Robustifying Spoken Language Identification on Noisy Condition Using Curriculum Learning

Curriculum learning shows promise in improving machine learning tasks like spoken language identification (LID) in challenging scenarios. We propose an end-to-end language identification framework, the x-vector Transformer-based End-to-End Language Identification system, leveraging curriculum learning. The x-vector Transformer (x-Trans) combines x-vector and transformer-based language modeling to capture rich speaker embeddings and extract high-level linguistic features from raw audio signals, enabling robust language identification. Evaluation on the MLS14 set of NIST LRE 2017 demonstrates significant improvements, achieving equal error rates of 8.91 %, 4.83%, and 3.89% on 3s, 10s, and 30s test speech, which is respectively 34.96%, 47.50% and 52.03% reduction compared to the referred work's performance using x-vector and BLSTM. To assess noise robustness, the clean LRE data undergoes full corruption with radio channel noises extracted from the RATS corpus at various Signal-to-Noise Ratio levels. The subse-quent phase employs curriculum learning, facilitating progressive learning from simple to complex data, resulting in improved performance in both clean and noisy conditions. Finally, we conduct multilingual Asia dataset experiments, achieving over 97.5 % accuracy for in-domain languages, even under highly noisy conditions.