How it works
A PeachDesk voice agent does four things on every call: it listens, it thinks, it speaks, and it stays connected. PeachDesk is the voice AI platform that runs those four jobs, and lets you decide where each one executes.
In plain English: ears, brain, voice, and a line to the outside world.
The ears.
Speech-to-text turns the caller's audio into text the agent can reason over, as the conversation happens.
The brain.
A language model follows your workflow: your prompts, your tools, your knowledge base, and your logic for transfers, handoffs, and ending the call.
The voice.
Text-to-speech turns the agent's reply into natural-sounding audio, in the voice you chose or cloned.
The line.
Telephony carriers or browser WebRTC carry the audio in both directions, on phone numbers and voice surfaces you control.
Every stage is a separate decision. For each agent, speech-to-text, the language model, and text-to-speech can each run on Frontier mode (leading cloud providers, metered), on your own provider keys (BYOK), or on Local mode (a full open-model stack on your infrastructure, with no per-minute provider charge; you pay for GPU capacity).
Describe your use case and let the AI Agent Builder draft a starting agent, pick one of 26 pre-built templates, or design the flow from scratch on a visual canvas of 16 node types. Choose multi-step conversation flows or a single-prompt agent.
Attach knowledge-base documents, add custom HTTP API tools, call transfer, agent-to-agent handoff, in-call SMS, and keypad navigation.
Test by text chat, AI-simulated callers, in-browser voice, or real phone calls. Text chat skips the speech stages, so use voice tests before go-live.
Publish an immutable version, route your own phone numbers to the agent through your telephony provider, or add the agent to your website with one script tag.
Every call is pinned to the exact agent version that served it. A configurable QA step scores real calls automatically, and analytics break down spend, cost of goods, margin, and sourcing mode per call. Call recordings and transcripts live in object storage your deployment controls, behind signed, time-limited URLs, with configurable retention windows and automated purge.
This page is the plain-English version. For the technical model of pipelines, nodes, and sourcing, read the voice agent architecture explainer, or go straight to the product documentation.
A voice AI agent is software that holds a spoken conversation with a person over a phone call or browser audio. It listens with speech-to-text, decides what to say with a language model, speaks with text-to-speech, and connects to the telephone network or WebRTC to carry the audio. Production voice agents also need orchestration, telephony, testing, and cost controls, which is what a voice AI platform provides.
The caller speaks and speech-to-text transcribes the audio. Your workflow, built from typed nodes, conditional edges, prompts, tools, and knowledge bases, decides the next step with the language model. Text-to-speech turns the reply into audio, and your telephony provider or browser WebRTC carries the sound. Each stage runs on the sourcing mode you chose for that agent.
A chatbot exchanges text on a screen; a voice agent holds a real-time spoken conversation. That adds hard problems text never faces: transcription accuracy, turn-taking and interruptions, speech synthesis, telephony, and a latency budget small enough for natural conversation. A voice AI platform exists to solve exactly those problems.
The documentation walks through the same pipeline with real configuration.