Sync HTTP ASR Lets Voice Agent Teams Use Their Own Turn Detection at 134ms

Some voice-agent development teams prefer to handle turn detection themselves rather than relying on built-in logic from streaming speech-to-text providers. For these teams, a synchronous HTTP-based ASR approach lets their own voice-activity detection (VAD) signal the end of a user's turn, after which a completed audio clip is sent directly for transcription. The Sync API returns a transcript in approximately 134ms at the 50th percentile for clips up to two minutes, at a cost of $0.45 per hour. This contrasts with streaming STT models, which continuously monitor audio and add their own end-of-turn judgment — around 300ms for some providers — before committing a transcript. Teams targeting a sub-500ms latency window from end-of-speech to the start of agent reasoning can use this architecture to keep the transcription step close to pure processing time.