MiMo-Audio: Few-Shot Audio Learning with Large Language Models

This paper introduces MiMo-Audio, a large-scale audio language model demonstrating few-shot learning capabilities. It addresses the limitations of task-specific fine-tuning in existing audio models by leveraging the scaling paradigm seen in text-based language models like GPT-3. The paper highlights the model's strong performance on various benchmarks and its ability to generalize to unseen tasks, showcasing the potential of large-scale pretraining in the audio domain. The availability of model checkpoints and evaluation suite is a significant contribution.

• •MiMo-Audio is a large-scale audio language model.
• •It demonstrates few-shot learning capabilities.
• •Achieves SOTA performance on various benchmarks.
• •Generalizes to unseen audio tasks.
• •Model checkpoints and evaluation suite are publicly available.