How We Made a Text-to-Speech Model Respond in Sub-50 ms

Low-latency, realtime multimodal model serving, starting with speech at 50 ms time to first audio.