The Deepgram Voice Agent API is one real-time API that combines speech recognition, language model orchestration and speech synthesis, so developers do not have to stitch separate services together. Built-in barge-in detection, turn-taking prediction, function calling and mid-session control keep conversations moving without awkward pauses. Teams can bring their own LLM or text-to-speech provider while keeping Deepgram's orchestration and streaming pipeline. The Voice Agent API is priced at a flat $4.50 per hour on the full Deepgram stack, with built-in rate reductions when bringing your own models.
It is aimed at software developers, AI engineers and enterprise platform teams that need production voice agents, especially in industries with strict compliance requirements. Deployment choices include fully managed, dedicated single-tenant, inside a customer VPC or self-hosted, with support for HIPAA, GDPR and regional data residency. Startups can begin on pay-as-you-go with a $200 credit, while large organizations can move to prepaid Growth credits from $4,000 a year or an Enterprise agreement.
Practical uses include phone-based customer support agents that handle interruptions and take actions, ordering and scheduling assistants, healthcare intake lines, and in-app conversational assistants. Because Deepgram runs speech-to-text, text-to-speech and orchestration itself, latency can be tuned across the whole path, which matters for natural turn-taking. Usage-based pricing means a prototype costs little, and the same pipeline carries over when a project moves toward production. A playground and demo let developers hear an agent before writing code, and teams can swap in their own models when they want to control cost or behavior at scale.
