Please provide a method to NOT download the models automatically.

[Nampo]

Checks

• This template is only for feature request.
• I have thoroughly reviewed the project documentation but couldn't find any relevant information that meets my needs.
• I have searched for existing issues, including closed ones, and found not discussion yet.
• I am using English to submit this issue to facilitate community communication.

1. Is this request related to a challenge you're experiencing? Tell us your story.

We can't keep the installation in a single place or make it portable on windows, and there is no need to have yet another copy of Whisper in the C: drive.
Let the users place the models or symlink to them where they want.
Thanks.

2. What is your suggested solution?

Just a config with the path would be enough.

3. Additional context or comments

Auto-download should always be optional, especially if there is NO WAY to choose the path of the models.
Imagine if every one who release a new model do the same, we end up with unmanageable mess and models duplication on multiple drives.

4. Can you help us with this feature?

• I am interested in contributing to this feature.

[zhu-han]

If you want to use a local model, just change the model name to your local model path. For example, if you are using python API, change "k2-fsa/OmniVoice" to your local path.

model = OmniVoice.from_pretrained(
    "k2-fsa/OmniVoice",
    device_map="cuda:0",
    dtype=torch.float16
)