Model Environments¶

RVCBench integrates many third-party voice cloning and TTS models. These projects often pin incompatible versions of PyTorch, Transformers, ONNX Runtime, tokenizers, or model-specific helper packages, so one global Python environment is not expected to run every model.

The model environment specs are maintained in this repository under envs/. Use the environment that matches the model you are launching.

Creating an Environment¶

The Qwen environment installs only the core package and Qwen runtime. Other base templates install the core package and may require a separate upstream checkout or package. They are installation starting points, not tested lock files. Run from envs/ so editable package paths resolve correctly:

cd envs
conda env create -f qwen3-tts.yml
conda activate qwen3
cd ..

Then run the corresponding benchmark command, for example:

python scripts/run_qwen3tts_quickstart.py --max-samples 5

To update an existing environment:

cd envs
conda env update -f qwen3-tts.yml --prune
cd ..

Model-to-Environment Map¶

Model family	Environment file	Conda environment name
General benchmark / fallback	`envs/audiobench.yml`	`audiobench`
VoxCPM2	`envs/voxcpm.yml` (base recipe; requires the pinned upstream checkout)	`voxcpm`
FireRedTTS2	`envs/fireredtts2.yml` (native monologue canary validated; requires pinned upstream checkout)	`fireredtts2`
BertVITS2 / SafeSpeech surrogate	`envs/bertvits2.yml`	`bertvits2`
CosyVoice	`envs/cosyvoice.yml`	`cosyvoice`
dots.tts	`envs/dots-tts.yml`	`dots-tts`
Fish Audio S2 native	`envs/fish-s2-native.yml` (setup; base recipe)	`fish-s2-native`
F5-TTS	`envs/f5-tts.yml` (validated Python 3.11 / Torch 2.8 / SDK 1.1.18 pins; base recipe)	`f5-tts`
Fish Audio S2 (local API server)	`envs/fish-speech-s2.yml` (base env; serves the model, see README)	`fish-speech-s2`
FishSpeech	`envs/fishspeech.yml`	`fishspeech`
GLM-TTS	`envs/glm-tts.yml`	`glm-tts`
GlowTTS	`envs/glowtts.yml`	`glowtts`
Higgs Audio	`envs/higgs-audio.yml`	`higgs-audio`
Higgs TTS 3	`envs/vllm-omni-cu129.yml` (base env; serves the model, see README)	`vllm-omni-cu129`
IndexTTS	`envs/indextts.yml`	`indextts`
Kimi Audio	`envs/kimi-audio.yml`	`kimi-audio`
MaskGCT	`envs/maskgct.yml`	`maskgct`
MGM-Omni	`envs/mgm-omni.yml`	`mgm-omni`
MOSS-TTS	`envs/moss-tts.yml`	`moss-tts`
MOSS-TTSD (native MossTTSD checkpoint)	`envs/moss-ttsd.yml`	`moss-ttsd`
MOSS-TTSD (legacy Asteroid checkpoint)	`envs/moss.yml`	`moss`
OpenVoice	`envs/openvoice.yml`	`openvoice`
OZSpeech	`envs/ozspeech.yml`	`ozspeech`
PlayDiffusion	`envs/playdiffusion.yml`	`playdiffusion`
Qwen3-Omni	`envs/qwen3-omni.yml`	`qwen3-omni`
Qwen3-TTS	`envs/qwen3-tts.yml`	`qwen3`
Spark-TTS	`envs/sparktts.yml`	`sparktts`
StyleTTS2	`envs/styletts2.yml`	`styletts2`
VALL-E	`envs/vall-e.yml`	`vall-e`
Bark Voice Clone	`envs/bark-native.yml` (native LibriTTS16 generated/scored; base recipe; setup)	`bark-native`
Amphion VALL-E (separate implementation)	`envs/amphion-valle.yml` (native LibriTTS16 generated/scored; base recipe)	`amphion-valle`
VibeVoice	`envs/vibevoice.yml`	`vibevoice`
XTTS-v2	`envs/xtts-v2.yml`	`xtts-v2`
ZipVoice	`envs/zipvoice.yml`	`zipvoice`
ZONOS2	not a conda env — see `envs/zonos2.yml` (`uv`-managed `checkpoints/ZONOS2-repo/.venv`)	—

Reproducibility Options¶

The Qwen quickstart and CPU smoke path are covered by offline CI. The other model families have adapter integrations and base environment templates; they have not all been revalidated with current upstream packages. Dependencies are declared in pyproject.toml; follow each model runtime setup before using its adapter.

Generation-only usage needs the core plus model runtime. Install .[eval] explicitly for full evaluation; run rvcbench doctor --model qwen3 --eval --imports to check dependency availability before launching. This check does not load models or certify CUDA compatibility. Training/protection runtimes need their own upstream dependencies.

For stronger reproducibility, generate platform-specific lock files from these YAML files with conda-lock, or publish prebuilt Docker/Apptainer images per model family. Containers are usually the most reliable option for CUDA-heavy third-party inference stacks, while Conda or Micromamba specs are easier for users who need to adapt paths, CUDA versions, or local checkpoint locations.

For models requiring a different Torch version, generate with +vc.generate_only=true in that model's environment and evaluate in a separate environment containing .[eval]. The Qwen, Enkidu and evaluation extras accept PyTorch 2.3–2.9 and keep an existing build; envs/qwen3-tts.yml and envs/enkidu.yml pin the exact versions used for the published runs. Other model base templates do not install the evaluation stack automatically.

Generation provenance records statically discovered unmapped_imports separately from distribution versions. These names may be authored upstream namespaces, missing modules or packages without distribution metadata. The inventory includes literal __import__() and import_module() calls; it does not resolve nonliteral imports or certify a complete environment lock.

For SparkTTS, run rvcbench doctor --model sparktts --imports in the chosen runtime before generation. It checks the main native dependencies, including einx, without loading weights. Dependency availability does not validate the upstream source tree, checkpoint compatibility, CUDA or full model inference.

OpenVoice's converter-only CPU check uses native configuration and weights, but its full audio API also needs a compatible NumPy/Numba/Librosa combination. Import the audio stack in the intended runtime before loading checkpoints. Read-only environment installations may require NUMBA_CACHE_DIR pointing to a writable local directory. This addresses cache placement; it does not repair NumPy/Numba version incompatibility or establish a clean OpenVoice/Melo lock.

Building the evaluation extras from source¶

pysptk (needed for MCD) is published only as a source package. Its build appends the commit of any Git repository that contains the build directory to its version number, which makes pip discard the build. If pip install -e '.[eval]' reports that no pysptk version matches, build it outside a Git checkout or pin the version explicitly:

PYSPTK_BUILD_VERSION=1.0.1 python -m pip install 'pysptk==1.0.1'

Where scorer models are stored¶

The speaker, emotion and DNSMOS scorers keep their model files in one directory per scorer: $RVCBENCH_ASSET_DIR/<name> when that variable is set, otherwise ./checkpoints/<name> when it exists (as in a source checkout), otherwise ~/.cache/rvcbench/<name>. A speaker model fetched from the Hugging Face Hub is copied there rather than linked into the Hub cache, so moving or clearing that cache does not break it.