remove 'providers' option for melspec and embedding InferenceSession

[fquirin]

Ty for this very interesting open-source wake-word project! I've done some experiments for SEPIA open assistant framework and it looks like it could finally become a real alternative to Porcupine 🙂 (with the exception of custom ww creation for now ^^).

Just a small comment. I'm getting warnings because of 'CUDAExecutionProvider' ('CUDAExecutionProvider' is not in available provider names).
I think you can simply remove the providers option in the utils class because ONNX runtime will set this to the available providers automatically.

Cu,
Florian

[dscripka]

Thanks, I'm glad you've found the library useful! That's an excellent point about the ONNX providers, and at minimum I can set the default to be the CPU provider so most people are less likely to get this warning. I'll plan to make this change in an upcoming release.

As for custom models, you can already train custom wakeword models now (see an example notebook here), though I recognize that the process is a bit complex. I'm hoping to make the custom model training process much simpler in the future.

[fquirin]

Looking forward to it 👍.

I spent some time to try and find out if I could transport the runtime over to ONNX web, but stopped for now when I realized there are actually 3 ONNX models involved ^^. I'll keep this on my to-do list though since my SEPIA clients are mostly running as web-apps (even the headless one).
Do you have any plans to work on that as well? :-)

[dscripka]

This issue will be fixed when #31 is merged, so closing.

And yes, there are three separate models, which I admit is a bit unwieldy when porting to other deployment environments. I have considered making a javascript library before, but it would take some time as much of the functionality in the Python library would need to be ported as well.

Feel free to open a Discussion topic for this though, happy to continue the conversation there.