Fast ASR for Voice Agents: Bring Your Own Turn Detection

Short summary

Teams that have already tuned their own voice-activity detection can bypass a streaming STT provider's built-in turn detection and use sync HTTP ASR instead. Sync transcription returns a finished clip's text in ~134ms p50 at $0.45/hr, eliminating the latency the STT layer would spend re-deciding turn boundaries. The key architectural insight is that turn detection is a decision about ownership, not speed — and keeping HTTP connections warm preserves the latency budget.

• •Turn detection is an architectural decision: own it or delegate it to the STT provider
• •Sync ASR returns transcripts in ~134ms p50 with no end-of-turn deliberation, ideal when your VAD already fired
• •Keep HTTP connections warm between turns to avoid DNS/TCP/TLS overhead on the critical path

Generated with AI, which can make mistakes.