What strategies would you employ to adapt a pre-trained multi-lingual ASR model to a low-resource language for Deepgram-style speech transcription?