Voice AI Providers Supported by Micdrop
A Micdrop server hands each part of the call to a provider of your choice, so you can mix them for cost and quality, run them all locally, or write your own.
- Speech to text transcribes the user.
- LLM (Agent) writes the answers.
- Text to speech speaks them.
- Realtime models hear and speak in one connection, in place of the three above.
- Classifiers turn each turn of the user into typed decisions next to the agent.
Each provider ships as its own npm package. The table shows which part of a voice agent each one covers, with a link to its page.
| Provider | Package | Speech to text | Text to speech | LLM | Realtime | Classifier |
|---|---|---|---|---|---|---|
| Vercel AI SDK | @micdrop/ai-sdk | Yes | ||||
| Cartesia | @micdrop/cartesia | Yes | ||||
| ElevenLabs | @micdrop/elevenlabs | Yes | ||||
| Gemini | @micdrop/gemini | Yes | Yes | Yes | Gemini Live | Decisions |
| Gladia | @micdrop/gladia | Yes | ||||
| Gradium | @micdrop/gradium | Yes | Yes | |||
| Kokoro (local) | @micdrop/kokoro | Yes | ||||
| Mistral | @micdrop/mistral | Voxtral | Voxtral | Yes | ||
| OpenAI | @micdrop/openai | Yes | Yes | Yes | Realtime | Decisions |
| Phonon (local) | @micdrop/phonon | Yes | ||||
| Piper (local) | @micdrop/piper | Yes | ||||
| Pocket TTS (local) | @micdrop/pocket-tts | Yes | ||||
| Qwen3-TTS (local) | @micdrop/qwen-tts | Yes | ||||
| TypeSafe | @micdrop/typesafe | Jev | ||||
| Whisper (local) | @micdrop/whisper | Yes |
The AI SDK package opens the LLM column to every model the Vercel AI SDK supports, Anthropic, Groq or a local Ollama server included. Any other provider plugs in through a custom speech to text, text to speech, agent or classifier.
Mistral, Gladia and Gradium cover the three parts of a call with European companies, as the sovereign voice AI guide shows.
Mixing providers
Each part is a separate option of the server, so every one can come from a different provider:
import { MicdropServer } from '@micdrop/server'import { MistralAgent } from '@micdrop/mistral'import { GladiaSTT } from '@micdrop/gladia'import { CartesiaTTS } from '@micdrop/cartesia'
new MicdropServer(socket, { stt: new GladiaSTT({ apiKey: process.env.GLADIA_API_KEY, }), agent: new MistralAgent({ apiKey: process.env.MISTRAL_API_KEY, model: 'mistral-large-latest', }), tts: new CartesiaTTS({ apiKey: process.env.CARTESIA_API_KEY, voiceId: 'cartesia-voice-id', }),})The speech to text, the agent and the text to speech all have a local counterpart, so a whole conversation can also run on your machine with the local models guide.