VoiceStudio has an optional native CosyVoice 3 backend for machines where the
official Python/PyTorch route is not a good fit. It runs the community
Lourdle/cosyvoice.cpp GGML port as
a local server. The tested AMD route is a Radeon RX 6700 XT using Vulkan. This
confirms that card and build; it is not a guarantee for every AMD GPU.
This is a separate engine named CosyVoice 3 GGUF. The existing Python CosyVoice 3 engine and VoiceStudio default do not change. It requires a reference recording and its transcript, supports CosyVoice 3's nine languages, and passes free-form instructions such as “soft, warm, and breathy” to the model. The native route also accepts speaking-speed and seed controls. It returns 24 kHz speech. It does not provide text-only voice design or promise ElevenLabs feature parity. Quantized output may sound muffled on some lines; review the upstream model card and listen to the actual renders.
The executable is not bundled or downloaded by VoiceStudio. Build the pinned community source yourself and retain its notices and all GGML/Vulkan/ONNX Runtime notices. On Windows, install Git, CMake, Visual Studio C++ build tools, and the Vulkan SDK, then from a Developer PowerShell:
git clone --recursive https://github.com/Lourdle/cosyvoice.cpp.git
cd cosyvoice.cpp
git checkout 67d45feb2f654bbc66cca25bff78464b3ec08fa1
git submodule update --init --recursive
cmake -S . -B build -DGGML_VULKAN=ON -DCOSYVOICE_NO_ICU=ON -DCOSYVOICE_SERVER_NO_WEBUI=ON
cmake --build build --config Release --parallelUse the build output containing cosyvoice-server.exe and cosyvoice-cli.exe.
Set OMNIVOICE_COSYVOICE_GGUF_DIR to that output folder, or set the individual
OMNIVOICE_COSYVOICE_GGUF_SERVER and OMNIVOICE_COSYVOICE_GGUF_CLI variables.
The engine will report which native devices the runtime actually exposes. The
folder may be the CMake output itself (build\bin\Release); the engine also
looks in its bin and Release subfolders.
The engine list names the kind of problem: the native runtime is not set up or
will not start, or its model files are missing. The exact cause (which
folder or file was tried, the runtime's exit code or timeout, and the last
lines of its output with private paths removed) is written once to the backend
log as CosyVoice GGUF unavailable: ...; paste that line in a bug report.
If the runtime's device probe (cosyvoice-cli --list-backends, up to 120 s on
the first run) fails or crashes without listing a device, auto uses the
runtime's CPU path and says so in the routing reason. When the probe succeeds
but lists no Vulkan device, the reason names the runtime's exit code and its
own Vulkan message (for example ggml_vulkan: No devices found.). Set
OMNIVOICE_COSYVOICE_GGUF_BACKEND=CPU to skip the probe entirely. An explicit
Vulkan0 never falls back silently.
Open Settings → Models and install CosyVoice 3 0.5B Q4_K_M (native
GGUF/Vulkan). The entry is pinned to an immutable model revision and installs
the GGUF and two ONNX frontend assets. Generation never downloads a model or
runtime. Existing local model files can instead be selected with
OMNIVOICE_COSYVOICE_GGUF_MODEL and OMNIVOICE_COSYVOICE_GGUF_FRONTEND. A
set override must point at real files: a missing model or frontend file is an
error, never a silent fallback to the installed assets.
The backend selector is OMNIVOICE_COSYVOICE_GGUF_BACKEND:
| Value | Behavior |
|---|---|
auto (default) |
Select a Vulkan device reported by the native runtime; otherwise use CPU and disclose the fallback. |
Vulkan0 |
Require that exact detected device. Missing Vulkan is an error, never a silent CPU fallback. |
CPU |
Force CPU execution. |
For the tested AMD workstation, set OMNIVOICE_COSYVOICE_GGUF_BACKEND=Vulkan0
and confirm the engine details list Vulkan0 with AMD Radeon RX 6700 XT.
Voice prompts are cached locally by reference-audio content plus transcript;
requests and profile changes are serialized so a render cannot use another
voice's prompt. The server binds to loopback, uses a per-process API key, and is
stopped when the engine unloads or stops responding; a request the server
rejects (an HTTP error) is reported with its status and leaves it running.
The pinned runtime accepts its API key only as a command-line argument, so other accounts on the same computer that can list process arguments can read it while the server runs. The key is regenerated on every launch and the port accepts loopback connections only; on a shared multi-user machine, prefer another engine.
The upstream model card reports Apache-2.0. VoiceStudio records the model as unreviewed: component provenance, commercial inference, output terms, and redistribution obligations have not been cleared. The runtime is an unofficial community port; retain its MIT and bundled third-party notices. These notices do not grant rights to cloned reference voices. Use only recordings you have permission to process. See the model inventory and upstream model card.