The thesis is: Low-latency speech depends on coordinating incremental generation with the serving path, not only on the speech model’s nominal speed. The mechanism is: Delayed Streams Modeling aligns text and audio on an 80 ms clock, letting audio trail incoming text rather than waiting for a complete sentence.
Baseten reports a median 214 ms time-to-first-audio on Coval’s TTS benchmark as of September 2026. It also says audio can begin after the first few tokens and describes kernel fusion and CUDA Graphs to reduce GPU launch overhead. A separate deployment claim is that the Baseten Delivery Network delivers 2–3x faster cold starts by mirroring and caching model weights. The why_now is: Real-time voice systems are sensitive to both first-audio delay and the ability to add capacity during short traffic spikes. These are reported results and explanations in Baseten’s post, not an independent assessment of end-to-end agent latency.
The evidence_boundary is: The latency and scaling results and technical explanation are reported in Baseten’s post; the supplied source does not independently verify them or establish end-to-end agent latency. The first implication is: Teams evaluating voice agents should measure time-to-first-audio and cold-start behavior separately from model quality and total response time. The first watch condition is: Whether independent workload tests reproduce the reported latency and cold-start gains under bursty production traffic. The report identifies an engineering surface worth testing, but does not show that all voice-agent interactions or deployments will achieve the reported figures.