One Voice Agent API for the whole conversation
Speech-to-text, text-to-speech, and voice agent orchestration all run through one API, so you don’t have to build and maintain the voice stack yourself.
Try for freeLeading voice teams build with Deepgram
Inside the Voice Agent API
Turn-taking that feels natural, fallback that keeps you online, models you can bring yourself, and deployment that fits your stack.
Skip the DIY setup
WebSocket instead of wiring together speech-to-text, an LLM, and text-to-speech. Audio never round-trips through your infrastructure.
Runs turn-taking for you
End-of-turn detection, interruption handling, and conversation state live inside the API, not in silence timers you build and re-tune yourself.
Knows when to speak
Your agent fills silence when someone trails off, stays quiet through pauses, and checks in on a silent line.
Change the agent mid-conversation
Add context, swap the model or voice, and inject exact wording without interrupting the flow. Automatic fallback keeps an LLM or text-to-speech failure from ending the call.
Bring your own models
Bring your own LLM and voices or use Deepgram's.
Deploy where you need to
Global, EU, India, or Australia endpoints
Multilingual coverage
Speak their language
Reach more customers with languages your agent speaks end-to-end.
Trusted by builders, ready for scale
Over 1,000 organizations have built on Deepgram Voice Agent API, with enterprise-grade reliability and a foundation proven on billions of minutes of transcribed conversation.
- 1,000+Organizations building voice agents on Deepgram
- 99.99%Enterprise-grade reliability
- 47MMinutes of voice agent conversation
Related articles
Deepgram
Deepgram Voice Agent API vs OpenAI Realtime API: Building voice agents in 2026
AI Engineering & Research
Deepgram
Turn detection in voice agents: silence timeouts vs VAD vs semantic models
AI Engineering & Research
Deepgram
Voice agent function calling: real-time data lookup during a call
AI Engineering & Research
Hasan Jilani
Put a Deepgram voice agent on any web app in minutes
AI Agents
FAQs
One WebSocket that runs speech-to-text, LLM orchestration, and text-to-speech together. You send audio in and get audio out. Deepgram handles end-of-turn detection, barge-in, and function calling; you supply the prompt, the tools, and your application logic.
Yes, for both. Bring a language model through your own OpenAI-compatible endpoint or a Bedrock deployment, with ordered fallback if one fails. Text-to-speech can be Deepgram's or a third-party provider you already use. Deepgram's documentation lists every supported provider.
When a user starts speaking, Deepgram Voice Agent uses Flux's model-integrated turn detection to interrupt the agent's audio and listen. When the user finishes, Flux detects the end of the turn in about 260 ms, without requiring a separate VAD for endpointing. You can also use ForceEndTurn to end a turn manually when you need explicit control.
Yes. Deploy the voice agent on your own Kubernetes cluster with Deepgram’s Helm chart. Speech-to-text, text-to-speech, and turn detection run in your environment. Connect to your own endpoint for the language model.
Voice agents can understand 60+ languages with Deepgram Speech-to-Text. For voice output, use Deepgram Text-to-Speech or bring your own TTS provider. Together, these options let you build the complete conversational loop in the languages your agent needs.
Billing is per minute of WebSocket connection time. Pay-as-you-go starts at $0.075/min, and that rate covers speech-to-text, a managed language model, and text-to-speech, with no separate provider bills. Bring your own LLM or voice and it drops to as low as $0.050/min. The Advanced model tier is $0.163/min. Volume pricing and full tiers are on the pricing page.
Every conversation turn returns a latency report, broken down by speech-to-text, LLM time to first token, tool time, and text-to-speech. It’s on by default, with no flag required. Conversation text, agent thinking, and function calls all arrive as inspectable events.