MiMo-Audio
MiMo-Audio is an audio language model trained on more than one hundred million hours of audio that shows few-shot generalization across audio tasks it was not explicitly trained for, such as voice conversion, style transfer and speech editing.
MiMo-Audio is an audio language model built to generalize to new audio tasks from only a few examples or simple instructions, rather than needing task-specific fine-tuning the way most existing audio language models do. Scaling its pretraining data to more than one hundred million hours produced few-shot learning capability across a diverse set of audio tasks.
Its tokenizer is a 1.2-billion-parameter transformer running at 25 Hz that uses an eight-layer residual vector-quantization stack to produce 200 tokens per second, jointly optimized for semantic and reconstruction quality. A patch encoder and decoder downsample this for the language model, and the base model generalizes to tasks such as voice conversion, style transfer and speech editing that were absent from its training data, while an instruction-tuned version adds thinking mechanisms for both understanding and generation.