lilianemomeni / kws-net Goto Github PK

Thanks for this amazing work!

May I know what is your GPU environment and how long it takes for training?

The implementation for audio-specific training and audio-visual combined training seems to be missing in this repository. Could you please point me to where I can access the same and reproduce the experiments from the paper.

Cheers!

Thanks for this amazing work!

The data pre-processing stages are a bit confusing for me.
Could you please kindly share the processed features? So I am avoid the results difference from the feature part.

First I have to say this is truly amazing work, I can see use cases beyond what you present.

I know you have a very extensive README, but I was wondering if you could make it a bit simpler to run, specifically, have a Google Colab notebook that:

1. installs the necessary apt and pip packages
2. downloads the pre-trained models
3. allows inputting a video path/url and a list of words to spot
4. runs the demo and spots these words

Such a notebook would be ideal for quick experimentations with your models, and for me, for example, allow testing additional languages in which I'm interested (besides English, French, and German).

Hi, thx for sharing this amazing work!

Actually, after finding that the directory does not contain audio-only KWS pre-trained model or model class, I've been trying to train audio-only KWS model for my own.

I extracted mel-spectrogram, and I'm wondering if you did any kind of normalization, and I would be much appreciated if you share what type of normalization or rough range of input mel-spectrogram data.

Hope you have a great day, bye!