Is this similar to Alibaba's Fun-ASR?

[satya941]

How is the model here different from alibaba's fun ASR? https://www.alibabacloud.com/blog/alibaba-introduces-fun-asr1-5-advancing-multi-language-speech-recognition_603138

[LauraGPT]

They are from the same broader Tongyi Lab / Fun-ASR model family, but they are not the same published artifact or version number.

This repository currently publishes two downloadable 800M open checkpoints for local deployment:

• Fun-ASR-Nano-2512: Chinese, English, Japanese, plus Chinese dialects and accents; trained on tens of millions of hours.
• Fun-ASR-MLT-Nano-2512: a separate 31-language checkpoint trained on hundreds of thousands of hours.

Both can run locally through FunASR, vLLM, or the llama.cpp/GGUF runtime documented in this repository. The repository source is Apache-2.0, and each model card states the weight license.

The Alibaba Cloud article describes Fun-ASR 1.5 as a 30-language MoE service with automatic code-switching and Chinese text formatting. Its published access paths are the Alibaba Cloud Model Studio API and a ModelScope demo. The article does not identify “1.5” as either Nano-2512 checkpoint or announce a downloadable Fun-ASR-1.5 checkpoint.

So the practical choice is:

• need downloadable weights, offline/self-hosted inference, or customization: use the Nano checkpoint matching your language coverage from this repo;
• need the exact Fun-ASR 1.5 cloud model and its MoE/code-switching service behavior: use the Model Studio API/demo.

The “1.5” label should not be interpreted as this GitHub repository’s package or release version.