Support min_snr_in_db == max_snr_in_db in AddBackgroundNoise

[hbredin]

1. torch.distributions.Uniform raises a ValueError when low == high:

   ValueError: Uniform is not defined when low>= high

2. AddBackgroundNoise only raises when min_snr_in_db > max_snr_in_db, and not when min_snr_in_db == max_snr_in_db

There is therefore an inconsistency between the two behaviors.

I think supporting min == max is the way to go.

[iver56]

Thanks for filing this. Yeah, low == high isn't properly supported yet in this transform. I guess it could be patched by using torch.full, like here: https://github.com/asteroid-team/torch-audiomentations/blob/cf33679368125b0ff6db6680a7cf0b289386849f/torch_audiomentations/augmentations/shift.py#L119

As a workaround until it is fixed, one can do something like this, which for all practical intents and purposes gives the same result:

AddBackgroundNoise(
    bg_path, min_snr_in_db=12.0, max_snr_in_db=12.0000001
)

[iver56]

If you'd like to contribute, that would be a welcome PR 😄

[hbredin]

@FrenchKrab I think I noticed this max_snr = min_snr + epsilon hack in your code. Would you like to contribute this change?

[FrenchKrab]

I opened a PR that adds the suggested torch.full fix

[iver56]

Thanks