Deepgram Voice Agent API

What it does

Deepgram Voice Agent API is a WebSocket API that combines speech-to-text, language model orchestration and text-to-speech in a single connection. Developers send audio in and receive audio out, while the API handles end-of-turn detection, interruption handling and conversation state. The developer supplies the prompt, tools and application logic.

Teams use it to build phone and web voice agents without wiring separate speech, LLM and voice services together. It exposes each conversation turn as inspectable events and returns a latency breakdown by speech-to-text, LLM time to first token, tool time and text-to-speech. Teams can bring their own LLM or voice provider, or use Deepgram's, and can change the model or voice mid-conversation.

Suited to developers and enterprises building real-time voice applications who can write code against a WebSocket API. Access requires an account and API key through the Deepgram console, and billing is per minute of connection time. Limits include concurrency caps that vary by plan and region, and self-hosting on Kubernetes is described for enterprise deployments with NVIDIA-based requirements.

What it can do

• Turn-taking with end-of-turn detection and interruption handling inside the API
• Automatic fallback so an LLM or text-to-speech failure does not end the call
• Change agent context, model or voice mid-conversation without interrupting the flow
• Bring your own OpenAI-compatible or Bedrock LLM, and Deepgram or third-party TTS
• Per-turn latency report and inspectable conversation events, on by default
• Multilingual speech input across 60+ languages with Deepgram Speech-to-Text

Works with

• OpenAI-compatible LLM endpoints
• Amazon Bedrock
• Third-party text-to-speech providers

Example workflows

Phone support voice agent

Illustrative example. Input: caller audio over WebSocket. Steps: transcribe speech with turn detection, send text to a customer-chosen LLM with function calling for account lookup, speak the reply with Deepgram or a chosen TTS voice. Output: spoken responses and per-turn latency events.

Web app voice assistant

Illustrative example. Input: browser microphone audio streamed to the WebSocket. Steps: detect user end-of-turn, call a configured model, handle barge-in by stopping agent audio. Output: synthesized audio stream and inspectable conversation events for debugging.

Before you choose

• Billing is per minute of WebSocket connection time, so idle or long calls count toward cost.
• Concurrency limits apply per project and differ by plan and region.
• Self-hosting requires your own Kubernetes cluster and a separately connected language model endpoint.

Listed under

• Support

Get started

Free starting credit is listed on the pricing page.

1. Create a Deepgram console account and generate an API key.
2. Connect to the Voice Agent WebSocket with the Deepgram SDK or API and send a prompt, tools and voice configuration.
3. Optionally configure an OpenAI-compatible or Bedrock LLM endpoint and a third-party TTS provider with ordered fallback.

Billing uses WebSocket connection minutes, including idle connected time. Bring-your-own LLM or TTS configurations have separate provider costs and different Deepgram rates. The $200 starting credit is a one-time allowance, not a recurring free tier. Check the current configuration-specific rate table.

Source: deepgram.com/pricing, checked 29 Oct 2026. Prices can change; check before you buy.

Sources for this overview

Product pages, documentation, and pricing checked 29 Oct 2026. Each source supports part of the overview; availability and prices can change.

• Voice Agent API | Deepgram ↗deepgram.com
• Deepgram pricing | Scalable speech-to-text, text-to-speech & Voice Agent APIs | Deepgram ↗deepgram.com
• Welcome to Deepgram's docs | Deepgram docs ↗developers.deepgram.com

Company website

deepgram.com: product, pricing and policy pages

YC directory mirror

Batch and company name

Social accounts

Accounts the company links, and its posts on X

Human review

Checked by a person

knowledge_base

Last checked 29 Oct 2026. Every field keeps its source and date; unknown means we couldn’t confirm it.