TS CosyVoice Save Speaker

TS CosyVoice Save Speaker is the one-time cost that makes the rest of the pack painless. Feed it a reference clip and it extracts the speaker features that define the voice, then writes them to a .pt file in ComfyUI/models/cosyvoice/speaker/. From then on, Speaker Text To Voice reads that file instead of you re-supplying audio, and every line you generate keeps the same voice. It's the difference between loading a reference clip into every single workflow and loading it once, ever.

The mental model: this is the "save the character" button. Narrator, villain, sidekick - you save each once, name it properly, and build a little voice library that every future project draws from.

How it works

The node trims your reference_audio to 30 seconds, downsamples to mono 24 kHz, and runs the model's zero-shot speaker encoder over it to pull out the timbre embedding. Then it needs the text being spoken in that clip - because CosyVoice's cloning uses the reference transcript as a prompt anchor. If you leave reference_text empty, it auto-transcribes the clip with Whisper, which triggers a ~140 MB base model download into ComfyUI/models/whisper/ the first time. If Whisper or ffmpeg isn't installed, it still saves the preset but with an empty reference_text - and the resulting clone is noticeably worse. The pack logs a warning about exactly this. Type the text yourself for anything important.

The inputs that matter

• reference_audio - the voice. Same quality rules as everywhere: clean, one speaker, 5–15 seconds, no music or reverb.
• reference_text - leave blank to let Whisper handle it, or type it for reliability. If the language is rare or the speech is mumbled, definitely type it.
• speaker_name - the filename, no extension. Name it like you'll find it in six months: narrator_female_warm, villain_male_deep, kid_neutral_v2. The default my_speaker is how voice libraries die.

Output

A saved_path string telling you where the preset landed. The node is marked as an output node, so it also just works as a terminal step in the graph.

The workflow it belongs to

Model Loader ──→ Save Speaker (once) ──→ Speaker Text To Voice (every line) → Save Audio

Save once per voice, then never feed a reference clip into a workflow again.

Installing

Same pack as the whole family - ComfyUI Manager (search TS CosyVoice), or:

cd ComfyUI/custom_nodes
git clone https://github.com/AlexYez/comfyui-ts-cosyvoice.git
cd comfyui-ts-cosyvoice
pip install -r requirements.txt

Plus one ONNX Runtime (onnxruntime-gpu / onnxruntime) and CUDA torch installed before the pack deps. If you want auto-transcription, also pip install openai-whisper and make sure ffmpeg is on PATH. Neither is required - the node degrades gracefully without them, just with a worse clone.

Troubleshooting

• Preset saves but the clone sounds off - almost certainly an empty reference_text. Check the log for the whisper warning and re-save with the transcript typed in.
• It saves with saved_path but the dropdown in Speaker Text To Voice doesn't show it - the preset went somewhere other than models/cosyvoice/speaker/, or ComfyUI's schema cache is stale; restart ComfyUI.
• Whisper never runs - openai-whisper missing, or ffmpeg not on PATH. Install both, restart, retry.
• First save downloads a 140 MB whisper model and feels slow - that's one-time.

The whole pack is built around this two-step rhythm. Skip it and you're re-encoding references forever; use it and the hard part of voice consistency is done.