MiLorE-SSL: Scaling Multilingual Capabilities in Self-Supervised Models without Forgetting

TL;DR

MiLorE-SSL is a lightweight continual learning framework for multilingual speech models that efficiently incorporates new languages without forgetting existing ones, using LoRA modules, soft MoE, and limited replay data.

Contribution

It introduces a novel combination of LoRA, soft MoE, and limited replay data for scalable multilingual SSL without catastrophic forgetting.

Findings

Achieves strong performance in new languages

Improves existing language capabilities

Uses only 2.14% trainable parameters

Abstract

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.