Command line

Run loudkit --help or loudkit <command> --help for the full options.

Speak and listen

loudkit speak "Hello from loudkit." --voice joe --play -o hello.wav

speak saves a WAV. --play then waits for the system player to finish: afplay on macOS, aplay or paplay on Linux, and PowerShell on Windows. Without the flag, it prints a playback command. The Python synthesize method never plays audio.

If no player is available, the command names what to install and exits with an error. The saved WAV remains available, including when playback fails. Linux packages are alsa-utils or pulseaudio-utils; Windows needs PowerShell.

Use --checkpoint to select the model, --language to override the voice’s language, and --seed for repeatable generation. A positional - reads stdin.

Preview speech text

loudkit text 'Dr. Smith paid $12 on 2024-01-02. [12]' --language en

text runs the same text funnel as synthesis without loading or downloading a model. It prints prepared text to stdout, so it can be redirected or piped. The language defaults to en; a positional - reads stdin.

Stderr shows a word diff: replaced, removed, or inserted spans. These include number and date expansions and dropped annotations. The diff compares input with output; it does not attribute changes to individual processing stages. Whitespace-only changes are omitted. Empty or entirely dropped input prints an empty line.

Clone a voice

loudkit clone recording.wav --name mine --language en writes a portable voice profile. --device onnx and --device coreml use the corresponding enrollment graphs without torch. Install the matching backend and audio extras, and prepare cloning assets with download --with-cloning --for onnx or download --with-cloning --for coreml. See cloning.