Will the fine-tuning codes for the VAD and AED models be open-sourced later?

[angenge]

[kaituoxu]

If you need this, we'll consider it.
Thx :)

[eschmidbauer]

🤚 also think this would be really cool

[kyr0]

@kaituoxu That would be amazing. I found FireRedVAD to be the best in all top-tier VAD models, BUT it has one serious issue. It classifies "shhh" sounds from the mouth as voice. This is a serious problem in real-world scenarios and a simple FIR filter or other DSP doesn't help as the frequency range overlaps female human voice frequency range. It needs to be fixed in the model. I implemented all 3 top-tier VAD in my defuss project:

https://github.com/kyr0/defuss/tree/main/packages/vad

Please give me a few days more to have the interactive web demo using them working more stable:

https://github.com/kyr0/defuss/tree/main/examples/with-three.js

It should work so far, but it might give some trouble setting it up. I want to push it online soon, so everybody can enojy comparing the models live - with decent acceleration via ONNX/WebGPU etc. I also made them all work via WASM in Node.js - defuss-vad is isomorphic.

Well.. just wanted to say: Congrats to the great work and model! It's just these "shhh" sounds should not be detected as voice for latin languages. I'm aware that it SHOULD be classified as voice for the very beautiful Mandarin / Kandonese language :)

[kaituoxu]

@kyr0 Thanks for your feedback!