More aggressive filtering for both voice embeddings as well as text

[Darwinkel]

Currently, we only apply naive filtering of audio samples and dialogue strings. This covers most of the corrupted samples, but not all of them.