MacOS's say command is a surprisingly useful addition for local shell scripts:

say "$(total) packages updated!"

It uses your configured Settings > Accessibility > Spoken Content voice to speak anything you pass it. But even the best voices (IMO, the "Siri" ones) still sound disappointingly robotic.

Better

Enter mlx-audio and kokoro:

alias tts='mlx_audio.tts.generate --model mlx-community/Kokoro-82M-4bit --stream --speed 1.25 --voice af_sky'

# Then...
tts --text "Do you recognize this voice? I bet you do."

# Or...
cat some-file.txt | tts

This alias uses a local MLX model to generate remarkably realistic sounding voices, in nearly realtime, on any recent M-series mac.

Installation

As of July 2025, installation requires some magic version numbers:

uv tool install mlx-audio --with pip --with mlx-vlm==0.1.26 --with mlx-lm==0.24.1

Which adds mlx-audio.* commands.

Running any command with --model and --voice will automatically download the necessary files from Hugging Face. These will be stored at ~/.cache/huggingface/hub.

Links

• https://github.com/Blaizzy/mlx-audio
• https://huggingface.co/hexgrad/Kokoro-82M
• https://huggingface.co/mlx-community/Kokoro-82M-4bit