AI Gateway now supports streaming transcription

AI Gateway from Vercel now supports streaming transcription. Previously, transcription on the gateway required a complete audio file and returned the full transcript in a single response. The update lets applications send audio as it is captured and receive transcript updates as the model produces them, instead of waiting for an end-to-end batch job.

The product mechanics are straightforward: audio is streamed in during capture, and partial transcript results are streamed out as they become available. That keeps latency low compared with the old complete-file-then-full-transcript path. Streaming transcription is in beta and is aimed at continuous or near-real-time flows rather than offline file jobs.

For engineers and builders, this matters for live captioning and voice input, where waiting on a full file before any text appears breaks the user experience. You can wire capture, gateway, and UI so partial text lands as soon as the model emits it, which better matches how people speak and how UIs should update.

In market terms, this moves AI Gateway from batch-style transcription toward the interactive pattern many voice and multimodal products already expect. Teams that already use the gateway for other AI traffic can keep transcription behind the same surface instead of bolting on a separate streaming pipeline for captions and speech-to-text.

Practical takeaway: if you are building live captions or always-on voice input on Vercel’s AI Gateway, evaluate the beta streaming path against your current complete-file flow and measure end-to-end latency and partial-result quality under real capture conditions. Watch how the beta hardens—especially reliability of partial updates, error handling mid-stream, and any limits on session length or concurrent streams—before you treat it as the default for production voice features.