Nodes/omnivoice_comfy/OmniVoice TTS
ComfyUI Node

OmniVoice TTS

A ComfyUI node in audio/tts/OmniVoice with 15 inputs and 1 output.

By komikndr·Created 5 months ago·Updated 5 months ago· 13
OmniVoice TTS
  • model
  • ref_audio
  • audio
textHello from OmniVoice.
languageEnglish
instruct
speed1.00
duration0.0
num_step28
cfg3.0
seed0
t_shift1.00
denoisefalse
preprocess_prompttrue
postprocess_outputtrue
ref_text
Categoryaudio/tts/OmniVoice

Inputs (15)

NameTypeDefaultDescription
modelOMNIVOICE_MODEL
textSTRINGHello from OmniVoice.The actual words OmniVoice should generate as speech. For voice cloning, put the target sentence here and put the transcript of the reference clip into ref_text.
languageSTRINGEnglishLanguage name for the generated speech, for example English or Chinese. Leave at English unless you are intentionally switching languages.
instructSTRINGOptional speaking style or character direction. This affects delivery and tone, not the literal transcript.
speedFLOAT1.000.25–4Playback/speaking speed target. 1.0 is normal, lower is slower, higher is faster.
durationFLOAT0.00–300Optional target duration in seconds. Set to 0 to let OmniVoice estimate duration automatically from the text.
num_stepINT281–128Iterative generation steps. The node default is 28 as the current quality baseline; higher values are slower and may or may not help further depending on the prompt.
cfgFLOAT3.00–20Classifier-Free Guidance scale. The node default is 3.0 for the current tuned baseline; this pushes the result to follow the conditioning more strongly and can change timbre and articulation noticeably.
seedINT00–9223372036854776000Random seed for OmniVoice generation. With the current deterministic defaults this mainly matters if stochastic settings are re-enabled later, but it is exposed for reproducibility and future tuning.
t_shiftFLOAT1.000–2Mask-schedule shift used by OmniVoice's iterative decoding. The node default is 1.0 as the current tuned baseline for this custom node.
denoiseBOOLEANfalseAdds OmniVoice's denoise control token during generation. The node default is disabled because that is currently giving better audio on this setup.
preprocess_promptBOOLEANtruePreprocess the reference prompt before tokenization. This can trim long reference audio, remove silences, and normalize the prompt path for voice cloning.
postprocess_outputBOOLEANtruePostprocess the generated audio after decoding. This can remove long silences, normalize level behavior, and add fade/padding to avoid abrupt starts or ends.
ref_audiooptAUDIO
ref_textoptSTRINGRequired only for voice cloning. Enter the transcript of ref_audio, not the target text to generate.

Outputs (1)

NameTypeDescription
audioAUDIO