Text to speech that runs on your machine.

Hear it

Twenty voices, ten languages.

English, Spanish, French, German, Italian, Portuguese, Polish, Dutch, Swedish and Danish. Every voice is enrolled from a consented donation or a CC0 / CC-BY corpus, with the source named.

Hear two now.

Both clips read the same sentence from Alice’s Adventures in Wonderland.

joeEnglish · male

kathleenEnglish · female

One seeded sample per voice, playable in the browser. The sources and the licences are on the same page. Open the demo to hear each enrollment recording next to the speech it produced, and the other eighteen voices.

We have evaluated English by ear. We do not speak the other nine languages well enough to judge them reliably, so feedback from native speakers is very welcome.

Where it runs

Laptops, GPUs and edge devices.

Faster than real time on datacentre GPUs, on desktop GPUs, on Apple Silicon and on an embedded board. Every figure is one loudkit bench run on the named machine. The hardware, the method and the commands.

requires an NVIDIA GPU

NVIDIA CUDA

7.47×real time, RTX 3090 with CUDA graphs

PyTorch, from an embedded board to a datacentre card.

Host	Eager	CUDA graphs
A100	2.16×	8.12×
L4	2.20×	7.56×
RTX 3090	3.39×	7.47×
T4	2.10×	5.46×
GTX 1080 Ti	1.95×	1.89× *
Jetson Orin	0.73×	1.83×

 * Pascal cannot capture graphs: on the 1080 Ti the flag runs eager.

Batch throughput, RTX 3090 with CUDA graphs

Many texts at once. Each number is the total audio produced per second of wall clock across the whole batch, not the speed of one synthesis.

20.1×batch 1

80.8×batch 8

110.8×batch 16

137.3×batch 32

153.1×batch 64

What it measures, and the full table. The highest measured aggregate is 170.8× on an A100 at batch 64.

requires Apple Silicon

Apple GPU

3.43×real time, M-series laptop, --device mps

The generator runs on the CPU and the renderer on the GPU. A CoreML renderer ships for Swift.

requires no PyTorch

ONNX Runtime

1.21×real time, same laptop, fp32 graphs

Runs without PyTorch, on ONNX Runtime.

an embedded NVIDIA GPU

Jetson Orin Nano

1.83×real time, --device cuda --cuda-graphs, measured on the board

The same engine, the same conformance fixture, on a board with no discrete GPU.

Every figure is a multiple of real time. 3× real time means one minute of audio in twenty seconds. Under 1× is slower than real time.

Five languages, one engine

Ports, not wrappers.

Python is the reference implementation. Swift, Go, Rust and TypeScript are full ports of it, and all five are held to the same conformance fixture, so the same text and the same voice give the same speech tokens in every one.

PythonPyTorch, ONNX Runtime or CoreML backendreference

SwiftCoreML backendSwift Package Manager

GoONNX Runtime backendgo get

RustONNX Runtime backendcargo

TypeScriptONNX Runtime backendnpm i

Install

Install, then speak.

Add the package, load a checkpoint, pick a voice, get a waveform. Python fetches what it needs on the first call. The ONNX ports read files from disk, so they fetch first into loudr-1/, which is what the paths below are relative to.

Python

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

import loudkit as lk

engine = lk.load("loudreader/loudr-1")
voice = lk.voice("kathleen", repo="loudreader/loudr-1")
engine.synthesize("Hello from loudkit.", voice, seed=7).save("hello.wav")

TypeScript

npm i loudkit

pip install "loudkit[hub]"
loudkit download loudreader/loudr-1 --for onnx --local-dir loudr-1

import { Engine, loadVoice } from "loudkit";

const engine = await Engine.load(
  "loudr-1/loudr-1.safetensors", "loudr-1/onnx", "loudr-1/tokenizer.json");
const voice = loadVoice("loudr-1/voices/kathleen.safetensors");

const result = await engine.synthesize("Hello from loudkit.", voice, 7);
// result.audio is a Float32Array at result.sampleRate.

Go

go get github.com/loudreader/loudkit/go

pip install "loudkit[hub]"
loudkit download loudreader/loudr-1 --for onnx --local-dir loudr-1

// Point onnxruntime_go at libonnxruntime first. See the Go guide.
eng, _ := engine.Load(
    "loudr-1/loudr-1.safetensors", "loudr-1/onnx", "loudr-1/tokenizer.json")
defer eng.Close()
v, _ := voice.Load("loudr-1/voices/kathleen.safetensors")

// After the seed: language, speed, previousTokens, shouldCancel.
audio, _, _, _, sr, _, _ := eng.SynthesizeLong(
    "Hello from loudkit.", v, 7, "", 1.0, nil, nil)

Rust

loudkit = "0.1"

pip install "loudkit[hub]"
loudkit download loudreader/loudr-1 --for onnx --local-dir loudr-1

let mut eng = Engine::load(
    "loudr-1/loudr-1.safetensors", "loudr-1/onnx", "loudr-1/tokenizer.json")?;
let v = voice::load("loudr-1/voices/kathleen.safetensors")?;

// After the seed: language, speed, previous_tokens, should_cancel.
let (audio, tokens, _, sr, _, _) =
    eng.synthesize_long("Hello from loudkit.", &v, 7, None, 1.0, None, None)?;

Swift

// Swift Package Manager
.package(url: "https://github.com/loudreader/loudkit", from: "0.1.0")

pip install "loudkit[hub]"
loudkit download loudreader/loudr-1 --for coreml --local-dir loudr-1

import Foundation
import LoudKit

let engine = try Engine.load(
    checkpoint: URL(fileURLWithPath: "loudr-1/loudr-1.safetensors"))
let voice = try VoiceProfile.load(
    url: URL(fileURLWithPath: "loudr-1/voices/kathleen.safetensors"))

let result = try engine.synthesize("Hello from loudkit.", voice: voice, seed: 7)
try result.save(to: URL(fileURLWithPath: "hello.wav"))

The same seed and voice give the same speech tokens in all five. The four ports load graphs exported once from a Python checkout: ONNX for TypeScript, Go and Rust, CoreML for Swift. Each guide has that step.

Where to go next

What it promises.

Documentation indexThe same index the repository shows: guides, reference, platforms and the model.What 0.1 supportsWhat is supported, what is experimental, and what 0.1 does not attempt.Identity contractSame text, same voice, same seed. What that guarantees, and where it stops.BenchmarksMeasured figures, the hardware, and the command that reproduces each row.Model cardWhat loudr-1 is, where it came from, its limits and what each download contains.DemoEnrollment reference and generated speech side by side, plus the twenty voices.

Responsible use.

The tool clones a voice from ten seconds of audio and includes C2PA Content Credentials in saved WAVs by default. Consent is yours to obtain. Read the terms.

LoudReader, the iOS reading app, runs on this engine. loudkit is the engine it ships.