On July 23rd, 2026, Vercel announced that its AI Gateway now supports streaming transcription, a feature that lets developers send live audio to the platform and receive the transcript back as the speaker talks. The capability is exposed through a new SDK method called streamTranscribe, which the company highlighted in a post on its official X account, Vercel Developers @vercel_dev (source tweet).
The announcement links to Vercel’s changelog entry, which details the addition of the streaming transcription endpoint to the AI Gateway and provides a short code example showing how the method can be called from the AI SDK (changelog).
What the feature does
Traditional speech‑to‑text services operate in a request‑response model: a developer uploads a recorded audio file, the service processes it, and returns the full transcript after the file is uploaded and the inference completes. By contrast, Vercel’s streamTranscribe function accepts a continuous audio stream, processes chunks in near real‑time, and pushes the partial transcription back to the caller as each segment is recognized. In practical terms, a developer can embed a microphone feed in a web app, pipe the audio to Vercel’s AI Gateway, and display the live transcript to users with only a few hundred milliseconds of latency.
Why it matters for developers
Vercel’s platform is built around edge‑first deployments. Its serverless Edge Functions run on a globally distributed network, enabling low‑latency responses for front‑end applications. Adding streaming transcription to the AI Gateway aligns the speech‑to‑text workflow with Vercel’s edge strategy: audio can be captured in the browser, sent to the nearest edge location, transcribed, and the result streamed back to the user without round‑tripping to a central data center. This reduces both latency and bandwidth costs for interactive voice experiences such as real‑time captioning, voice‑controlled dashboards, and conversational agents that need immediate feedback.
Developers can now build voice agents that live entirely on Vercel’s stack. By leveraging the same deployment and CI/CD pipelines they already use for static sites and serverless functions, teams can avoid the operational overhead of stitching together separate cloud services for audio ingestion, transcription, and response handling. The SDK’s unified method abstracts away the underlying infrastructure, letting engineers focus on conversation design rather than managing authentication, scaling, or regional routing.
How it fits into the broader AI market
Real‑time speech recognition is a core capability offered by the major cloud providers – Amazon Transcribe Streaming, Google Cloud Speech‑to‑Text, and Azure Speech Services all provide similar low‑latency APIs. Vercel’s entry differentiates itself by being tightly coupled with its edge‑centric developer experience. While the big providers often require developers to manage separate compute resources or configure complex networking to keep the transcription close to the end user, Vercel bundles the service with its existing edge function model, promising a simpler integration path for teams already on the platform.
Potential use cases
Live captioning for webinars and virtual events – Organizers can embed a Vercel‑hosted player that shows near‑instant subtitles without a third‑party service. Voice‑driven analytics dashboards – Business users can ask questions aloud and see real‑time transcriptions that feed directly into data query pipelines. Interactive tutorials and coding assistants – Developers can speak commands that are transcribed on the fly, enabling a more natural coding workflow. Accessibility tools – Real‑time transcription can power assistive technologies for users with hearing impairments, especially when hosted at the edge for minimal delay.
Technical considerations
The streaming transcription endpoint processes audio in small frames, typically 20‑30 ms, and returns incremental text. Vercel’s documentation (linked in the changelog) notes that the method supports common audio codecs such as PCM and Opus, and that developers can control language models via optional parameters. Because the service runs on Vercel’s edge network, developers should be mindful of regional latency variations and may need to handle fallback logic if a particular edge location experiences degradation.
Vercel’s AI roadmap
Vercel introduced its AI Gateway earlier this year as a unified front for accessing large‑language models, embeddings, and other generative AI primitives from the edge. The addition of streaming transcription signals a broader ambition: to make Vercel a one‑stop shop for all AI‑powered interactions that happen on the client side. By extending the gateway to cover audio, Vercel is positioning itself to compete not just on static content delivery but also on interactive, multimodal experiences.
What developers should do next
Developers interested in experimenting can sign up for a free Vercel account, enable the AI Gateway feature, and start calling streamTranscribe from the AI SDK in a sandbox environment. The changelog provides a quick‑start example that demonstrates how to pipe a Web Audio API stream into the function and render the transcript in the UI. Early adopters are encouraged to benchmark latency against existing cloud transcription services to evaluate the edge advantage in their specific use case.
Overall, Vercel’s move to support live audio streaming transcription on its AI Gateway lowers the barrier for building voice‑first applications on the edge, potentially reshaping how front‑end teams incorporate speech interfaces into their products.