How to easily convert Meta's MMS-ASR models to ggml?

[allandclive]

Help needed to convert MMS-ASR models to ggml. It supports more low resource languages. https://github.com/facebookresearch/fairseq/tree/main/examples/mms
https://huggingface.co/facebook/mms-1b-all

[maxng07]

Taking a peek, the MMS-ASR looks very similar to meta Seamlessm4t. The interface binding does expose the ability to run it on CPU. If I look at the script in ASR folder, mms_infer.py calls infer.py in https://github.com/facebookresearch/fairseq/blob/main/examples/speech_recognition/new/infer.py and in line 402, it clearly sets torch to run in CPU if GPU is not found. There are other lines in the script that sets to cpu, which indicate it is possible to run this on CPU without GPU. But iine 418-419, seems to indicate a feature that requires GPU.

The only way to find out is to try this on CPU machine. I'm pretty positive that it should work just like seamlessm4t.