XB-BOX - 🌐 CosyVoice3 跨语言合成

The flagship trick of CosyVoice3 is cross-lingual synthesis: you feed it a reference clip of a voice in one language - English, Mandarin, whatever - and it speaks your target-language text in that same voice, retaining the timbre and cadence. XB_CosyVoice3_CrossLingual is XB_ToolBox's wrapper for exactly that: reference voice in, text in, AUDIO of that voice speaking your words out. It's the node that turns "one decent sample of a person" into "a local multi-language voice actor."

It's part of the CosyVoice3 suite bundled in the pack, alongside zero-shot cloning, instruct control, and dialog generation. If you've ever envied the API TTS voices, this is the local alternative the community has been running for a while now - CosyVoice clones have real reddit presence in the ComfyUI space, and multi-language is one of its calling cards.

How it works

The node takes the loaded model (a COSYVOICE_MODEL object from the pack's XB_CosyVoice3_ModelLoader), extracts speaker characteristics from reference_audio, and synthesizes text in the target language conditioned on that voice. The knobs that matter:

• target_language - auto (detect from the text) or explicit zh, en, ja, ko, de, es, fr, it, ru. auto is the default and usually right.
• speed - 0.5–2.0×, default 1.0. Small changes read as energy; big ones can get artifacty.
• seed - default 42; -1 for random. TTS models are famously seed-finitely - if you get a take with a weird pronunciation, bump the seed before rephrasing.
• text_frontend - on by default: normalizes text (numbers → words, etc.). Turn it off if you're feeding in CMU phonemes or special tags like <slow>, which the frontend would mangle.

Output: audio - the synthesized AUDIO, ready for a save/preview node or the rest of your pipeline.

The inputs that matter

• model - from XB_CosyVoice3_ModelLoader. There's no standalone model load here; wire the loader's output in.
• text - the words to speak, in the target language.
• reference_audio - the voice to clone. Keep it clean and short (the family recommendation is 3–10s).

Installing

In XB_ToolBox:

cd ComfyUI/custom_nodes
git clone https://github.com/WJLUOXIAO/XB_ToolBox.git

then restart, and install the pack's dependencies - this is the real install story for all CosyVoice3 nodes. requirements.txt pulls in transformers, librosa, soundfile, modelscope/huggingface_hub, conformer, omegaconf, x-transformers, diffusers, pyworld, openai-whisper and friends. The ModelLoader downloads the model on first use (default Fun-CosyVoice3-0.5B, from HuggingFace or ModelScope - pick ModelScope if HF is throttled in your region). First run will quietly download a few GB; don't panic at the wait.

Common issues

• Reference too long - the nodes validate duration (max ~30s). Crop with XB_CosyVoice3_AudioCrop first.
• Weird pronunciation on a great seed - try another seed. These models can nail or butcher a sentence on adjacent seeds.
• Numbers/text read wrong - keep text_frontend on unless you know you need phonemes.
• "Model not loaded" / install failure - the dependency stack is the #1 CosyVoice pain point in the community; a version conflict in transformers usually surfaces as an import error. A clean venv for the pack or aligning versions with ComfyUI's own env is the standard fix.

The honest caveat: quality is voice- and seed-dependent - CosyVoice's multi-language output is genuinely good, but not ElevenLabs-smooth on every clip. Feed it a clean reference and it's a legit local multi-language voice; feed it a noisy one and no wrapper will save you.