loudkit documentation

Twelve pages cover using loudkit. Everything else is in reference/ for someone who has already shipped, platforms/ for one machine, and design/ for anyone changing the engine.

Open the voice gallery to search, listen, compare English voices from both models. Voices explains which profiles are included and records their sources and licences.

Using loudkit

1. Getting started: Python, first WAV, voices, the seed, devices.
2. Swift: the same engine over CoreML.
3. Go: over ONNX Runtime.
4. Rust: over ONNX Runtime.
5. JavaScript and TypeScript: over onnxruntime-node.
6. Choosing a model: loudr-1 or loudr-1-turbo, and where each one runs.
7. Cloning a voice: a profile of your own from ten seconds of audio.
8. Long text and streaming: first audio before the passage finishes.
9. Server and agents: HTTP, gRPC, MCP and Speech Dispatcher over one warm engine.
10. Troubleshooting: symptoms, causes, fixes.
11. Model card: loudr-1.
12. Turbo model card: loudr-1-turbo.

Beside them: Voices, What 0.1 supports and Responsible use.

Reference

• Compatibility: what may change between releases, and how to pin one.
• Errors: what each implementation raises.
• Timestamps and speed: what a result carries and how playback speed works.
• Content Credentials: what a saved WAV records.
• Identity contract: what “same input, same audio” means across backends, and what it does not.
• Voice encoder licence chain.

Platforms

Apple (CoreML and the Swift package), Docker, Jetson.

Performance

Benchmarks: the measured figures, the machines and the commands. Measured parity: the cross-runtime report.

Design

design/ holds the notes for anyone changing the engine: the architecture map, text normalization, postprocess, the ONNX graphs, typing, embedding, the benchmark tools, silence classes, two-token decode and the evaluation method. None of it is needed to use loudkit.