loudkit

Why it's ranked here

Few local speech engines put this much of their contract in writing. A scope document separates supported, experimental and out-of-scope features, lists which runtime and hardware combinations were actually measured, and admits English is the only language checked by ear. The Python package declares itself Alpha, so the API is young. Against that: Apache 2.0 code and weights, 28 voices each with a recorded source and consent basis, and one engine reachable from a command line, an HTTP server, gRPC and a preview MCP server. For offline speech with reproducible output across several platforms, it is a serious option rather than a demo.

What's good

Determinism is designed, not hoped for. Each stage draws its own random stream derived from the caller's seed, so tuning one stage cannot change another stage's output, and warming the engine cannot shift a later render. The local server binds to loopback by default, refuses oversized request bodies before reading them, queues at most 32 requests, and will not bind publicly without a bearer token. A response header flags audio cut off at the length cap, so a severed sentence is not passed off as finished. Saved audio carries an unsigned note naming model, voice, seed and backend, and the docs say plainly it is not C2PA.

Tradeoffs

The first run downloads a 747 MB model, and the ONNX and CoreML releases run to roughly 2.4 to 3.1 GB. Go and Rust need the ONNX Runtime library installed separately. By the maintainers' own account, nine of the ten languages have not been judged by a native speaker. MCP is a preview whose tools may change shape, and it has no cloning tool. The server speaks plain HTTP with no TLS, quotas or multi-tenancy, so a public deployment needs a proxy in front. There is no emotion control, no fine-tuning, and no Android or WebAssembly target. Features land in Python first and reach the other ports later.

How to use it well

It fits apps that must speak text offline on a laptop, a Jetson board or an Apple device, and teams that need identical audio from identical input for tests or caching. Load the engine once and keep it, since reading the checkpoint takes seconds; that cost is why the server exists. Preview the text preparation step from the command line before blaming the model for a misread number. Pin a release revision for anything reproducible, and check the truncation header on every server reply. Agents that already call OpenAI's speech endpoint can point at the local server instead. It does not cover hosted multi-user speech, training or expressive control.

Observed

License

Apache-2.0 for code and the loudr-1 release; tokenizer and speaker encoder keep upstream MIT from Chatterbox

Declared maturity

Development Status :: 3 - Alpha (pyproject.toml)

Language

Python reference implementation, with ports in Swift, Go, Rust and TypeScript held to shared conformance fixtures

Python support

Requires Python 3.10 or newer; classifiers list 3.10 through 3.13

Packaging

PyPI package with extras: torch, audio, hub, onnx, coreml, enroll, server, mcp, grpc

Interfaces

CLI with eight commands; HTTP server with an OpenAI-compatible speech route and SSE streaming; gRPC; MCP over stdio (preview); Speech Dispatcher module

Backends

PyTorch (CPU, CUDA, MPS), ONNX Runtime, CoreML

Platform requirements

Swift needs macOS 14 or iOS 17; TypeScript targets Node 20; Go and Rust need libonnxruntime on the machine

Model size

747 MB synthesis checkpoint for loudr-1, fetched on first use

Voices

28 shipped voices across ten languages, each with source, licence and consent basis recorded

Server defaults

Binds loopback; 10 000 character text cap; queue of 32; non-loopback bind requires --allow-public and a bearer token; no built-in TLS

Out of scope

Model training and fine-tuning, emotion control, Android, WebAssembly, production multi-tenant serving