voice — a speech toolkit in Rust

Recognition, synthesis and voice conversion, each one a Unix filter, so they compose in a pipe:

voice -(stt)-> text -(translate)-> text -(tts)-> voice -(rvc)-> voice

Every engine reads raw mono f32le PCM or plain text on stdin and writes the same on stdout, with logs on stderr — so any subset is a valid pipeline.

No Python. Models are ported to Burn and load their original Hugging Face weights directly; training is native Rust too. One exception stays on purpose: the .safetensors → ONNX exporter under export/.

One binary per engine, plus one that prepares corpora and one that has them all

Running an engine's binary with no subcommand is the filter.

Install

export LIBTORCH=$PWD/libtorch      # 2.9.0; omit to build without the tch backend
cargo build --release              # all six binaries
cargo build --release -p stt-cli   # or just one

ONNX Runtime, LibTorch and ffmpeg are found the same way by all six binaries:

→ docs/setup.md

Engines

Not an engine, but the thing you run before the two that train:

Each binary's README has the rest — every flag, which weights are fetched and from where, and its fine-tuning loop.

rvc and seedvc do the same job on opposite terms: rvc learns one voice from a corpus and renders it very well, seedvc takes any voice from a 1–30 s clip and renders it adequately with nothing trained at all.

--backend is a run-time choice on every engine: rvc, stt and tts run under ONNX Runtime or under native Burn on LibTorch, CubeCL/CUDA or WebGPU, in one binary, with one set of spellings; seedvc also runs on ONNX Runtime, from a bundle of six graphs written by export/export_seedvc.py. Fine-tuning is always Burn.

Documentation

Usage lives with the tool and this file stays an index. Beyond the five binary READMEs and docs/setup.md:

The architecture papers are written for reviewing and maintaining a port rather than for driving the CLI. Four of them take one network apart; the preprocess one covers a phase instead — two networks and six stages that run none — so it is also where the corpus-preparation measurements live. The convention is a Typst source with its rendered PDF committed beside it, so reading one needs no toolchain; rebuild with typst compile docs/<name>.typ and commit both.

Crate layout

Three tiers, and the names say which is which.

Shared plumbing — no model, no engine, safe for anything to depend on:

Networks, named after the model, holding no app dependencies:

One -core and one -cli per binary, all the same shape:

Two rules keep it that way. No engine depends on another engine — anything two of them need moves into the plumbing tier first. And voice- is reserved for the top: a crate an engine depends on may not be named after the integration, which is why the shared crates are *-kit rather than voice-*.

Contributing

cargo clippy --workspace is kept clean and cargo fmt is enforced. Read CLAUDE.md first — it holds the constraints that are easy to break silently, weight compatibility and backend genericity among them.

Licence

Copyright (C) 2026 王翼翔.

This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License, version 3, as published by the Free Software Foundation. It is distributed in the hope that it will be useful, but without any warranty — without even the implied warranty of merchantability or fitness for a particular purpose. See LICENSE for the full text.

Releases up to and including the MIT-licensed ones stay MIT for whoever holds a copy of them; the change applies going forward. The move to copyleft is what porting Seed-VC requires — it is GPL-3.0, and a port carries that forward to everything linking it.