Local-first audio inference engine for TTS, ASR, and voice AI workflows.

Website • Documentation • Releases • Getting Started

Izwi is a privacy-focused audio AI platform that runs entirely on your machine. No cloud services, no API keys, no data leaving your device.

Core capabilities:

• Voice Mode — Real-time voice conversations with AI
• Text-to-Speech — Generate natural speech from text
• Studio — Build long-form TTS projects and exports
• Speech Recognition — Convert audio to text with high accuracy
• Speaker Diarization — Identify and separate multiple speakers
• Voice Cloning — Clone any voice from a short audio sample
• Voice Design — Create custom voices from text descriptions
• Forced Alignment — Word-level audio-text alignment
• Chat — Text-based AI conversations

The server exposes OpenAI-compatible API routes under /v1.

Download the latest .dmg from GitHub Releases:

1. Open the .dmg file
2. Drag Izwi.app to Applications
3. Launch Izwi

wget https://github.com/izwi-ai/izwi/releases/latest/download/izwi_amd64.deb
sudo dpkg -i izwi_amd64.deb

Download and run the installer from GitHub Releases.

Full installation guides: macOS • Linux • Windows • From Source

izwi serve

Open http://localhost:8080 in your browser.

izwi pull Qwen3-TTS-12Hz-0.6B-Base

izwi tts "Hello from Izwi!" --output hello.wav

izwi pull Parakeet-TDT-0.6B-v3
izwi transcribe audio.wav

Long-form ASR is handled automatically: Izwi now chunks long recordings, stitches overlapping transcripts, and returns a full transcript instead of only the first model window.

Optional tuning knobs:

IZWI_ASR_CHUNK_TARGET_SECS=24
IZWI_ASR_CHUNK_MAX_SECS=30
IZWI_ASR_CHUNK_OVERLAP_SECS=3
# Optional: preload models at server startup to reduce first-request cold latency.
# Comma-separated model IDs (for example Whisper-Large-v3-Turbo,Qwen3.5-4B)
IZWI_PRELOAD_MODELS=Whisper-Large-v3-Turbo
# Optional: run a short synthetic ASR warmup after preloading (enabled by default).
IZWI_WARMUP_PRELOADED_MODELS=1
IZWI_ASR_WARMUP_DURATION_MS=800
# Optional: tune text streaming queue depth when using per-character ASR streaming.
IZWI_STREAM_TEXT_QUEUE_CAPACITY=4096

Izwi desktop supports optional, opt-in anonymous usage analytics powered by Aptabase.

• Disabled by default until users explicitly opt in.
• Can be enabled during onboarding or later in Settings.
• Users can opt out at any time.
• No prompts, transcripts, audio payloads, local paths, or personal identifiers are sent.

To enable analytics transport in the desktop shell, set the app key in the runtime environment:

APTABASE_APP_KEY=A-US-XXXXXXXXXXXXXXX

Use the exact key from Aptabase (for example A-US-... or A-EU-...).

Without this variable, analytics calls are treated as no-op events.

Run izwi list to see all available models.

Full model documentation: Models Guide

Apache 2.0

• Qwen3-TTS by Alibaba
• Parakeet by NVIDIA
• Gemma by Google
• HuggingFace Hub for model hosting