v5.1

[snakers4]

Experimental PIP package release

• Experimental pip-package release;
• Community PRs to update the examples;

What's Changed

• Adamnsandle by @adamnsandle in Adamnsandle #481
• Update microphone_and_webRTC_integration.py by @eltociear in Update microphone_and_webRTC_integration.py #475
• cpp example by @filtercodes in cpp example #482
• Update Golang example to support model v5 by @streamer45 in Update Golang example to support model v5 #489
• Create python-publish.yml by @adamnsandle in Create python-publish.yml #492
• Adamnsandle by @adamnsandle in Adamnsandle #493

New Contributors

• @eltociear made their first contribution in Update microphone_and_webRTC_integration.py #475
• @filtercodes made their first contribution in cpp example #482

Full Changelog: v5.0...v5.1

---

This discussion was created from the release v5.1.

[lifeiteng]

@snakers4 Thanks for your work!

How do I knowing if I'm using the latest version?

model, utils = torch.hub.load(repo_or_dir='snakers4/silero-vad', model='silero_vad')

[snakers4]

Torch hub load without tags or branches just load the latest state of the repo

[ozancaglayan]

Thanks!

Just as a side info: I now tried this version on a sample of private audio files and compared v4 and v5. It seems that v5 rejected around 25% more content given the same settings. Upon listening to some of those rejected regions, I hear proper spoken content (sometimes unintelligible sometimes intelligible but children voices) It would be good if you can give us some high-level description of what kind of data the model was trained on and how did it change from v4

[snakers4]

It is always worth doing the following:

• Isolating the problematic audio;
• Plotting the probability chart (there is param in the get_speech_timestamps function);
• Seeing if the problem is systemic, or just improper hyper-paramers;

It would be good if you can give us some high-level description of what kind of data the model was trained on and how did it change from v4

We added as diverse data as possible for 6k languages, and put major focus on productions data, noises, edge cases.
We also tested the VAD on many more validation sets.

I hear proper spoken content (sometimes unintelligible sometimes intelligible but children voices)

Can you share some audios?

[ozancaglayan]

Yes, I'll do the analytics at some point, thanks!

Unfortunately, I can't share these audio files but I'm pretty sure I/we can find some examples with children speaking.

[leminhnguyen]

@snakers4 I have the same problem, how can I rollback to V4 version?

[ozancaglayan]

Hi. I can't share some audios but can share plots of two audio files, under same settings with v4 and v5.1. The plot is the probability difference e.g. v5_probs - v4_probs. Positive values show the regions where v5 assigns more voice probability than v4 and vice versa. Both models are run with window_size_samples=512, and under ONNX.

First one is especially interesting. But both plots show that the issue is not a simple threshold shift happening across the models. I wonder whether this is all because of the removed Adaptive Audio Normalization layer?

[TYMA22389]

Субтитры сделал DimaTorzok