ComfyUI-Qwen3-TTS-DarioFT Introduction
ComfyUI-Qwen3-TTS-DarioFT is a set of custom ComfyUI nodes for turning written text into speech with Qwen3-TTS. You can give a character one of nine preset voices, describe a new voice in words, or use a short recording as a reference for voice cloning. The extension also includes tools for training a model on your own voice recordings and comparing the result with a reference.
For an AI artist, this means you can create narration, character dialogue, or spoken lines within a ComfyUI workflow instead of treating audio as a separate step. Speech generation supports Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian.
How ComfyUI-Qwen3-TTS-DarioFT Works
Think of the workflow as choosing a performer, giving them a script, and directing how they should sound:
- Choose a model with Qwen3-TTS Loader. The model determines which kind of voice creation is available: preset speakers, voice descriptions, or reference-audio cloning.
- Connect a matching generation node and enter the text you want spoken.
- Provide voice direction. Depending on the node, you choose a speaker, describe the voice in words, or supply a reference recording.
- Run the workflow to generate audio. You can change the text or voice settings and run it again to explore alternatives.
Voice cloning and fine-tuning both use recordings, but they serve different purposes. Cloning uses a short reference clip to guide speech generation without a training session. Fine-tuning trains a dedicated model from a folder of paired recordings and transcripts, which you can then load for future generations.
On first use, the extension downloads only the model you select. Models are kept in ComfyUI/models/Qwen3-TTS/; saved voice prompts are kept in its prompts/ subfolder.
ComfyUI-Qwen3-TTS-DarioFT Features
Preset voices with Custom Voice
Connect a CustomVoice model to Qwen3-TTS Custom Voice when you want a ready-to-use speaker. Choose from nine presets, including Vivian and Ryan, then enter the spoken text. Set language to your target language or Auto. The optional instruct field lets you request a delivery such as “Happy” or “Whispering.”
This is a straightforward starting point when you need dialogue quickly and do not have a reference recording.
New voices with Voice Design
Connect a VoiceDesign model to the Voice Design generator. Enter the words to speak in text, and describe the desired sound in instruct. For example: “A deep, resonant male voice, narrator style, calm and professional.”
This is useful when you have a character concept but no recording to clone. Changing the description lets you audition different interpretations of the same line.
Reference-based Voice Clone
Connect a Base model to Voice Clone and provide:
ref_audio: A short recording of the voice to use as a reference. The usage guidance recommends 1–10 seconds; shorter clips around 5–15 seconds may also help if generation hangs.ref_text: The words spoken in the reference clip. An accurate transcription can improve quality.text: The new words you want the voice to say.
Two controls help manage longer generations. max_new_tokens limits generated tokens and defaults to 2048; a lower value such as 1024 may be suitable for short lines if generation gets stuck. ref_audio_max_seconds automatically trims the reference to 30 seconds by default. Setting it to -1 disables trimming, but long reference audio can increase the risk of a hang.
Reuse a voice with prompt caching
If you are making many lines in the same cloned voice, Qwen3-TTS Prompt Maker can calculate the voice features from a reference clip once. Connect its Qwen3_Prompt output to Voice Clone rather than recalculating those features for every line.
You can keep those features as a reusable voice prompt:
- Connect a
QWEN3_PROMPTto Qwen3-TTS Save Prompt. It saves a.safetensorsfile underComfyUI/models/Qwen3-TTS/prompts/. - Later, select the saved file in Qwen3-TTS Load Prompt and connect its output to Voice Clone.
For example, you can generate a voice with Voice Design, use its audio with Prompt Maker, save the prompt, and reload it for subsequent cloned lines.
Train and assess a dedicated voice
The fine-tuning nodes support both 1.7B and 0.6B Base models:
- Prepare recordings and transcripts. Put
.wavfiles and matching.txtfiles in a folder, using the same filename for each audio–text pair. Include a representativeref.wav, or specify a reference in the node. - Build the dataset. Point Qwen3-TTS Dataset Maker at the folder to create
dataset.jsonl. - Process the audio. Use Qwen3-TTS Data Prep to create a
*_codes.jsonlfile. Itsbatch_sizedefaults to 16; lowering it can reduce memory use if processing runs out of GPU memory. Unchanged datasets can be skipped using SHA256-based caching. - Train. Connect the processed file to Qwen3-TTS Finetune, choose a Base
init_model, set anoutput_dir, and enter aspeaker_name. Queue the workflow to start training. The node saves checkpoints after each training epoch and supports resuming from a checkpoint. - Try the trained voice. In Qwen3-TTS Loader, set
local_model_pathto the trainedoutput_dir/epoch_X. Use Qwen3-TTS Custom Voice and type the exact trainedspeaker_nameinto its input. It will not appear in the preset-speaker dropdown.
