VRGDG VoxCPM2 Voice Clone / TTS

VoxCPM is the open-weights speech model that's been quietly winning the "actually good local TTS" crowd, and VRGDG_VoxCPM2Generate wraps it for ComfyUI with a modes menu that covers everything from plain text-to-speech to serious voice cloning. It's part of the VRGameDevGirl pack, which means it's aimed at the music-video pipeline - clone a voice for your AI singer, generate narration, or just have a character say something new in an existing voice. No API key, no cloud, no per-character fee. It's all local.

The node loads the model with VoxCPM.from_pretrained(...) on first run, so your first generation is really a model download. Once it's cached, it's fast - the community's verdict on VoxCPM is "really fast, very accurate," and it handles far more than English.

The modes menu (this is the whole game)

• text_to_speech - plain TTS, no voice reference. The text field is what it says.
• voice_design - zero-shot styled speech guided by your wording alone. No reference clip needed, you steer the character through your prompt text.
• prompt_continuation - continue speaking from a prompt clip; give it prompt_audio plus the exact prompt_text.
• controllable_clone - the easiest real voice clone. Feed a clean 10–30 second single-speaker reference clip and it talks in that voice.
• ultimate_clone - the strongest clone. Same idea, but you must also provide the exact transcript in prompt_text; matching the words closely matters.

The inputs that matter

• text - what you want the model to say. For cloning modes this is the new target speech, not a transcript of the reference.
• reference_audio / reference_audio_path - your speaker reference. Clean, single speaker, 10–30 seconds is the sweet spot the tooltip asks for.
• prompt_text - required for prompt_continuation and ultimate_clone: the exact transcript of the prompt/reference clip.
• cfg_value - CFG strength, default 2.0. Lower sounds looser, higher sounds more forced. 2.0 is a genuinely good start.
• inference_timesteps - default 10. Higher is slightly better, notably slower.
• max_len - if long text gets cut off, raise it. Costs VRAM and time.
• retry_badcase - on by default; auto-reruns obviously bad generations up to retry_badcase_max_times (default 3). Slower but much more reliable.

Outputs: audio (the AUDIO tensor), saved_audio_path (a WAV written to ComfyUI/output/VRGDG_AudioFiles/, prefix from save_filename_prefix), sample_rate, and a status string.

Install

The pack install plus the VoxCPM dependency:

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl
pip install -r requirements.txt   # includes voxcpm, transformers, torchaudio

Or search vrgamedev in ComfyUI Manager. First run downloads the model - give it time. device defaults to auto (CUDA if available), and the tooltip is blunt: CUDA is strongly recommended. This is not a CPU-friendly node.

Common issues

• First run "hangs": model download. Watch the console.
• Cloned voice doesn't sound right: the reference clip is the weak link. Hiss, room tone, or a second speaker wreck cloning - the load_denoiser option exists precisely to clean up noisy source clips before cloning.
• Bad generations that don't retry: check retry_badcase is on and the threshold (retry_badcase_ratio_threshold) is at its default - don't touch that one unless you're tuning.
• Accent bleeding / language quirks: community reports say it's a strong multi-language model but can carry an accent from your reference. That's a feature when you want it, a bug when you don't - use a neutral, clean reference.

One honest note: VoxCPM2's own ecosystem is young, so expect occasional rough edges, and the model card should be your second read after this page. But for fully local, genuinely good voice cloning with no API in the loop, this is one of the better nodes you can drop into a ComfyUI graph today.