Add realtime transcription using vllm /v1/realtime websocket connection

[ad-astra-video]

Is your feature request related to a problem?

Over the last months we've been improving the gateway experience by disaggregating the go-livepeer gateway, which is heavy and forces operators to run media processing and payment signing in one node (see transformation SPE retrospective).

• We shipped a standalone remote signer that holds the Ethereum key and signs PM tickets out of the media path, plus local SDKs (e.g. livepeer-python-gateway) so a gateway becomes an embeddable library — point it at a signer URL + key, media stays local. The gateway ↔ orchestrator payment flow is unchanged (PM, settling in ETH).
• Building on that, Josh built a new live runner that allows full container passthrough over HTTPS, SSE, WebSockets, and Trickle. Builders ship a standard container, orchestrators deploy it with a single docker pull + docker run, and it becomes discoverable on the network. The goal: a builder takes the same container they already run on their existing cloud infra and runs it on Livepeer instead — making the network a direct, decentralized replacement for that infra, with payments handled by the protocol. Relevant branches: go-livepeer#3938 (Live Runner) and livepeer-python-gateway#20 (Live Runner + Scope support).

We created livepeer/live-runner-example-apps to showcase one example per communication schema (some are not yet merged and will ship later this week).

To get people trying this out and to gather first-hand feedback, we'd like a builder to add realtime transcription using (vllm realtime) as another example using the Trickle and/or websocket protocol — and to report on the developer experience and any problems encountered.

Describe the solution you'd like

A realtime transcription runner in a standalone vllm-realtime/ example folder that runs realtime transcription models supported by vllm and use the realtime transcription result to derive something from the audio (e.g. count words, sentiment, etc).

Implementation should provide:

• ingest audio with websocket or trickle
• websocket for output
• Both offchain (free) and on-chain (paid, via the remote signer).

See the streamdiffusion for example of trickle input (GPU, realtime trickle, static registration, self-hosted weights, offchain + on-chain)

At minimum:

• runner.py — the app serving the vllm realtime engine via the SDK.
• client.py — publishes transcription result and allows for settings adjustments live where applicable (--signer for on-chain).
• Dockerfile + docker-compose.yml (offchain) + docker-compose.onchain.yml (paid) — bring up the orchestrator + app with one command each.
• README.md — how to run it offchain and on-chain.

Also include written feedback on the developer experience and any issues hit along the way.

Describe alternatives you've considered

None

Additional context

• https://vllm.ai/blog/2026-01-31-streaming-realtime
• https://docs.vllm.ai/en/latest/examples/speech_to_text/realtime/
• Example apps repo: https://github.com/livepeer/live-runner-example-apps
• go-livepeer Live Runner branch: Live Runner go-livepeer#3938
• Local SDK branch: Live Runner + Scope support livepeer-python-gateway#20

Retroactive Grant Notice

Important

Compensation cap: $500

Before picking this up, read the Retroactive Grants post in full. Key points:

• Not pay-for-delivery. Compensation is based on demonstrated adoption and community impact — not on shipping. Real, named community contributors must use your work and publicly attest to it. Building in isolation, then asking for a grant, does not qualify.
• Talk to the community first. A public discussion (Discord or Forum) about the problem must precede the work, and evidence of it must be part of your application. No prior discussion, no tempcheck, no public signal = won't pass review.

Required steps before investing more time:

1. Read the original Retroactive Grants post in full.
2. Start a public discussion (Discord or Forum). Get input. Confirm direction.
3. Secure at least one community proof point — a Discord thread with positive signal, a Review Team acknowledgement, or evidence of active adoption.
4. Identify, by name, who has this problem.
5. Build, ship, and confirm it's actually being used.
6. Apply with at least one named community contributor willing to publicly attest to your work's impact on the forum post.

Skipping steps 2, 3, 4, or 6 will halt the application. Reach out to jj.a.ss.o.nn_everestnode on Discord before you start building.

[linear-code]

ENG-126