Thanks — I hear you: you want transcription to start immediately when a user joins the LiveKit room, not 2–3 seconds later. Below are targeted steps that will help remove that lag and get AWS Transcribe streaming as soon as possible. Which part should we focus on first? (pick one) • Immediate subscription flow (server/agent changes) • Client pre-connect audio buffering • AWS Transcribe / infra tuning • I can walk through all of them if you prefer Recommendations to reduce the 2–3s lag 1) Ensure immediate audio subscription • Have your agent join the room with auto-subscribe enabled so it subscribes to microphone tracks as soon as they are available. • If you handle track events, call subscribe/set_subscribed immediately on the

track_published

event for

MICROPHONE

sources. • This avoids waiting for publish/subscription handshakes that add seconds. 2) Pre-warm agent resources (VAD / models / dependencies) • Load VAD or other audio-processing models at agent process startup (not on first session). Keep them in memory and reuse across sessions. • Reusing the warmed process avoids per-call model initialization time that often causes the first-second delay. 3) Use client-side preconnect audio (buffer before agent connects) • Start microphone capture and buffer audio on the client before or while connecting the agent. • Use the client pattern that captures audio (e.g.,

withPreConnectAudio()

or equivalent) so early user audio is buffered and delivered to the agent as soon as it subscribes. This makes the start of transcription feel instantaneous because early audio is already available. 4) Optimize AWS Transcribe streaming configuration and proximity • Confirm your STT is in streaming mode (not batch). Streaming is required for low-latency transcription. • Run your agent server in the same AWS region as the Transcribe endpoint (reduce network RTT). • If possible, keep a persistent Transcribe session or reuse streaming connections rather than starting a new Transcribe session per call. 5) Consider LiveKit Inference or co-located routing (if available to you) • If you use LiveKit’s inference/co-located routing, it reduces extra hops to AI services and can reduce latency further by keeping audio and STT close together. Quick checklist to implement now 1. Turn on

auto_subscribe

for the agent connection (audio-only or all). 2. Pre-warm VAD and STT client once at process start. 3. Implement client-side preconnect audio buffer so audio is already flowing when the agent arrives. 4. Verify streaming STT config and match AWS region to agent deployment region. 5. Monitor TTFB / first-transcript time and iterate. If you want, I can: • Show example code for immediate subscription (your language: JS, Python, or TypeScript?) • Show the preconnect audio client snippet for your client stack • Help you decide where to deploy the agent to match AWS Transcribe region Which example would help you most right now?