loudkit

Natural-sounding text-to-speech for products, scripts and experiments.

loudkit runs on your own hardware. It includes twenty voices across 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.

Hear all 20 voices | Open in Colab | Model | Documentation

• 20 voices in 10 languages: English, Spanish, French, German, Italian, Polish, Portuguese, Dutch, Swedish and Danish.
• Voice cloning: a reusable profile of about 150 KB from roughly ten seconds of audio.
• Runs locally: PyTorch, ONNX Runtime and CoreML.
• Five native SDKs: Python, Swift, Go, Rust and TypeScript.

Make a WAV

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

import loudkit as lk

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

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

From the shell:

loudkit speak --checkpoint loudreader/loudr-1 --voice joe \
  "Hello from loudkit." -o hello.wav

The first run downloads the 747 MB synthesis checkpoint and the voices. Later runs use the local cache. See Getting started for revision pinning, local directories and a complete first-run walkthrough.

Voices

Open the voice gallery to compare all twenty shipped voices. Each generated sample is shown next to the enrollment reference used to build its profile.

The voice roster records the source, licence and consent basis for every profile. The voices come from recordings donated for speech technology or from CC0 and CC-BY speech corpora. No scraped celebrity voices ship with the project.

We have evaluated English by ear. We do not speak the other nine languages well enough to judge their naturalness reliably. If you do, please listen and tell us what sounds good or wrong. Reports from native speakers are especially welcome.

Clone a voice

Use a recording that 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 --checkpoint loudreader/loudr-1 \
  --voice voices/my-voice.safetensors "Now in a cloned voice." -o cloned.wav

The result is a portable profile, not another copy of the model. Python users can call the same path through lk.enroll. See Cloning a voice for recording advice and Responsible use for the consent rules.

Download only the runtime you need

The Hugging Face repository contains every supported format. The downloader selects one runtime and leaves the rest behind.

Add --with-cloning if that installation also needs enrollment. The model card lists exact download sizes and files.

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. CPU performance depends heavily on the runtime: ONNX Runtime is faster than real time on the measured M3 Pro, while the PyTorch CPU reference path is not.

For batched workloads, the token generator reaches 20.1x aggregate throughput at batch 1 and 153.1x at batch 64 on the RTX 3090. The highest measured result is 170.8x on an A100 at batch 64. These are generator-only throughput numbers, not single-request latency or end-to-end RTF. Full commands, hardware and caveats are in Benchmarks.

SDKs and local APIs

Python is the reference implementation. Swift, Go, Rust and TypeScript are native ports checked against the same conformance fixtures.

For processes that should keep one engine warm, loudkit also includes:

• a local HTTP server with an OpenAI-compatible speech endpoint;
• a small MCP preview server;
• typed gRPC streaming with backpressure;
• a Speech Dispatcher module for Linux screen readers.

Servers and agents documents the exact contract of each integration.

Output and reproducibility

Synthesis results support chunk and estimated word timestamps, pitch-preserving speed from 0.5x to 2.0x, and C2PA Content Credentials. Saved WAVs and server responses carry an unsigned claim-only C2PA manifest by default. It records the model fingerprint, checkpoint and voice digests, seed, backend and audio digest.

For a fixed build, device and backend, the same text, voice and seed produce the same waveform. Different devices or backends can differ at the waveform level because of floating-point arithmetic. The precise guarantees and measurements live in the Identity contract and Measured parity.

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.

The project ships twenty permitted voice profiles and the code to enroll your own. It will not help with undisclosed impersonation, bypassing voice authentication or removing provenance from generated audio.

Documentation

• Getting started: first synthesis, caches and local files.
• Guides: long form, cloning, servers and every SDK.
• Model card: model lineage, limitations and release layout.
• Supported in v0.1: the public compatibility boundary.
• Troubleshooting: common failures and concrete fixes.
• Architecture: components and package layout for contributors.

Licence

The code and loudr-1 release are Apache-2.0. The tokenizer and speaker encoder retain their upstream MIT licence from Chatterbox. NOTICE lists every upstream component and licence. The voice encoder's chain is in its provenance record.