loudkit

Text-to-speech that runs on your own hardware.

loudkit has 28 voices in ten languages, clones a voice from about ten seconds of audio, and has native SDKs for Python, Swift, Go, Rust and TypeScript. After the first model download it runs offline. It needs no account, sends no telemetry and has no usage fees.

LoudReader, a reading app, uses loudkit as its speech engine. loudkit is licensed under Apache-2.0.

Hear the voices | Try the online demo | Open in Colab | Model | Documentation

Quickstart

pip install "loudkit[torch,audio,hub]"
loudkit speak --voice joe "Hello from loudkit." --play

The first run downloads the 747 MB model and the 28 voices. After that, it runs offline. speak writes out.wav by default, and --play plays it with the system player. Use -o hello.wav to choose the file name. Other English voices work the same way, for example --voice kathleen or --voice henry. A voice reads text in its own language unless you pass --language.

Python

import loudkit as lk

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

engine.synthesize("Hello from loudkit.", voice, seed=7).save("hello.wav")

engine.voices() lists the 28 names. All 28 voices are in both model downloads and load by name. synthesize splits long text into chunks and returns one waveform. engine.stream(...) yields the same audio chunk by chunk, so playback can start early.

Read Getting started next.

Swift, Go, Rust and TypeScript

Each port loads the model by name. The first call downloads and verifies it, and later calls run offline. The snippets need version 0.1.1 of each package.

Swift (CoreML, macOS 14 or iOS 17). Add .package(url: "https://github.com/loudreader/loudkit", from: "0.1.1") to Package.swift. Guide.

import LoudKit

let engine = try await Engine.load("loudreader/loudr-1")
let voice = try engine.voice(named: "joe")
try engine.synthesize("Hello from loudkit.", voice: voice, seed: 7).saveWav("hello.wav")

Go (ONNX Runtime). go get github.com/loudreader/loudkit/go@v0.1.1. Guide.

eng, err := loudkit.Load("loudreader/loudr-1")
if err != nil { log.Fatal(err) }
defer eng.Close()
v, err := eng.Voice("joe")
if err != nil { log.Fatal(err) }
res, err := eng.Synthesize("Hello from loudkit.", v, loudkit.Options{Seed: 7})
if err != nil { log.Fatal(err) }
if err := res.SaveWav("hello.wav"); err != nil { log.Fatal(err) }

Rust (ONNX Runtime). cargo add loudkit@0.1.1. Guide.

use loudkit::{Engine, Options};

let mut engine = Engine::load("loudreader/loudr-1")?;
let voice = engine.voice("joe")?;
let options = Options { seed: 7, ..Default::default() };
engine.synthesize("Hello from loudkit.", &voice, &options)?.save_wav("hello.wav")?;

TypeScript (ONNX Runtime, Node 20). npm install loudkit@0.1.1. Guide.

import { Engine } from "loudkit";

const engine = await Engine.load("loudreader/loudr-1");
const voice = engine.voice("joe");
(await engine.synthesize("Hello from loudkit.", voice, { seed: 7 })).saveWav("hello.wav");
await engine.close();

Go and Rust need libonnxruntime on the machine (brew install onnxruntime, or a build from the ONNX Runtime releases); the guides say where each port looks for it. All five implementations pass the same conformance fixture. The identity contract states where their speech tokens match and where they can differ.

Two models

loudreader/loudr-1 is the default. loudreader/loudr-1-turbo is faster and sounds slightly less natural. Both run on every backend and in all five SDKs. Select the model with the string you pass to load. See Choosing a model.

Voices

The voice gallery plays all 28 voices next to the recording each was enrolled from. The ten languages are English, Spanish, French, German, Italian, Polish, Portuguese, Dutch, Swedish and Danish. English has ten voices, and each other language has two. VOICES.md records the source, licence and consent basis of each voice. The voices come from recordings donated for speech technology, or from CC0 and CC-BY corpora.

Only English has been evaluated by ear. The other nine languages have automated checks but no native-speaker review. If you speak one of them, please listen and report what sounds wrong.

