>_
← Back to projects Voice AI

Fluency.ai

A full on-device voice pipeline — speak in English, get a spoken reply back. No internet, no accounts, no cloud.

Python Moonshine STT Mistral 7B Kokoro TTS Gradio
FLUENCY — DEMO
Fluency AI demo video thumbnail

A conversation, entirely offline

Fluency AI is a voice conversation app: talk to it in English, and it listens, thinks, and talks back — entirely on your own computer, with no internet connection required after setup.

It chains three local models into one pipeline — Moonshine for speech-to-text, Mistral 7B for the reply, and Kokoro for text-to-speech — served over a self-signed HTTPS Gradio app so it's reachable from a phone on the same Wi-Fi, not just the machine it runs on.

What it does

Voice In, Voice Out
  • Click the mic, speak in English, get a spoken and written reply back
  • Conversation history stays in the chat panel until cleared
Fully On-Device
  • Speech recognition, LLM reasoning, and voice synthesis all run locally
  • No accounts, no API keys, no cloud calls — nothing leaves the computer
Cross-Device on LAN
  • Self-signed HTTPS server, reachable from a phone or tablet on the same Wi-Fi
  • Just open the PC's local IP on port 7860 from any device
Local Model Stack
  • Moonshine handles speech-to-text, Mistral 7B (via Ollama) the reply, Kokoro the voice
  • All three models run on CPU via ONNX — no GPU required

The pipeline, step by step

1

You speak — the browser records audio through the microphone.

2

Moonshine (STT) transcribes the audio to text, running on CPU via ONNX through FastRTC.

3

The transcript goes to Mistral 7B Instruct via a local Ollama server, which generates a conversational reply.

4

Kokoro (TTS) converts the reply text to natural-sounding speech, again on CPU via ONNX.

5

The browser plays the audio and shows the text reply — the whole round trip happens without an internet connection.


All local, by design. Moonshine, Mistral 7B, and Kokoro all run on the machine itself. The Gradio server uses a self-signed HTTPS certificate specifically so microphone access works from other devices — phones, tablets — on the same network.

Tech Stack

Speech-to-Text
Moonshine (via FastRTC)
LLM
Mistral 7B Instruct Ollama
Text-to-Speech
Kokoro (via FastRTC)
Web UI
Gradio
Tooling
uv Self-signed HTTPS (cryptography)

Check out the code

Full source, setup instructions, and architecture notes are on GitHub.