Voximplant has native clients for major Voice AI providers. For text-based LLMs outside that list, both Chat Completions and Responses API clients accept a baseUrl parameter. Point it at any OpenAI-compatible endpoint, and your core scenario logic can stay the same, though some provider-specific parameters may need adjusting.
Choosing between the two
Chat Completions is the most widely supported OpenAI-compatible API shape. It’s a good starting point if you need broad provider coverage or want to reuse existing bot logic. It works across the largest set of providers and handles streaming output text naturally in a cascaded pipeline.
See a working example with Together AI.
Responses API is often the better fit for low-latency cascaded voice AI. Streaming text events map directly into realtime TTS, which keeps latency low in a cascaded pipeline. It also supports built-in tools and structured response items where the provider offers them. Responses API is worth starting with, if it is supported by your chosen provider.
See a working cascaded pipeline example with Deepgram, Groq, and Inworld.
Both interfaces accept baseUrl, so provider switching works the same way regardless of which you use.
Keeping LLM requests in your region
When a call transcript goes to an LLM endpoint, it goes wherever that endpoint is hosted — a public cloud region, a private cloud deployment, or your own datacenter. For clients with data localization requirements, that's a compliance question: a transcript may contain personal data, and cross-border transfers require legal justification.
Some LLM providers offer EU-based endpoints or let you deploy models on EU cloud infrastructure. Others support deployment on specific cloud infrastructure, or on premises, if data cannot leave your perimeter at all. In each case, baseUrl is what points VoxEngine at the right endpoint. The scenario code stays the same regardless of where the model runs.
For providers that offer regional deployments, the BYO LLM documentation lists available base URLs.
More reasons to switch providers
- Cost optimization: run a lighter model for simple tasks like FAQ handling or appointment confirmation, and route complex calls to a more capable model, all within the same application.
- Model testing: benchmark providers against each other on real call data without rebuilding your pipeline for each test.
This is the first post in a series on building compliant voice AI on Voximplant. More on running a fully sovereign voice AI stack is coming
Resources
BYO LLM documentation
Chat Completions example: Together AI
Responses API example: Deepgram-Groq-Inworld
Sign up for Voximplant
