MLX Speech Daemon

msd is a manual-session, TTS-only daemon for MLX-Audio Qwen3 CustomVoice models.

V1 exposes one public command:

msd say "Hello." --voice Aiden --instruct "clear, warm, brisk"
msd up --model cv17-q8
msd status
msd render --text "Offline file." --output out.wav
msd hermes --input /tmp/in.txt --output /tmp/out.wav
msd stop
msd down
msd serve --foreground

The daemon is started automatically by say, up, and hermes when needed. It listens on a Unix socket, keeps one selected CustomVoice model warm, streams generated chunks to the default audio output for say, and interrupts active playback by default when a newer say request arrives.

V1 intentionally does not include STT, launchd, cron, containers, VMs, an autostart gateway, or Hermes source patches.

Install

uv tool install --force --editable ".[dev]"

If ~/.local/bin is on PATH, this exposes msd.

MP3 output requires ffmpeg on PATH. WAV output does not.

Model Aliases

cv17-q8 -> mlx-community/Qwen3-TTS-12Hz-1.7B-CustomVoice-8bit
cv17-q6 -> mlx-community/Qwen3-TTS-12Hz-1.7B-CustomVoice-6bit
cv06-q8 -> mlx-community/Qwen3-TTS-12Hz-0.6B-CustomVoice-8bit

Aliases and speaker names are case-insensitive.

Test Doubles

The real engine is used by default. Tests and dry smokes can use:

MSD_ENGINE=fake MSD_PLAYBACK=fake msd serve --foreground