SenseVoice-Small 增量微调:在保留普通话+粤语的基础上增加潮汕话识别,如何避免灾难性遗忘?

Background We use SenseVoice-Small for Mandarin + Cantonese recognition (deployed) online, and output standard Chinese. Now, we want to add Hokkien recognition without losing the existing Mandarin/Cantonese recognition capabilities. The difficulty is that Hokkien is not in the existing languages of SenseVoice, and we cannot get the original Mandarin/Cantonese training data, so we are worried that direct fine-tuning will lead to catastrophic forgetting of Mandarin/Cantonese capabilities.