Cloning a voice

A voice in loudkit is a handful of tensors, not a model: a file of about 150 KB you can copy, ship and version on its own. Making one from five to ten seconds of clean audio is the enroll extra:

pip install "loudkit[torch,audio,enroll,hub]"

Two front ends, one enrollment: loudkit clone in the shell, and enroll in each of the five SDKs. There is no MCP clone tool.

Enrollment runs the speaker encoder and the speech tokenizer, models that synthesis never loads. Their weights are a second file beside the 747 MB synthesis checkpoint, 523 MB, fetched the first time you clone or with loudkit download ... --with-cloning.

From the shell

loudkit clone my-recording.wav --checkpoint loudreader/loudr-1 \
  --name my-voice --language en

That writes voices/my-voice.safetensors and prints the speak command that reads it. The file is mode 0600 on POSIX.

The command reads one local WAV or FLAC file. It does not record, fetch a URL, denoise or batch. The input is checked before any model runs: 5 to 10 seconds of one speaker is right, more than 30 seconds is refused, and so are silence, NaN or Inf samples, and a clip under one second.

By default, the command cuts at the last suitable pause in the first ten seconds and appends 0.4 seconds of silence. A pause near the ten-second limit leaves only part of that silence inside the prompt. If no pause fits, the command cuts at 9.6 seconds when needed to make room for silence. This can help soften speech onsets, but the fallback can cut through a word. Use --no-end-in-silence to enroll the recording as given.

flag	what it does
--checkpoint	a repo id, a release directory or a checkpoint file. Must be a cloning-capable release.
--name	what to call the voice. Carried in the profile, and the default filename.
--language	the language the voice speaks. Stated, never guessed.
-o, --output	where to write. Default voices/<name>.safetensors.
--revision	commit, tag or branch to pin --checkpoint to.
--device	enrollment runtime: cpu, cuda, cuda:<index>, mps, onnx or coreml. Default cpu.
--no-end-in-silence	keep the recording as given, without cutting and padding its ending.
--force	overwrite the output. Without it an existing file is left alone and the command exits 1.

Name the language. A Polish voice cloned without --language pl reads its text through the English rules:

loudkit clone nagranie.wav --checkpoint loudreader/loudr-1 \
  --name gosia --language pl

From Python

One call. It takes a recording and the same reference lk.load takes, and returns the profile.

import loudkit as lk

mine = lk.enroll("my-recording.wav", "loudreader/loudr-1", name="my-voice")
mine.save("voices/my-voice.safetensors")  # ~150 KB

The prompt is built from the first ten seconds and the speaker embedding reads the whole clip, which is why anything over 30 seconds is refused rather than cut: trim the recording to its best 5 to 10 seconds yourself.

Python leaves the recording as given by default. Pass end_in_silence=True to use the shell command’s pause selection and silence padding. The original recording must pass the input checks before it is changed.

Name the language when the voice is not English:

mine = lk.enroll("nagranie.wav", "loudreader/loudr-1", name="my-voice", language="pl")

VoiceProfile.language is what the engine uses when a call names no language, so a Polish voice enrolled without this reads its text through the English rules.

The other arguments: device places the enrollment models (cpu is enough); revision pins the release, as on lk.load. Reading an audio file uses soundfile, which the enroll extra brings; pass mono samples in [-1, 1] instead and no reader is involved.

Clone without PyTorch

Install the runtime and audio reader without the torch-based enroll extra:

pip install "loudkit[onnx,audio,hub]"
# On a Mac, for CoreML instead: pip install "loudkit[coreml,audio,hub]"

Use device="onnx" or device="coreml" with a complete release that includes enrollment graphs. The output is the same portable profile format:

mine = lk.enroll("nagranie.wav", "loudreader/loudr-1-turbo",
                 name="my-voice", language="pl", device="onnx")
mine.save("voices/my-voice.safetensors")

Both models use the canonical enrollment assets. You do not need to clone again when changing the synthesis model.

Use it

import loudkit as lk

voice = lk.voice("voices/my-voice.safetensors")
engine = lk.load("loudreader/loudr-1")
engine.synthesize("Now it is my voice speaking.", voice, seed=1).save("mine.wav")

What a voice carries

A VoiceProfile holds the speaker embedding for the token generator, the conditioning tensors for the renderer, and the reference-prompt tokens and mel the enrollment produced. A voice does not change when the machine or the backend changes, so the file travels, and it works with both models.

Good input, honest limits

Five to ten seconds of clean, single-speaker, noise-free audio is enough. The hard limits: more than 30 seconds, silence, non-finite samples and clips under one second are refused, with a message that says what a good input looks like. The best ten seconds is rarely the first ten you try; how the shipped voices were chosen, and how to check a new profile before shipping it, is in choosing a reference.

Consent is yours to obtain. Do not clone a voice you have no permission to use. See RESPONSIBLE_USE.md. The shipped voices are enrollments of donated or openly licensed recordings named per voice in VOICES.md, not clones of private individuals.

Next

Server and agents: a warm server and an MCP tool for any agent on the machine.