Baseten launches Voice Agent Stack with sub-second latency at ~$0.01 per audio minute

baseten · x · 2026-11-17

Baseten launched its Voice Agent Stack and a reference agent, Baseten Realtime Voice (both in private preview), for building customizable, production-grade voice agents on Dedicated Inference. The orchestrator and every model run over a tight in-cluster network for low end-to-end latency, with independent scaling and full freedom to swap or post-train any model in the pipeline.

A demo agent pairs Qwen3-ASR for speech recognition, Gemma 4 31B for reasoning, and Qwen3-TTS for speech generation. Realtime Voice achieves sub-second median turn latency at high concurrency and costs about $0.01 per audio minute.

Related event: Baseten launches sub-second voice agent stack at ~$0.01/min(2 posts)→

Original post →

More from coding & agent

coding & agent channel →