Switching to float32 for tracks by default
as reported on Mattermost by @mfernoux, the float16
precision can cause troubles when calculating/applying the scaling and shifting.
We could switch back to float32
or add a check that we don't divide by 0