Evaluate noise suppression models for required ops

[anssiko]

Noise suppression use case was adopted by WebNN API in #61 (thanks @ibelem!):

https://webmachinelearning.github.io/webnn/#usecase-noise-suppression

This issue is to discuss the next steps:

• Evaluate RNNoise model for required ops, to be added to the first-wave models
• Gauge PoC interest, possible collaboration with related groups e.g. WebRTC
• Sharing resources, e.g. RNNoise paper and demo

[kdavis-mozilla]

Hey @jmvalin your work is coming up yet again!

@wchao1115, the WebNN API spec co-editor, was interested in exploring RNNoise a bit further in this context. So, to connect the dots, I thought it apropos to ping you here if any RNNoise questions arise, as, after all, you know more about RNNoise than anyone.

[jmvalin]

@wchao1115 @kdavis-mozilla @anssiko Sure, let me know how I can help.

[wchao1115]

Thanks @anssiko, @kdavis-mozilla, and @jmvalin. In addition to RNNoise I'd also like to propose NSNet, the baseline model for the Interspeech DNS Challenge.

[wchao1115]

@jmvalin. From the implementation standpoint, both RNNoise and NSNet are hybrid DSP/ML based solution. Given that webnn is focusing on the deep learning part, how do you see the solution integrated into the web platform, especially on the signal processing part?

I'm not up to date on the state of Web Audio implementation in the web browsers; do you think the level of support for FFT windowing in the browser is sufficient for RNNoise?

[jmvalin]

I don't actually know the state of DSP in the browser. Being able to compute an FFT would certainly be nice, but it's not a showstopper. Similarly, having fast kernels for convolutions/correlations would be equally useful since RNNoise performs pitch analysis. Honestly what I've like to see at some point is a WebBLAS (plus FFT and convolution/correlation). That would probably cover most use cases -- including a big chunk of WebML.

[wchao1115]

To date, we still need to define split and clamp to satisfy the model's requirements.

[huningxin]

To date, we still need to define split and clamp to satisfy the model's requirements.

+1. And the clamp op would be useful to other first-wave models. For instance, when working on the MobileNet example, I found the TFLite version has the Relu6 op. According to the TensorFlow implementation, it can be lowered to XLA Clamp op.

@wchao1115 , if no duplication, I can work on a PR to add the split and clamp ops. Please let me know.

[wchao1115]

Sounds good. Thanks @huningxin