Does anyone benchmark these models for text-to-speech using traditional word-error-rates? It seems audio-input Gemini is a lot cheaper than Google Speech-to-text.