Inference possibilities

[amil-rp-work]

Hey,
Thanks for providing such an amazing easy-to-use webapp code.

I was wondering if this pre-trained model could be used for any random non-youtube video file(audio or video)?

[jramcast]

The models do not process video, only audio.

You can use any 16 bit PCM wav file, which is then preprocessed by the vggish network here (this is a prerequisite for models trained on Audioset):

https://github.com/jramcast/mgr-service/blob/master/mgr/infrastructure/audioset/vggish/loader.py#L46
https://github.com/jramcast/mgr-service/blob/master/mgr/infrastructure/audioset/vggish/model/vggish_input.py#L75

If you need to use other audio formats, e.g. mp3, you need to covert the audio to wav before feeding it into the model. I used ffmpeg here to extract the wav audio from youtube videos, but you can also use it to convert mp3 to wav.