Skip to main content
API7.ai Logo

Voice Agent Platforms

Voice applications bring together several latency-sensitive stages: audio transport, speech recognition, turn detection, language-model inference, tool execution, and speech synthesis. Voice agent platforms coordinate those stages so applications can hold realtime conversations over the web, in an application, or on a telephone call.

The language model is only one part of that runtime, but it is often the part that organizations need to govern centrally. When a voice platform accepts an OpenAI-compatible LLM endpoint, its text model requests can pass through AISIX for caller authentication, model aliases, routing, policy, and telemetry. Teams can then manage model access in AISIX without replacing the platform that delivers the voice experience.

The voice application or hosted platform remains responsible for the media session and converts speech into conversation content before calling AISIX. It sends that content and any tool definitions to the gateway. AISIX then authenticates the caller, resolves the model alias, applies configured policy, records telemetry, and dispatches the request to the selected model provider.

Choose a Platform Guide​

All guides use the same AISIX values, but each platform exposes them differently:

PlatformIntegration surfaceGuide
ElevenLabs AgentsHosted Custom LLM configurationElevenLabs Agents
LiveKit AgentsPython OpenAI LLM pluginLiveKit Agents
PipecatPython OpenAILLMServicePipecat
VapiHosted custom-llm model configurationVapi

ElevenLabs Agents and Vapi host the voice runtime. LiveKit Agents and Pipecat are application frameworks, so your application also configures its transport, speech-to-text service, text-to-speech service, and turn detection.

Prepare AISIX​

Each platform needs these gateway-facing values:

  • An HTTPS AISIX proxy URL the platform or application can reach.
  • An AISIX caller API key dedicated to the voice application.
  • A model alias that the caller key can access through POST /v1/chat/completions.

If AISIX is already deployed for your organization, obtain these values from the team that manages it. Otherwise, follow the Open-Source AISIX Gateway Quickstart or AISIX Cloud Quickstart, or contact API7 for Hybrid Cloud access.

Use a caller key scoped to only the aliases the voice application needs. Configure request and token limits for expected conversation volume, and set an expiration time for temporary evaluations.

Understand the Boundary​

AISIX receives the model request after the voice platform has converted speech into conversation content. AISIX therefore does not replace the platform's audio transport, speech recognition, speech synthesis, voice selection, interruption handling, or telephony.

Gateway guardrails can inspect supported request and response text on the Chat Completions path. They do not inspect audio that remains inside the voice platform. See Guardrail Behavior for exact endpoint and content coverage, and review each platform's recording, transcript, retention, and regional processing settings separately.

Voice applications normally stream the model response so speech synthesis can begin before the complete answer is available. Confirm the model alias supports Chat Completions streaming, and account for network distance between the voice runtime, AISIX gateway, and model provider when measuring response latency.

Verify an Integration​

Use one short conversation to confirm the complete path:

  1. The voice platform accepts a spoken or text test message.
  2. AISIX records POST /v1/chat/completions for the expected caller key and model alias.
  3. The platform receives the streamed model text and returns it to the user as audio or text.

After the first success, test interruption behavior and any function tools the application depends on. Tool calling requires the selected upstream model and the platform integration to preserve the OpenAI-compatible tool-call stream.

Next Steps​