ComfyUI Node
VRGDG VoxCPM2 Voice Clone / TTS
A ComfyUI node in VRGDG/Audio with 18 inputs and 4 outputs.
VRGDG VoxCPM2 Voice Clone / TTS
- reference_audio
- prompt_audio
- audio
- saved_audio_path
- sample_rate
- status
◄textHello from VoxCPM2.►
◄modetext_to_speech►
◄deviceauto►
◄cfg_value2.0►
◄inference_timesteps10►
◄load_denoiserfalse►
◄normalize_textfalse►
◄retry_badcasetrue►
◄retry_badcase_max_times3►
◄retry_badcase_ratio_threshold6.0►
◄min_len2►
◄max_len4096►
◄save_filename_prefixvoxcpm2►
◄reference_audio_path►
◄prompt_audio_path►
◄prompt_text►
CategoryVRGDG/Audio
Inputs (18)
| Name | Type | Default | Description |
|---|---|---|---|
| text | STRING | Hello from VoxCPM2. | What you want the model to say. For cloning modes, this is the new target speech, not the transcript of your reference clip. |
| mode | COMBO | text_to_speech | Choose how VoxCPM2 should speak: - text_to_speech: regular TTS with no voice reference. - voice_design: zero-shot styled speech guided by your wording alone. - prompt_continuation: continue from a prompt clip using prompt_audio plus prompt_text. - controllable_clone: easiest voice clone. Provide a clean reference clip. - ultimate_clone: strongest cloning mode. Provide a reference clip and the exact transcript in prompt_text. |
| device | COMBO | auto | auto picks CUDA when available and falls back to CPU. CUDA is strongly recommended for VoxCPM2. |
| cfg_value | FLOAT | 2.00–20 | Classifier-free guidance strength. A good starting point is 2.0. Lower can sound looser; higher can sound more forced. |
| inference_timesteps | INT | 101–200 | Number of inference steps. 10 is a solid default. Higher may improve quality slightly but is slower. |
| load_denoiser | BOOLEAN | false | Optional cleanup for noisy prompt or reference audio before cloning. Useful if your source clip has hiss, room noise, or light background noise. |
| normalize_text | BOOLEAN | false | Normalize text before generation. Can help with number/date expansion and cleaner pronunciation on some inputs. |
| retry_badcase | BOOLEAN | true | Retry obviously bad generations automatically. This can improve reliability but may take longer. |
| retry_badcase_max_times | INT | 31–10 | Maximum retry attempts when retry_badcase is enabled. |
| retry_badcase_ratio_threshold | FLOAT | 6.01–20 | Bad-case threshold for retry logic. Leave this alone unless you are tuning generation behavior. |
| min_len | INT | 21–128 | Minimum generated audio token length. Usually safe to leave at the default. |
| max_len | INT | 409632–16384 | Maximum generated token length. Increase if long text is getting cut off. Higher values can use more VRAM and time. |
| save_filename_prefix | STRING | voxcpm2 | Base name for the saved WAV file in ComfyUI/output/VRGDG_AudioFiles. |
| reference_audioopt | AUDIO | Speaker reference for voice cloning. Use a clean single-speaker clip, ideally 10 to 30 seconds. | |
| reference_audio_pathopt | STRING | Speaker reference for voice cloning. Use a clean single-speaker clip, ideally 10 to 30 seconds. | |
| prompt_audioopt | AUDIO | Prompt/continuation audio. Use this for prompt_continuation, or for ultimate_clone when you want the model to match a specific spoken example. | |
| prompt_audio_pathopt | STRING | Prompt/continuation audio. Use this for prompt_continuation, or for ultimate_clone when you want the model to match a specific spoken example. | |
| prompt_textopt | STRING | Exact transcript of the prompt/reference clip. Required for prompt_continuation and ultimate_clone. Match the spoken words as closely as possible. |
Outputs (4)
| Name | Type | Description |
|---|---|---|
| audio | AUDIO | — |
| saved_audio_path | STRING | — |
| sample_rate | INT | — |
| status | STRING | — |