Why do you expect someone's voice to give clues as to wether or not s/he's a criminal?

More likely than not, the model would simply learn to detect minority accents and further help institutional discrimination against minorities (which are already significantly more likely to be sentenced to prison, etc).