duplex.gg
Dashboard
Realtime model playground

Compare PersonaPlex, OpenAI, Google, and other SOTA voice candidates.

Pick a provider, choose a role, and preview how each candidate would handle full-duplex receptionist routing before wiring it into Discord, agent workspaces, or customer workflows.

Realtime provider

Select the candidate stack to benchmark.

SOTA bench
Live room preview

OpenAI Realtime

Wise Teacher on gpt-realtime
Voice marin / cedar / alloy · ready
idle
OpenAI Realtime is the cleanest high-quality demo path: browser mic to ephemeral WebRTC session, then route to agents.
You are a wise and friendly teacher. Answer questions or provide advice in a clear and engaging way.
sub-second target
target latency
premium managed API
cost model
WebRTC ephemeral session
deployment
Open real OpenAI WebRTC demo See routing use case

Agent preset

persona script

Candidate notes

Strengths
Strong reasoning
official Realtime SDK
server VAD and interruption
browser WebRTC path already wired
Caveats
Provider cost
managed API dependency
requires API key and ephemeral session broker

PersonaPlex voice library

Used when benchmarking the self-hosted PersonaPlex/Moshi-style path. Managed providers use their own voice catalogs.

Deploy next

Save a provider candidate, attach it to a receptionist route, connect Discord, then benchmark it against the same receptionist handoff script.

Create agent

Benchmark rubric

Interruption / barge-in

Can users cut off the agent and redirect intent naturally?

Routing accuracy

Does the provider preserve intent when switching from Receptionist to specialist?

Tool-call reliability

Can the voice session safely trigger workflow tools behind approvals?

Transcript quality

Can summaries survive noisy Discord rooms and multi-speaker contexts?

Cost per useful minute

Not raw audio cost — cost after retries, latency, and failed sessions.

Deployment control

Can the customer run it self-hosted, managed, or hybrid by tier?