WebRTC
Browser voice transport
Use WebRTC when you need direct low-latency microphone input and audio output in a web product.
Best fit: browser-based voice assistants with responsive turn-taking.
Build realtime voice agents with GPT Realtime 2
Design low-latency speech-to-speech assistants, live translation workflows, streaming transcription, and tool-enabled voice experiences for production teams.
Independent builder-focused workspace for GPT Realtime 2, WebRTC, tools, and voice operations.
Voice Studio
Voice generation for narration, dialogue, and transcripts
Generate narration, dialogue, or transcriptions with AI voice tools.
Voice
AI voice
Realtime voice agent planning
GPT Realtime 2 is for teams planning low-latency voice agents with the OpenAI Realtime API
GPT Realtime 2 is an independent workspace for mapping realtime speech-to-speech agents, browser WebRTC sessions, server-side WebSocket audio, streaming transcription, live translation, tool calls, and production usage controls before a team ships a voice workflow.
Last updated:
Key takeaways
- Use WebRTC when a browser or mobile client needs responsive microphone input and audio output.
- Use WebSocket when a backend service owns audio routing, recording, telephony, or policy enforcement.
- Treat tool calls, escalation, monitoring, and cost controls as launch requirements, not afterthoughts.
Architecture fit table
| Workflow signal | Recommended setup | Why it matters |
|---|---|---|
| Browser voice assistant | WebRTC session with short-lived client access | Keeps microphone and playback latency low while avoiding long-lived secrets in the client. |
| Call center or telephony path | Server-controlled realtime audio with explicit handoff rules | Lets the backend manage routing, logs, compliance review, and human escalation. |
| Live translation or transcription | Separate session settings, transcript review, and usage budget | Keeps language handling, quality checks, and cost forecasting visible to operators. |
Primary references
Voice-agent capabilities for serious workflows
Keep the homepage focused on what teams actually need: responsive conversations, controlled sessions, useful transcripts, and tool actions that fit existing systems.
Speech-to-speech agents
Natural realtime conversations for support, coaching, intake, and guided operations.
Streaming transcription
Capture spoken sessions as structured text for review, search, QA, and follow-up.
Live translation workflows
Route multilingual conversations through a voice experience that stays usable in the moment.
Tool-enabled conversations
Let voice agents check records, create tickets, update systems, or trigger approved actions.
Session control
Tune instructions, voice behavior, context, and handoff rules for repeatable outcomes.
Usage visibility
Plan around session length, model choice, tools, and context so teams can budget confidently.
Workflow
From voice idea to production-ready agent
A clean operating model for building realtime voice systems without making the first screen feel like an experiment.
1
Define the agent
Write the role, boundaries, escalation rules, and success criteria before touching transport or tools.
2
Configure realtime sessions
Choose voice behavior, input modes, turn handling, and context strategy for the target channel.
3
Connect tools and data
Attach only the systems the agent needs, with explicit permissions and predictable failure paths.
4
Review usage and launch
Track transcript quality, latency, tool activity, and credit consumption before scaling traffic.
Architecture
Implementation patterns for realtime voice
Use the right transport and session shape for the channel: browser voice, server-side audio, secure client access, and tool-backed conversations.
Explore Architecture
Audio pipeline
Server-side streams
Use server-controlled audio streams when backend orchestration, recording, or telephony integration matters more.
Best fit: controlled infrastructure, call routing, compliance review, and server-owned state.
Security
Ephemeral access
Issue short-lived client secrets from your server so browsers can connect without exposing privileged credentials.
Best fit: production clients that need secure session startup and policy enforcement.
Tooling
Tools and policies
Connect function calls, business rules, retrieval, and handoff paths so the voice agent can act safely.
Best fit: support, sales, training, operations, and internal copilots.
Built around practical production constraints
The homepage should signal that this is a professional voice-agent platform, not a novelty page.
128K context window for long-running realtime workflows
128K
context window for long-running realtime workflows
WebRTC browser voice transport for low-latency interaction
WebRTC
browser voice transport for low-latency interaction
Tools function calling for workflow actions and system handoff
Tools
function calling for workflow actions and system handoff
Where realtime voice agents fit
Position GPT Realtime 2 around concrete business conversations instead of generic chat features.
Customer support voice agents
Answer routine questions, collect context, and hand off cleanly when a human should step in.
Sales qualification calls
Capture needs, route leads, and update pipeline tools while keeping the conversation natural.
Language learning tutors
Run spoken practice with corrections, summaries, and adaptive lesson flow.
Live translation assistants
Help multilingual teams communicate across calls, field work, travel, and operations.
Meeting and field copilots
Turn spoken updates into structured notes, tasks, and follow-up records.
Internal operations assistants
Guide staff through checklists, policy questions, and system actions hands-free.
GPT Realtime 2 FAQ
Clear answers for teams evaluating realtime voice agents.