Duplex
Compute
Duplex Compute is building affordable voice infrastructure for AI that listens, speaks, and acts.
We’re working toward a full-duplex API with tool calling, configurable behavior, and room for longer conversations. Built for developers bringing voice to their applications.
Request early accessGood conversation.
Smaller compute bill.
Voice interfaces are becoming capable enough to teach, help, and get work done. The cost of keeping a conversation open still limits what people can build.
We’re exploring open models and efficient serving to bring that cost down. Our focus is the complete interaction: interruptions, useful tool calls, and conversations that can keep going.
Listen while speaking
Handle a correction without making someone wait for the end of a sentence.
Act on what’s said
Connect conversation to your application’s tools, knowledge, and reasoning.
Measure the whole system
Compare task success, timing, and cost—not just how a voice sounds.
Small systems.
Serious possibilities.
Open voice systems are approaching commercial APIs on some interaction benchmarks. The next question is how reliably—and at what serving cost.
Capability and cost, side by side
COMPARISONTwo measures of performance. One hourly price. No hidden third axis.
| System | Task success ↑FDB-v3 Pass@1 · 0–100% | Reply quality ↑Transcript-judged · 0–100% | Voice cost / hour ↓USD · separate backend costs excluded |
|---|---|---|---|
| Duplex ComputeOur API | 57% | 69% | $0.60Per hour · includes tool calling |
| GPT-Live-1OpenAI · hosted API | 55% | 71.7% | $3.00Published session rate |
| Gemini Live 3.1Google · benchmark version | 54% | 71.8% | Not comparedNo matched hourly quote here |
| TML-smallTML-Interaction-Small | 68% | ≈82.8% | Not comparedNo matched hourly quote here |
| NemotronLabs 11BNVIDIA · open model | 33% | Not reported | Self-hostedDepends on GPU and utilization |
| .wave NemotronHosted beta · 2-minute sessions | Tools not supported | Not reported | $0.30Introductory rate, hourly equivalent |
Methodology and more systems
Performance bars share a 0–100% scale; cost bars share a $0–$3/hour scale. Pass@1 measures strict first-attempt task success, including tool selection and arguments. Reply quality measures task fulfillment judged from the spoken transcript, not sound quality. All systems were evaluated on the same FDB-v3 benchmark run. Small differences do not establish superiority; confidence intervals are omitted here for readability.
The .wave API does not expose the underlying model’s tools, so we do not attach NemotronLabs’ tool-use score to that hosted service. Gemini’s benchmark version is 3.1; usage measured on a different Gemini version is not a matched price for this row. “Not compared” does not mean free.
| System | Task success | Reply quality |
|---|---|---|
| NVIDIA front/back, 235B | 48% | ≈67% |
| Cascaded GPT-4o | 45% | 60% |
| NVIDIA front/back, 30B | 44% | ≈54% |
| Ultravox v0.7 | 41% | 51% |
| Gander 9B | 40% | ≈49% |
The NVIDIA 30B and 235B entries use external ASR and Qwen3 backends. Cascaded GPT-4o combines Whisper, GPT-4o, and TTS. Read the research note →
More conversation.
Less overhead.
Our target is $0.01 per connected minute for voice with tool calling. We’re validating the quality and serving economics before offering an API.
Tell us what you need80% lower voice-session cost than GPT-Live-1. OpenAI rate: $0.05/min. Separate backend reasoning, tools, and telephony excluded. Source ↗
What would
you build?
A tutor that has time to listen. An assistant that can act. A character that doesn’t wait for its cue.
Tell us about your application. We’ll get in touch as we open access to the API.
Pre-launch · No payment requiredWorking on voice models
or inference?
We’d like to compare notes. Researchers, model builders, and potential design partners can reach us through the interest form. We’re an early-stage project, not currently advertising open roles.