izwi-ai/izwi

Voice AI runtime. Local first transcription, speaker diarization, TTS, and voice cloning with an OpenAI compatible API.

What it is: Izwi is a local-first voice AI toolkit. It's a single project that bundles a desktop app, a web interface, a command-line tool (CLI), and a local inference server. It runs entirely on your own machine—no cloud services or API keys are required.

What it does:

  • Real-time voice conversations: Uses local speech recognition (ASR), chat, and text-to-speech (TTS) models to let you talk to an AI.
  • Speech generation & editing: Text-to-speech, long-form Studio projects, voice cloning, voice design, and saved voices.
  • Audio analysis: Transcription, speaker diarization (identifying who spoke when), forced alignment (matching text to audio timestamps), and realtime speech-to-text.
  • Local chat & model management: Chat with local models, download/load/unload/delete models, view history, export data, and adjust settings.
  • OpenAI-compatible API: It exposes /v1 API routes for models, chat completions, audio speech, audio transcriptions, and preview Responses support. This means existing tools built for OpenAI's API can be pointed at Izwi instead.

How it works: You start the local server with izwi serve --mode web, then use the CLI to pull models (e.g., izwi pull Kokoro-82M) and run tasks like izwi tts "Hello from Izwi." or izwi transcribe audio.wav. The web app is then available at http://localhost:8080, with an API reference at /docs.

Supported models: It manages a catalog of model families including TTS (Qwen3-TTS, Kokoro-82M, Voxtral TTS, etc.), ASR (Parakeet, Whisper, Qwen3-ASR, etc.), diarization/alignment (Sortformer, Qwen3 ForcedAligner), and chat (Qwen3.5, Qwen3.6, etc.).

Privacy & hardware: Inference data stays local. Optional anonymous desktop analytics are disabled unless explicitly opted in, and they never send prompts, transcripts, audio payloads, local paths, or personal identifiers. macOS Apple Silicon builds use Metal; Linux and Windows release builds are CPU-only; CUDA is supported via Docker or source builds.

Tech stack: Built with Rust (Cargo binaries like izwi-cli and izwi-server) and Tauri for the desktop shell. Licensed under Apache 2.0.

Related

  • Project
    OpenWhispr/openwhispr

    A privacy-first, open-source voice-to-text application providing local AI dictation, meeting transcription with speaker diarization, and AI-powered notes.

  • Project
    kamjin3086/chatless

    A local-first desktop AI chat client that supports cloud and local models, featuring local RAG for document-based knowledge retrieval and MCP protocol integration.

  • Project
    katipally/openlive

    OpenLive is an open-source, local-first voice and vision layer that enables real-time, private interaction with AI models and coding agents without cloud-based audio fees.

  • Project
    isair/jarvis

    A private, local-first voice and text AI assistant that maintains a persistent memory of conversations and integrates with local tools to act as a personal companion.

  • Project
    Natively-AI-assistant/natively-cluely-ai-assistant

    A local-first, source-available AI meeting assistant that provides real-time transcription, context-aware answers, and advanced stealth features for interviews and professional calls.