Mix-and-match voice AI without the wiring

What it does

Verbi is a voice assistant pipeline written in Python. It records audio from your microphone, sends it to a transcription provider, passes the resulting text to a language model, then feeds the reply to a text-to-speech service and plays it back. The whole chain is controlled through a central config.py file, so switching from, say, Groq transcription to OpenAI TTS is mostly a matter of changing a string.

The interesting bit

The project is essentially well-organized glue code. Its value lies in doing the boring plumbing—audio I/O, API key juggling, and format shuffling—so you can focus on comparing how different models sound and respond. For researchers or developers evaluating voice stacks, that boilerplate is usually the part that eats an afternoon.

Key highlights

• Swap transcription, response generation, and TTS providers via config.py without rewriting the pipeline.
• Supports cloud APIs including OpenAI, Groq, Deepgram, and ElevenLabs, plus local options via Ollama for LLMs and Piper/MeloTTS for speech.
• Includes helper servers for local TTS models (Piper and MeloTTS) that expose them through FastAPI endpoints.
• Explicitly designed as a research and development testbed rather than a consumer product.

Caveats

• “Local” transcription and local response generation are currently placeholders in the model list, not fully implemented.
• Running local models requires manually launching separate servers (Ollama, Piper, or MeloTTS) before starting the assistant.
• Real-time streaming is on the roadmap but not yet supported.

Verdict

Grab this if you are building a voice product and need to benchmark providers or iterate on pipeline architecture quickly. Skip it if you want a polished, batteries-included voice assistant that works out of the box.