Overview
SubVerse AI lets you choose the AI models that power each agent. You pick models in the agent’s configuration. For example, the LLM Model, Transcript Model and TTS Model fields in the Voice Agent configuration. Models fall into four categories:Speech-to-text (STT)
Transcribes what the caller says. 10 models from Deepgram, OpenAI, Google, ElevenLabs and Sarvam.
LLMs
Understands the conversation and decides what to say or do. 27 models from OpenAI, Google and Anthropic.
Text-to-speech (TTS)
Turns the agent’s reply into speech. 16 models from ElevenLabs, Cartesia, Google, OpenAI, Azure, Inworld, Murf and Sarvam.
Realtime / Live models
Single speech-to-speech models that listen and speak directly. 7 models from OpenAI and Google.
How agents use models
Chat, email and background agents use an LLM only. Voice agents can work in one of two ways:How to read these pages
- Official figures come from the provider’s own documentation, model cards or announcements, and each one links to its source.
- Third-party figures are used only when the provider doesn’t publish benchmark numbers for a model, or publishes them only as charts. They are marked Third-party and name the independent publisher.
- Languages available in SubVerse AI lists (on the STT and TTS pages) show the languages and regional variants you can select for that model on the platform. Counts include regional variants, such as English (US) and English (India), so they can differ from the provider’s own language count. The provider may also support languages that are not yet available in SubVerse AI.
- Benchmarks are not directly comparable across providers. Each provider chooses its own benchmarks, settings and test conditions. Use them as a guide, and test models with your own use case.