holle

[6ccccc]

How can I get in touch with you? I’d like to talk with you about developing Windows-driven lip movement for digital humans.

[6ccccc]

Looking forward to your reply.

[arghyasur1991]

can you pls explain here what is it you are looking for?

[6ccccc]

First of all, thank you for your sharing. I want to build a virtual human (video digital human) dialogue system in Unity to create a virtual digital human. I’m using the Windows system with Unity version 2022.3LTS. I tried to load the ONNX model required for MuseTalk with Microsoft.ML.OnnxRuntime, and the ONNX download link is the Google link you shared, but the initialization never succeeds. I wonder if your solution is specifically developed for Mac? Do you have any suggestions for Windows? Or could it be an issue with my Unity version? I initially thought it was a graphics card problem and have replaced the graphics card. Currently, we do not support Unity 6. Thanks for your reply.

…

---Original--- From: "Arghya ***@***.***> Date: Sun, Mar 29, 2026 02:55 AM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [arghyasur1991/LiveTalk-Unity] holle (Issue #10) arghyasur1991 left a comment (arghyasur1991/LiveTalk-Unity#10) can you pls explain here what is it you are looking for? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[arghyasur1991]

Hi, this package is not currently tested on windows (as mentioned in Readme requirement section) so its possible that there is some issue in the model loading on windows.
one hack you can try to troubleshoot -
in https://github.com/arghyasur1991/LiveTalk-Unity/blob/main/Runtime/Core/Model.cs, line 73, you can try adding
preferredExecutionProvider = ExecutionProvider.CUDA // or ExecutionProvider.CPU hardcoded. I believe it always currently tries to load on CoreML and then fails.

You are also welcome to fix the CUDA path on the lines of
https://github.com/arghyasur1991/Spark-TTS-Unity/pull/6/changes

[6ccccc]

Actually, I don't need TTS because I've already completed the TTS part. I just want to get the Windows-specific MuseTalk.onnx file, then integrate it by referring to your solution. I noticed that the version exported in your solution is optimized for Mac.

[6ccccc]

What should I do to run it properly on the Windows platform? By the way, does this support real-time lip-sync driving, right?

[arghyasur1991]

does this support real-time lip-sync driving

Yes

What should I do to run it properly on the Windows platform?

As I mentioned, you can try setting correct execution provider (line mentioned in above comment) to either CUDA or CPU.. in my package, its hardcoded to CoreML which is likely the issue you are not able to initialize. The model is optimized for mac, yes but it should at least run fine on CPU and CUDA in windows as well with correct initialization steps.. The PR I shared is for TTS but it has patterns of enabling cuda execution provider so maybe you can get some reference there.

[6ccccc]

I'll give it a try. Are there any other areas that need modification?

[6ccccc]

[LiveTalk-ERROR] [ModelUtils] Failed to initialize ONNX Runtime logging: The type initializer for 'Microsoft.ML.OnnxRuntime.NativeMethods' threw an exception.

[ORTModel] Failed to initialize ONNX Runtime logging: The type initializer for 'Microsoft.ML.OnnxRuntime.NativeMethods' threw an exception.

[6ccccc]

Haha, I've been working on this for days, and I'm kinda falling apart.

[6ccccc]

Does this exported file not support the Windows platform? I used a tool to overwrite it, and the file size is still the same.

[6ccccc]

您那边这套系统延迟大吗whisper_encoder_dynamic.onnx Does audio feature extraction take about 5 seconds?