The Izwi desktop app supports macOS 13.3 and later, and the terminal bundle supports macOS 12 and later. Apple Silicon Macs use Metal acceleration on macOS 15+; older versions remain supported with the CPU backend.

Requirements

  • macOS 13.3 (Ventura) or later for the desktop app; macOS 12.0 (Monterey) or later for the terminal bundle
  • 8 GB RAM minimum (16 GB recommended)
  • 10 GB free disk space (more for models)

Step 1: Download

Download the latest Izwi_<version>_aarch64.dmg from GitHub Releases. The DMG is for Apple Silicon Macs; there is no Intel build.

Step 2: Install

  1. Open the downloaded .dmg file
  2. Drag Izwi.app into your Applications folder
  3. Eject the disk image

Step 3: First Launch

  1. Open Izwi from Applications
  2. If prompted about an unidentified developer:
    • Go to System Settings → Privacy & Security
    • Click Open Anyway next to the Izwi message
  3. To use izwi in Terminal, open Settings > System > Command line tool and select Install. Izwi links izwi and izwi-server into /opt/homebrew/bin or /usr/local/bin and may ask for an administrator password.

Step 4: Verify

Open Terminal and run:

Optional: Install espeak-ng for Kokoro-82M

Kokoro-82M uses espeak-ng for phonemization. Install it before using Kokoro voices.

Install with Homebrew

Verify


Command-Line Only Installation

If you only need the CLI tools without the desktop app:
This installs izwi, izwi-server, and izwi-desktop to ~/.local/bin.

Apple Silicon Optimization

Izwi automatically selects CPU on macOS 12-14 and Metal on Apple Silicon with macOS 15+. On macOS 15+, you can explicitly request Metal:
Or set the environment variable:

Metal precision

Each model family picks its own precision on Metal. IZWI_METAL_DTYPE overrides that choice for every family: A per-family variable such as IZWI_WHISPER_DTYPE is more specific and takes precedence over IZWI_METAL_DTYPE. On Metal, a per-family f16 or bf16 setting is now applied as written instead of being replaced with F32. On Metal, Parakeet and Nemotron store their encoder’s linear and pointwise-convolution weights in F16 and keep activations, normalization, convolutions and the decoder in F32, as whisper.cpp does. This roughly halves their weight memory (about 2.5 GB to 1.3 GB each). A numeric self-check runs at load; if it fails, the model loads in F32 and the reason appears in its diagnostics under weight_precision. Set IZWI_PARAKEET_ASR_DTYPE=f32 or IZWI_NEMOTRON_ASR_DTYPE=f32 (or IZWI_METAL_DTYPE=f32) to keep every weight in F32. Granite Speech applies the same split to its conformer audio encoder (about 0.8 GiB less weight memory); its self-check result appears under execution.encoder_weight_precision, and IZWI_GRANITE_SPEECH_DTYPE=f32 (or IZWI_METAL_DTYPE=f32) keeps the encoder in F32. Granite’s language model keeps its F16 projections and attention on Metal behind a load-time range probe; if the probe fails it loads in F32, and the reason appears under execution.lm_precision. The same two settings now also keep the language model fully in F32. Voxtral Realtime keeps its F16 default on Metal and now runs a short encoder-and-decoder self-check at load; if it fails, Voxtral loads in F32 and the reason appears under execution.precision. IZWI_METAL_DTYPE=f32 runs it in F32. Qwen3-ASR (GGUF) checks its audio encoder and its text model the same way, each falling back to F32 on its own (execution.audio_precision and execution.text_precision); IZWI_METAL_DTYPE=f32 now also applies to it. VibeVoice-ASR now runs its language model in F16 on Metal (it was F32) behind the same kind of check, and keeps its audio tokenizers and connectors in F32 with F16 feed-forward weights, about 16 GiB of weights instead of 31 GiB. The results appear under execution.lm_precision and execution.tokenizer_weight_precision, and IZWI_VIBEVOICE_ASR_DTYPE=f32 (or IZWI_METAL_DTYPE=f32) keeps it in F32. On Apple M3 and newer GPUs, BF16 is the intended default for Granite, Voxtral, VibeVoice and Qwen3-ASR, but it stays off until izwi’s Metal attention kernels support BF16; those GPUs use F16 for now. Whisper runs its F16 checkpoint in F16 on every Apple GPU, as whisper.cpp, mlx-whisper and OpenAI Whisper do. The log-mel input, attention softmax, LayerNorm statistics and logits stay in F32. This halves Whisper’s weight memory (about 3.0 GiB to 1.5 GiB). A numeric self-check runs one encoder window and one decoder step at load; if it fails, Whisper loads in F32 and the reason appears in its diagnostics under device.precision. Set IZWI_WHISPER_DTYPE=f32 (or IZWI_METAL_DTYPE=f32) to run Whisper in F32.

Data Locations


Uninstall

Quick Uninstall

From a cloned repository:
Or download the uninstall script from releases:

Manual Uninstall

Verify Removal


Troubleshooting

”Izwi can’t be opened because it is from an unidentified developer”

  1. Go to System Settings → Privacy & Security
  2. Scroll down to find the Izwi message
  3. Click Open Anyway

Command not found: izwi

If you installed the desktop app, open Settings > System > Command line tool, select Install, then open a new Terminal window. If you installed with ./scripts/install-cli.sh, the tools are in ~/.local/bin, which may not be in your PATH. Add this to your ~/.zshrc:
Then reload:

Kokoro error: espeak-ng not found

Install espeak-ng:

Metal acceleration not working

Ensure you’re on Apple Silicon and running macOS 15.0+:

Next Steps