Training offers memory-related controls including gradient_checkpointing (approximately 30–40% VRAM savings), use_8bit_adam (requires bitsandbytes), and gradient_accumulation, which simulates a larger batch size. warmup_ratio controls learning-rate warmup and defaults to 0.1. You can also use Qwen3-TTS Audio Compare to assess generated speech against a reference through speaker similarity, mel-spectrogram distance, and speaking-rate ratio.
Model loading and attention options
In Qwen3-TTS Loader, repo_id selects a model. source chooses HuggingFace or ModelScope when a download is needed, while optional local_model_path overrides repo_id for a model already on your computer, including a fine-tuned one.
You can usually leave attention on auto: it tries Flash Attention 2 and falls back to PyTorch’s sdpa when needed. If you encounter an attention-related issue, you can select sdpa or try eager as a fallback.
ComfyUI-Qwen3-TTS-DarioFT Models
Choose the model family that matches the voice-generation node you want to use:
| Model | Use it for |
|---|---|
Qwen3-TTS-12Hz-1.7B-CustomVoice | Generating speech with preset speakers using Custom Voice. |
Qwen3-TTS-12Hz-0.6B-CustomVoice | Generating speech with preset speakers using Custom Voice. |
Qwen3-TTS-12Hz-1.7B-VoiceDesign | Creating a voice from a written description using Voice Design. |
Qwen3-TTS-12Hz-1.7B-Base | Voice Clone generation or fine-tuning a dedicated voice. |
Qwen3-TTS-12Hz-0.6B-Base | Voice Clone generation or fine-tuning a dedicated voice. |
Qwen3-TTS-Tokenizer-12Hz | Preparing audio data for fine-tuning; it is not a voice-generation model. |
The 1.7B and 0.6B names identify two model sizes where both are available. The model family is the essential choice: for instance, select a Base model to clone a recorded voice, not a CustomVoice model. No 0.6B VoiceDesign variant is listed in the provided model set.
The first run downloads the selected model rather than every variant. If you previously used the extension and have cached models in HuggingFace or ModelScope cache locations, the extension can migrate them into the ComfyUI models folder.
Troubleshooting ComfyUI-Qwen3-TTS-DarioFT
Speech generation does not finish, and GPU usage stays high
Qwen3-TTS can occasionally continue generating without producing a stop signal. This is a known upstream issue, particularly relevant to long reference clips or long outputs.
- For Voice Clone, try a shorter
ref_audioclip. Start with a clear segment around 5–15 seconds, and keepref_audio_max_secondsat its default 30 rather than disabling trimming. - If you are generating a short line, lower
max_new_tokensfrom 2048 to 1024 and try again. - Try a different seed.
- If the current run remains stuck, stop the Python process and restart ComfyUI before rerunning the workflow.
Generation is slow on Windows
Flash Attention 2 may not be readily available on Windows.
- In Qwen3-TTS Loader, set
attentiontosdpaand test the workflow. - If
sdpacauses problems, tryeageras a fallback. - If you specifically need Flash Attention 2, consider a Linux environment through WSL2.
Frequently asked questions
Why did a model download when I ran the workflow?
The extension downloads a selected model on demand the first time it is needed. It does not download all model variants at once.
Which model should I choose to imitate a short recording?
Choose a Base model and use Voice Clone. Supply the recording as ref_audio and, where possible, its transcript as ref_text.
Can I reuse one cloned voice across many lines?
Yes. Use Qwen3-TTS Prompt Maker to calculate the reference voice features, then Save Prompt and Load Prompt to reuse them without recalculating each time.
Why is my fine-tuned speaker not in the speaker dropdown?
The trained name is not added to the preset list. Load the fine-tuned epoch with local_model_path, then type your exact trained speaker_name in the Custom Voice input.
What if dataset processing runs out of GPU memory?
Reduce the batch_size in Qwen3-TTS Data Prep from its default of 16 and run it again. For training, the fine-tuning node also provides memory-saving options such as gradient checkpointing.
Learn More about ComfyUI-Qwen3-TTS-DarioFT
- Qwen3-TTS is the underlying speech-generation library on which the author based this extension.
- If your audio workflow also needs transcription, the author provides ComfyUI-Qwen3-ASR, a separate speech-to-text extension with compatible outputs.
