A next-generation speech-language model