Spark_TTS_Clone

The sibling node in this pack does the trickier job. Spark_TTS_Clone takes a snippet of someone talking and a script, then reads the script back in that voice - zero-shot voice cloning, no fine-tuning, no training data beyond the reference clip. It turns "I want my video narrated like X" from a hiring decision into a Saturday-afternoon experiment.

Where the Creation node gives you gender and pitch sliders, this one gives you a person. It shares the same Spark-TTS 0.5B core, but instead of synthesizing a generic speaker it runs your reference audio through Spark-TTS's speaker encoder and conditions the generation on it. If you've already got a workflow that produces video, this is the node you wire in when you want it to talk.

How it works

Two ways to supply the reference. The easy one: pick a speaker_preset. That dropdown is populated live from a speakers_info.json file and ships with about 25 voices - ad_male_en, Benedict_male_en, Genshin_zhongli_male_zh, pitch-tagged variants like female_pitch_265_en, and more. Each preset maps to a bundled prompt audio file named {SpeakerName}_prompt.wav.

The better one: connect an Audio_reference from any node that outputs AUDIO - a Load Audio node, say. Audio_reference wins, no contest; the tooltip is blunt about it overriding speaker_preset. The node writes your clip to a temp file and uses it as the prompt speech. If you take that route, also fill custom_prompt_text with a transcript of the reference clip. The README and the tooltip both call it out as the thing that actually improves cloning accuracy, because Spark-TTS conditions on both the voice and its words.

Beyond the reference choice: text is your script. pitch and speed exist, but the tooltip hedges - "effect depends on cloning strength." Translation: they're weak knobs, mostly future-proofing, so don't count on them to fix a bad clone. The sampling quartet (temperature, top_k, top_p, max_new_tokens) and keep_model_loaded / use_cpu work exactly like the Creation node.

Outputs

Audio (standard AUDIO type - into your usual save/preview nodes) and Node Status (a string). Same pattern as its sibling: if the audio comes back silent, read the status string.

Installing it

Same pack, so same install: ComfyUI Manager (search "Spark-TTS") or:

cd ComfyUI/custom_nodes
git clone https://github.com/KERRY-YUAN/ComfyUI_Spark_TTS
cd ComfyUI_Spark_TTS
pip install -r requirements.txt   # with your ComfyUI Python

The one extra requirement is the Speaker_Preset folder, expected at ComfyUI/models/TTS/Speaker_Preset/ - it's a separate little repo, not part of the model weights. The auto-downloader fetches both, but it's the preset half that usually fails, because it's a git clone and needs Git in your PATH. If the dropdown shows a "(No speakers directory found)" placeholder, that's what happened - run Model_Download.bat in the node folder.

Where people get burned

Reference quality matters more than anything else here. Clean, single-speaker, ten-to-twenty seconds of reference audio beats a noisy five-second clip every time - it's the same lesson every ComfyUI TTS thread eventually teaches. If the clone drifts toward a generic tone, add or improve custom_prompt_text rather than fiddling with pitch. And you can grow the preset list yourself: drop a {Name}_prompt.wav into Speaker_Preset and add a matching line to speakers_info.json, and it shows up in the dropdown on the next refresh.

One honest note on expectations: the ComfyUI TTS crowd generally ranks dedicated cloners like EchoTTS and Chatterbox above Spark-TTS's 0.5B for sheer fidelity. What this node has going for it is that it's already in your graph, with a one-file reference setup and working auto-download - plenty for narration, tutorials, and game-style voices.

And since this is cloning a real human voice we're talking about: use it on your own voice, or with the person's blessing, not on the famous audio you found online. The model is good enough that the question stops being academic.