“A custom text-to-speech model capable of sub-100ms audio generation.”