Some of the best current voice tokenizers achieve ~12 Hz, that's many more tokens than a regular LLM would use for ultimately the same content.