Live voice provider marketplace (preview)
You'll pick the voice behind your agents — by price, latency, and quality, not by vendor lock-in. This preview shows the provider marketplace exactly as customers will see it: compare profiles in the table below and try the text-to-speech playground. The four mocked profiles run as deterministic offline simulations — no API keys, no signup. The realtime call below and the OpenAI/Google options in the playground are real: they send audio or text to our voice service, which relays it to the named provider.
Every voice provider, one comparison table
The same adapter interface sits under every provider, so swapping vendors never touches your agent's core loop. Four mocked profiles are runnable today with zero cost; stage-2 adapters for OpenAI and Google are installed and measured (below), with three more providers declared.
| Provider | Architecture | Quality tier | Latency budget | Cost (per min) | Status |
|---|
Mock costs are simulated estimates from public list prices. OpenAI and Google rows show MEASURED latencies from our 2026-08-25–27 bake-offs (non-streaming calls incl. connection setup; streaming architectures land lower). Real adapters fire only when their API key is set server-side — one pay-per-use call per request at list price.
Realtime voice call demo
This is the actual product surface, not a simulation. Click Start call, allow the microphone, and speak — a real agent answers in real time, transcribes both sides of the conversation live, and you can interrupt it mid-sentence just by talking. Prefer typing? Use the text box inside the demo. Cartesia streams text-to-speech.
Before you start: this call is not local. Your microphone audio streams to our voice service (voice.nexusagentworks.com) and on to the speech provider handling the turn (OpenAI or Google, with Cartesia for text-to-speech). We do not store the audio or the transcript; what each provider retains is governed by that provider’s own terms.
The microphone permission is scoped to this demo only — audio is streamed to our voice service for the conversation and is not stored by us. The demo also runs standalone at voice.nexusagentworks.com.
Text-to-speech playground
Choose a mocked provider and send text to your local proxy (127.0.0.1:8787). If the proxy isn't running, the page stays fully functional: it falls back to an identical client-side simulation so you can still compare profiles.
- Provider
- Audio reference
- Measured latency (proxy)
- Source
The classic fixed 700–1000 ms silence window adds 300–500 ms of dead air to every single turn. An adaptive gate closes a turn as soon as trailing silence is confidently speech-free — a research-backed projection per our latency study (−320 ms/turn vs the classic 800 ms window; projected, not yet live-measured).
Prefer live output over the built-in simulation? Run the bundled demo proxy from your own machine and reload this page.
One adapter interface, any provider
This preview is wired to the same provider-pluggable layer that will power real deployments — a single interface every voice vendor adapter must satisfy.
Cascade or speech-to-speech
Cascade providers expose transcription plus synthesis; level-3 S2S providers expose a streaming session instead. Your agent loop depends on the interface, never on a vendor SDK.
Keys stay server-side, always
API keys live in a gitignored env file and are read only inside your local proxy. The browser sees provider names and a configured flag — never a key value. Unconfigured providers answer with a clear error naming the missing variable.
Latency as a first-class metric
Every profile publishes its latency budget next to its price, because response gaps over about a second read as robotic. When real adapters land in stage 2, the marketplace table becomes the bake-off scoreboard.
Deterministic offline mocks today
The runnable profiles are pure functions — same input, same output, no network. That makes the whole surface testable end-to-end before a single dollar of API spend is committed.
Common questions
Do I need API keys to use this preview?
Not for the mocked profiles — those run fully offline and need no key, no sign-up and no purchase. The realtime call and the OpenAI/Google playground options run on our provider keys, server-side; you never supply one.
Is my text or audio sent anywhere?
It depends which control you use. Mocked profiles in the playground stay local: with the optional proxy running your browser posts to your own machine at 127.0.0.1 and a mock encodes the text there; without the proxy the simulation happens entirely in your browser. The OpenAI and Google playground options and the realtime call above do leave your machine — text or microphone audio goes to our voice service and on to that provider to be synthesised or answered. We do not store either; provider retention follows the provider’s terms.
Are the prices and latencies real quotes?
No — every figure is labeled a simulated estimate. Costs come from public list prices gathered in our market scan and latencies are planning budgets; both exist so you can reason about trade-offs before stage 2 runs measured benchmarks.
When can I hear real voices?
Installed for OpenAI (gpt-4o-mini-tts + Whisper) and Google (Gemini TTS + Flash ASR) — thin adapters behind the same interface, latency-measured in our 2026-08-25–27 bake-offs. Cartesia, ElevenLabs, and Deepgram are declared in the registry awaiting their adapters; invoking an unkeyed provider fails loudly rather than making a surprise network call.
Want this marketplace behind your phone lines?
Watch the full agent workforce run in the interactive demo, then see what a managed deployment costs. All figures are simulated estimates.