Clone a voice

Use five to ten seconds of one speaker, from a recording you own or have permission to use:

pip install "loudkit[torch,audio,enroll,hub]"
loudkit clone my-recording.wav --checkpoint loudreader/loudr-1 --name my-voice --language en
loudkit speak --voice voices/my-voice.safetensors "Now in a cloned voice." -o cloned.wav

The result is a portable voice profile of about 150 KB. By default, loudkit clone cuts the recording at a pause and pads its end with silence. lk.enroll(...) in Python and enroll in each port keep the recording as given. To get the loudkit clone result from Python, pass end_in_silence=True to lk.enroll. See Cloning a voice and Responsible use.

Measured speed

path	hardware	loudr-1	loudr-1-turbo
split PyTorch engine*	Apple M3 Pro	3.29x	5.77x
Swift, native generator plus CoreML renderer	Apple M3 Pro	2.49x	3.44x
ONNX Runtime, CPU provider	Apple M3 Pro	1.14x	1.59x
PyTorch with CUDA graphs	RTX 3090	8.55x	13.05x
PyTorch with CUDA graphs	Jetson Orin Nano Super, 25 W	1.85x	2.50x

* "Split" is a device placement of the same model. The token generator runs on the CPU, and the renderer (mel decoder and vocoder) runs on the Apple GPU through MPS. Adjacent windows can overlap across the two devices.

Higher is faster, and 1.0x is real time. Every row was measured on 0.1.1 on 2026-09-06, for both models, with the same passage, voice and seed. The Apple rows come from one M3 Pro laptop in ordinary use, with no isolation from background load. The A100, L4 and T4 rows are on the benchmark page.

For batched workloads, the token generator reaches 16.7x aggregate throughput at batch 1 and 57.3x at batch 64 on the RTX 3090 with loudr-1, and 42.0x to 155.0x with loudr-1-turbo. The highest measured result is 223.6x, turbo on an A100 at batch 64 (85.3x with loudr-1), measured on 0.1.1. These figures come from research/bench_batch.py, which replays a recorded token sequence through the token generator. Sampling, prefill and rendering are outside the timed loop, so the figures are neither single-request latency nor end-to-end RTF. Full commands, hardware and caveats are in Benchmarks.

Integrations

These keep one loaded engine across many requests. The contracts are in Server and agents.

• loudkit serve: an HTTP server with loudkit's own routes, Server-Sent Events streaming on /v1/synthesize/stream, and an OpenAI-compatible /v1/audio/speech that returns the whole utterance in one response. Agents that use OpenAI's speech API, such as Hermes Agent and OpenClaw, connect with configuration only.
• loudkit serve --grpc: the same engine behind a typed schema with backpressure.
• loudkit serve --mcp: an MCP server on stdio for agent hosts (preview).
• A Speech Dispatcher module for Linux screen readers, in integrations/.
• Docker images and compose for the server.

By default, a WAV saved from Python and a WAV reply from the server carry an unsigned, machine-readable note on how the audio was made. The note names the model, voice, seed and backend, and holds a checksum of the audio. loudkit verify reads it. The Go, Rust, JS and Swift ports write plain PCM without the note. The note is not C2PA: nothing signs it, and C2PA tools do not read it.

Scope

loudkit is an inference toolkit. It does not provide accounts, billing, multi-tenancy, model training or emotion control. For a public deployment, put your own authentication, rate limits and TLS in front of the server. The maintainers close issues and pull requests that ask for help with undisclosed impersonation, bypassing voice authentication, or stripping the note from generated audio. SUPPORTED.md lists what is supported.

Documentation

The documentation index lists the user guides: getting started (Python), one guide per port, choosing a model, cloning, long text and streaming, servers, troubleshooting and the two model cards. It also links the reference, platform and design pages. The model card covers lineage, limits and what each download contains.

Licence

The code and the loudr-1 and loudr-1-turbo releases are Apache-2.0. The tokenizer and speaker encoder keep their upstream MIT licence from Chatterbox. NOTICE lists every upstream component and licence.