MiLorE-SSL: Scaling Multilingual Capabilities in Self-Supervised Models without Forgetting
This addresses the computational expense and forgetting issues in scaling multilingual capabilities for speech representation learning, though it is incremental as it builds on existing SSL and adaptation techniques.
The paper tackles the problem of catastrophic forgetting in multilingual self-supervised learning models when adding new languages, proposing MiLorE-SSL, which achieves strong performance in new languages and improves existing ones with only 2.14% trainable parameters.
Self-supervised learning (SSL) has greatly advanced speech representation learning, but multilingual SSL models remain constrained to languages encountered during pretraining. Retraining from scratch to incorporate new languages is computationally expensive, while sequential training without migitation strategies often leads to catastrophic forgetting. To address this, we propose MiLorE-SSL, a lightweight framework that combines LoRA modules with a soft mixture-of-experts (MoE) mechanism for efficient continual multilingual training. LoRA provides efficient low-rank adaptation, while soft MoE promotes flexible expert sharing across languages, reducing cross-lingual interference. To further mitigate forgetting, we introduce limited replay data from existing languages, avoiding reliance on large historical corpora. Experiments on ML-SUPERB demonstrate that MiLorE-SSL achieves strong performance in new languages and improves the ability in existing ones with only 2.14% trainable parameters.