Qwen3-TTS Voice Design:
Qwen3VoiceDesign generates spoken audio from written text using a compatible Qwen3-TTS VoiceDesign model. You can describe the kind of voice or delivery you want in everyday language, then use the generated speech in an audio workflow. This makes the node useful when you want to create a voice from a description rather than provide a recording of a speaker. The node asks the model to synthesize speech and prepares the resulting waveform for use by other ComfyUI audio nodes.
Qwen3-TTS Voice Design Input Parameters:
model
A required QWEN3_MODEL connection that supplies the speech-generation model. This node calls the model’s voice-design capability, so the connected model must support that capability; a different Qwen3-TTS model type may fail even if it connects successfully. The source gives no default model or selectable model list. Load a VoiceDesign model, such as the example named in the node’s error message, before generating audio.
text
The required multiline STRING containing the words you want spoken. Its wording determines the speech content passed to the model. You can enter a sentence or a longer passage, but the node source specifies no default, minimum length, or maximum length. Keep the text aligned with what you want to hear, and review spelling and punctuation before generating.
instruct
The required multiline STRING describing the voice or speaking style you want the model to use. It is passed to the model as the voice-design instruction, separate from the words to be spoken. For example, you might describe a calm, clear delivery rather than put that direction into text. The node source specifies no default, minimum length, or maximum length, and does not guarantee that every descriptive detail will be reproduced exactly.
language
Selects the language passed to the model for speech generation. The available options are Auto, Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian; the default is Auto. With Auto, the node passes no explicit language selection and leaves language handling to the model. Choose a specific option when you know the intended language. The source defines no numeric minimum or maximum for this setting.
seed
An integer used to initialize PyTorch’s random generation before the model creates speech; the node also initializes CUDA random generation when CUDA is available. Its default is 42, its minimum is 1, and its maximum is 18446744073709551615. Keeping the same seed is useful when comparing changes to your text or instruction because it holds this source of randomness constant. A seed alone does not guarantee identical audio across different models, devices, or execution environments.
Qwen3-TTS Voice Design Output Parameters:
audio
An AUDIO value containing the generated speech. It includes a waveform, which holds the audio samples, and a sample rate, which tells downstream audio nodes how quickly to play those samples. The node uses the first waveform returned by the model and converts it to ComfyUI’s batched audio format. Connect this output to a compatible audio playback, processing, or saving node to hear or use the result.
Qwen3-TTS Voice Design Usage Tips:
- Load a model that supports VoiceDesign before connecting it to
model. The node explicitly reports an error when a model does not support voice-design generation. - Put only the words to be spoken in
textand place voice or delivery directions ininstruct. This keeps the requested speech separate from the description guiding its sound. - Choose a specific
languagewhen you know the language of the text; useAutowhen you want to leave that choice to the model. - Keep the
seedunchanged while comparing different instructions, then change it when you want to try another random generation. Compare results with the same model and environment when consistency matters.
Qwen3-TTS Voice Design Common Errors and Solutions:
Model Type Error: You are trying to use 'Voice Design' with an incompatible model. Please load a 'VoiceDesign' model (e.g. Qwen3-TTS-12Hz-1.7B-VoiceDesign).
- Explanation: The connected model reported that it does not support the voice-design operation this node requires.
- Solution: Connect a compatible VoiceDesign model to
model, then run the node again.
ValueError from voice-design generation
- Explanation: The model can raise a
ValueErrorfor a reason other than the incompatible-model case. The node passes those errors through without changing their messages, so the exact wording depends on the model and the problem it encountered. - Solution: Read the full error message. Check that you connected a compatible model and that
text,instruct, andlanguageare appropriate for it; address the condition named in the message before retrying.
Error while converting the generated waveform to audio
- Explanation: After generation, the node takes the first returned waveform and prepares it as ComfyUI audio. An error at this stage can occur if the model does not return a waveform in the form the conversion code expects. The source does not define one fixed error message for this case.
- Solution: Confirm that you are using a compatible VoiceDesign model. If the error persists, retain the complete traceback and model details for troubleshooting; changing the voice description alone may not resolve a waveform-format problem.
