Text-to-speech model speeds boosted 2-3.4x with new optimizations

Original: this is what real inference optimization looks like. everyone assumes TTS serving is just another LLM workload but the bottlenecks are completely different - delay patterns between text and audio toke

Source: x.com ↗