VAD parameter skipping loud segments and hallucinating during silence

[andremarko]

Hello guys,

In my scenario, I'm running Faster Whisper on a Brazilian Portuguese sports program, with VAD enabled.

With the default VAD parameters, some loud speech segments (especially in moments when the narrator shouts “goal!”) are not being transcribed.

However, when I lower the threshold and increase the negative threshold:

vad_parameters = dict(threshold=0, neg_threshold=0.1)

those segments are transcribed, but I start getting hallucinated text during silence, often repeated words.

Has anyone found an effective balance for this kind of content?