CMU Researchers Release XEUS: A Cross-lingual Speech Encoder Supporting 4000+ Languages, Trained on 1M+ Hours

Jul 4, 2024, 05:26 PM

Researchers from Carnegie Mellon University's WAVLab have announced the release of XEUS, a state-of-the-art cross-lingual speech encoder. XEUS, which is trained on over 1 million hours of speech data, supports more than 4000 languages. The model outperforms existing systems such as MMS 1B and w2v-BERT v2 2.0 on various tasks. It features 577 million parameters and employs a self-supervised learning objective using the E-Branchformer. XEUS also includes a new dereverberation technique. The code, checkpoints, and data for XEUS are being released under the MIT License.