Changes to Fast, lower-cost speech synthesis. Model fixed to speech-2.8-turbo. Returns a downloadable audio URL (valid 24h) in data.audio; stream fixed to false.