Getting started

The shortest path from nothing to sound, in Python. The other four languages have their own pages: Swift, Go, Rust, JavaScript.

Install

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

torch runs the model, audio writes WAV files, hub fetches releases by name.

Say something

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

The command saves a WAV and plays it through the system player. On Linux, install alsa-utils or pulseaudio-utils if neither player is available.

From Python:

import loudkit as lk

engine = lk.load("loudreader/loudr-1")
voice = engine.voice("joe")

result = engine.synthesize("Hello from loudkit.", voice, seed=7)
result.save("hello.wav")
print(result)  # Result(1.56s, 39 tokens, seed=7, RTF 1.0x)

Choose "loudreader/loudr-1-turbo" in lk.load to use the other model; the voice and synthesis calls stay the same. Available backends depend on the graphs included in the release. synthesize returns audio and never plays it.

The first call downloads the model (747 MB) and the 28 voices into the Hugging Face cache. Everything after that runs offline. The Swift, Go, Rust and JS ports keep a cache of their own, one directory for all four: ~/Library/Caches/loudkit/loudreader--loudr-1 on macOS, ~/.cache/loudkit/loudreader--loudr-1 on Linux, $LOUDKIT_CACHE moves it.

Load the engine once and keep it: reading the checkpoint takes a few seconds. synthesize takes text of any length and returns one Result. Its audio is a float32 array at result.sample_rate.

Voices

engine.voices() lists the 28 names. A voice carries the language it was enrolled in, so no language argument is needed:

loudkit speak --voice joe   "Hello from loudkit."     -o en.wav
loudkit speak --voice dave  "Hola desde loudkit."     -o es.wav
loudkit speak --voice henri "Bonjour depuis loudkit." -o fr.wav

The demo page plays every voice and VOICES.md names the source and licence of each. A voice file of your own loads with lk.voice("voices/mine.safetensors"); Cloning a voice makes one.

All 28 voices are included in both model downloads, including Henry, Oliver, Oscar and Sophie. No separate profile download is needed:

voice = engine.voice("henry")

The seed

The same text, voice and seed give the same WAV on the same machine and build. A different seed gives a different, equally valid reading.

Preview what will be spoken

loudkit text 'Dr. Smith paid $12.' --language en

Prepared speech goes to stdout. A word diff on stderr shows replacements and removals, such as the expansion of numbers and dates. This loads no model and downloads nothing. See the CLI reference.

For long text, engine.stream(...) yields each sentence as it is rendered: Long text and streaming.

Devices

lk.load picks the best device it finds: CUDA, then Apple silicon, then CPU. Name one to override it:

engine = lk.load("loudreader/loudr-1", device="cuda")  # or "cpu", "mps"
engine = lk.load("loudreader/loudr-1", device="onnx")  # ONNX Runtime, no torch

device="onnx" needs pip install "loudkit[onnx,audio,hub]" and fetches the exported graphs instead of the torch weights. loudkit doctor says what this machine can run; print(engine.describe()) says what the engine chose.

Fetching the model yourself

load fetched what it needed above. To fetch on purpose, or into a directory of your own:

loudkit download loudreader/loudr-1                                  # into the shared cache
loudkit download loudreader/loudr-1 --for onnx --local-dir loudr-1   # ONNX graphs, into ./loudr-1
loudkit voices loudreader/loudr-1                                    # the menu

A directory works everywhere a repo id does: lk.load("loudr-1"). To hold a production build to one exact release, see pinning a release.

Next

Long text and streaming, Cloning a voice, Choosing a model.

Download sizes for 0.1.1

Approximate decimal sizes for the release files; backend weights are included. Cloning adds enrollment assets only when requested. Both models ship separate synthesis and enrollment checkpoints.

Model	Torch	Torch + cloning	ONNX	ONNX + cloning	CoreML	CoreML + cloning
loudr-1	0.75 GB	1.28 GB	2.60 GB	3.13 GB	2.46 GB	2.99 GB
loudr-1-turbo	0.72 GB	1.25 GB	2.44 GB	2.97 GB	2.44 GB	2.97 GB