Realtime voice infrastructure — 20ms to first byte, 200+ languages, billed by usage
- Latency engineering is the right obsession for voice, and the stated figures are strong
- Model routing avoids architectural commitment to one provider
- Offering rival STT models alongside its own signals genuine neutrality
- It is components, not a product — you need engineers to build anything usable
- The pivot away from game character tooling strands developers who arrived for that