loudkit

Natural-sounding text-to-speech that runs on your own hardware.

Twenty-eight voices in ten languages, voice cloning from about ten seconds of audio, and native SDKs for Python, Swift, Go, Rust and TypeScript. Download the model once and run offline, with no account, telemetry or usage bill.

loudkit is the speech engine inside LoudReader, a reading app that speaks articles, PDFs and books on device. The engine is here under Apache-2.0 for anyone who wants to build with it directly.

Hear the voices | Try it in the browser | Open in Colab | Model | Documentation

Hear a voice in one command

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, then it runs offline. --play uses the system player; add -o hello.wav to keep the file. Every other voice works the same way: --voice kathleen, --voice dave.

The same from 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. synthesize takes text of any length, and engine.stream(...) delivers the same audio sentence by sentence so playback can start early. All 28 voices, including Henry, Oliver, Oscar and Sophie, are included in both model downloads and load by name.

Getting started is the page to read next.

The same in Swift, Go, Rust and TypeScript

Every port loads the model by name, fetches and verifies it on the first call, and runs offline after that. 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. The same text, voice and seed give the same speech tokens in all five languages.

Two models

loudreader/loudr-1 is the default and runs on every backend and in every language. loudreader/loudr-1-turbo is faster, at a small cost in naturalness, and in 0.1.1 runs in all five SDKs. The string passed to load is the whole choice: Choosing a model.

Voices

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

We have evaluated English by ear and do not speak the other nine languages well enough to judge them. If you do, please listen and tell us what sounds wrong.

Clone a voice

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

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 profile of about 150 KB, not another copy of the model. In Python the same path is lk.enroll(...); every port has enroll. See Cloning a voice and Responsible use.

Measured speed

* "Split" describes device placement, not a different model or checkpoint. The token generator runs on the CPU while the mel and vocoder renderer runs on the Apple GPU through MPS. Adjacent windows can overlap across the two devices.

Higher is faster, and 1.0x means real time. Every row was measured on 0.1.1 (2026-09-06, both models): the Apple rows on one laptop in ordinary use, the NVIDIA rows on the named parts, with the same passage, voice and seed. 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. They are generator-only throughput numbers, not single-request latency or end-to-end RTF. Full commands, hardware and caveats are in Benchmarks.

Integrations

For a process that keeps one engine warm rather than paying the load cost per call. The contracts are in Server and agents.

• loudkit serve: an HTTP server with loudkit's own routes and an OpenAI-compatible /v1/audio/speech, streaming over Server-Sent Events. Agents that speak OpenAI's speech API, Hermes Agent and OpenClaw among them, connect with configuration alone.
• 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.

WAVs saved from Python and the server's replies carry a machine-readable note about how the audio was made, naming the model, voice, seed and backend and carrying a checksum that ties the note to the sound; the Go, Rust, JS and Swift ports write plain PCM. loudkit writes that note and loudkit verify reads it. It is not C2PA Content Credentials: nothing signs it and no C2PA tool reads it.

Scope

loudkit is an inference toolbox, not a hosted speech platform. It does not provide accounts, billing, multi-tenancy, model training or an emotion control. The local server expects you to provide any public-facing authentication, rate limits and TLS. It will not help with undisclosed impersonation, bypassing voice authentication or stripping Content Credentials from generated audio. SUPPORTED.md states the boundary.

Documentation

The documentation index lists the twelve pages a user needs: getting started, one page per language, choosing a model, cloning, long text and streaming, servers, troubleshooting and the two model cards. The model card covers lineage, limits and what each download contains.

Licence

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