Vercel adds 14 ElevenLabs audio models to AI Gateway with single-key access

Developers using Vercel’s AI Gateway can now build spoken conversations, generate music and create multi-speaker recordings through the same service they use for other AI requests. Vercel and ElevenLabs have partnered to add 14 audio models, putting those tools behind one API key and shared usage, cost and request-tracking views.

Audio joins the shared gateway

The partnership covers six types of audio work: transcription, streaming speech generation, realtime voice agents, music, sound effects and dialogue with multiple voices. Transcription includes both recordings and live audio. Streaming speech lets an app begin reading an answer aloud while a language model is still writing it, rather than waiting for the complete response.

Vercel describes a workflow in which an app transcribes a user’s speech, sends it to a chosen language model, then answers in an ElevenLabs voice. Those requests appear alongside the app’s other gateway traffic. The shared service also provides API-key budgets, routing rules, retries and failover, rather than requiring audio usage to be tracked in a separate reporting system.

Separate voice steps—or one live session

For developers who want to choose each part, Vercel shows Scribe v2 Realtime handling transcription and Eleven Flash v2.5 reading the language model’s reply. Listening, reasoning and speaking remain separate steps. A developer can replace the transcription, language or speech model without changing the rest of that pipeline.

There is an important timing distinction in the example. Scribe transcribes speech as it arrives, but the code waits for the finished transcript before asking the language model for an answer. The answer then streams directly into speech generation. Streaming the spoken reply does not mean every stage of this particular example runs simultaneously.

The alternative is Eleven Realtime v1 Max or Mini. These models handle the spoken conversation through one live WebSocket connection, taking in audio and speaking a reply without separate transcription and speech steps to connect. Vercel positions that route for support calls and in-app assistants where users expect back-and-forth conversation.

• Browser access: The server creates a short-lived token, keeping the API key off the client. Vercel’s React helper handles the connection, microphone capture and playback.
• Turn-taking: Both realtime models support semantic voice activity detection to decide when someone has finished speaking. They do not support the server-vad setting.

Music, cast recordings and launch pricing

Beyond live conversations, Music v2.5 and Music v2 generate tracks from descriptions of genre, mood or instruments, with or without vocals. Developers can request a length or lay out individual song sections with their own styles and durations. Sound Effects v2 handles shorter sounds, including footsteps, slamming doors and looping ambience.

All three use the SDK’s generateSpeech function, but the text describes the desired sound instead of words to read aloud. For scripted dialogue, Eleven v4, v4 Turbo and v3 accept a voice for each line and return one continuous recording. That avoids stitching together separate clips for each speaker; script parts can also stream from a language model.

Vercel is offering all ElevenLabs models on the gateway at 50% off for the first 14 days, through November 9. That is an introductory offer, not a permanent price reduction. Separately, Vercel says AI Gateway reflects provider pricing with no markup and charges no platform fee on inference, including requests made with developers’ own provider keys.