can recognize click consonants from Khoisan languages…

#AI #SpeechRecognition #NLP #LowResourceLanguages
https://arxiv.org/abs/2606.11542

New paper introduces Model as a Library, an offline architecture that packages community-enrolled speech models as on-device dependencies for African languages, avoiding generative hallucination by drawing vocab from…

#AI #NLP #LowResourceLanguages #EdgeAI
https://arxiv.org/abs/2609.38574