For something as compressible as voice, I do not know how you would feel confident that data was not slipping through. Edge transcription models (eg Whisper) are continuing to get better, so it would be possible for malware to send a single bit if a user says a trigger word.