Nodes/FireRedTTS3-ComfyUI/FireRedTTS3 Voice Design
ComfyUI Node

FireRedTTS3 Voice Design

Design a brand-new voice from a text description (FireRedTTS3-Instruct).

By Saganaki22·Created 11 days ago·Updated 7 days ago· 17
FireRedTTS3 Voice Design
  • firered_model
  • audio
  • voice_plan
instructionA gentle young female voice, speaking a little slowly, with a playful touch.
textHello! This is FireRedTTS3 running natively inside ComfyUI.
languageauto
text_temperature0.70
text_top_p0.80
text_top_k20
text_repetition_penalty1.00
n_timesteps10
inference_cfg1.20
seed42
max_audio_seconds64
do_tntrue
do_splittrue
cross_fade_ms50
CategoryFireRedTTS3

Inputs (15)

NameTypeDefaultDescription
firered_modelFIREREDTTS3_MODEL
instructionSTRINGA gentle young female voice, speaking a little slowly, with a playful touch.Natural-language voice description (gender, age, timbre, emotion, pace, accent). Any language works; the model card's Chinese examples are also good templates. No reference audio needed.
textSTRINGHello! This is FireRedTTS3 running natively inside ComfyUI.Text to synthesize with the designed voice.
languageCOMBOautoLanguage of the text; used for text normalization and sentence splitting only. Voice design synthesis itself only works reliably in Chinese/English (upstream limitation: no language conditioning without a reference clip); other languages produce accented/garbled speech. For other languages, design the voice in English/Chinese, then clone the result with the base model and a language tag.
text_temperatureFLOAT0.700–2Sampling temperature for the model's voice-plan text (Chain-of-Thought).
text_top_pFLOAT0.800–1Top-p for the voice-plan text.
text_top_kINT200–500Top-k for the voice-plan text. 0 disables.
text_repetition_penaltyFLOAT1.001–2Repetition penalty for the voice-plan text. 1.0 is the official default.
n_timestepsINT101–50Flow-matching steps per generated audio patch. 10 is the official default; more is slower with diminishing returns.
inference_cfgFLOAT1.200–4Classifier-free guidance strength for the flow head. 0 disables CFG. Official defaults: 2.0 for cloning, 1.2 for design/edits.
seedINT420–21474836470 uses the current random state. A positive value is repeatable.
max_audio_secondsFLOAT644–160Hard cap on generated audio length per sentence (64s is the official maximum).
do_tnBOOLEANtrueRun text normalization (numbers, dates, units to spoken form). Chinese/English use local wetext; other languages get basic cleaning.
do_splitBOOLEANtrueSplit long text into sentences and generate them one by one (cross-faded together).
cross_fade_msFLOAT500–500Cross-fade between sentence segments in milliseconds.

Outputs (2)

NameTypeDescription
audioAUDIO
voice_planSTRING