TS CosyVoice Speaker Text To Voice

TS CosyVoice Speaker Text To Voice is the "I want this character to sound identical on every single line, forever" node. Plain Text to Voice re-encodes your reference clip every run; this one pulls a saved .pt speaker preset from disk instead. The voice stays stable between sessions, and the only thing you change per line is the text and the emotion instruction. For a series, a podcast, a long dub - anything where voice consistency across hundreds of lines matters - this is the workflow the pack's README calls out as arguably its most useful node.

It exists because of a two-step habit that's easy to build: Save Speaker once to lock in a voice, then this node forever after. The reference clip only ever touches the graph a single time.

How it works

Under the hood it's the same CosyVoice inference_instruct2 path as Text to Voice - text in, instruction guides delivery, out comes speech. The difference is where the timbre comes from. Instead of a raw audio prompt, it loads the precomputed speaker features you saved earlier via Save Speaker (those .pt files live in ComfyUI/models/cosyvoice/speaker/). Because the features are computed once and stored, you get the same voice identity run after run, and the per-line variation comes entirely from instruct_text or emotion_preset.

The inputs that matter

• speaker_preset - a combo that lists the .pt files in your speaker folder. This is the whole point: pick the voice, don't re-supply audio.
• text - the line.
• instruct_text / emotion_preset - same system as Text to Voice: 12 bundled delivery presets, or "Your instruction" plus free-form text. This is how one voice goes from warm to angry to exhausted across a scene.
• speed - 0.5× to 2×.
• seed - -1 for variety, fixed otherwise.

Output

One audio output, straight into Save Audio / Preview Audio.

The workflow it belongs to

Model Loader ──→ Save Speaker (once, per voice)
                 └──→ Speaker Text To Voice (every line) → Save Audio

Setup cost is a couple of minutes per voice; after that each line is just text plus an emotion. That's the entire appeal: clone once, direct forever.

Installing

Same pack install as the rest of the 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 the always-two-extras: one ONNX Runtime (onnxruntime-gpu on NVIDIA, onnxruntime on CPU/Apple Silicon) and CUDA torch installed before the pack deps. Model download is automatic on first Model Loader run.

Troubleshooting

• The preset dropdown is empty - you haven't run Save Speaker yet, or it saved somewhere other than models/cosyvoice/speaker/. Give Save Speaker a descriptive speaker_name and rerun.
• The voice drifts from the original reference - check what reference_text got baked into the preset when you saved it. If it was empty because Whisper wasn't installed, the clone is noticeably worse; re-save with the text typed in.
• Lines sound too similar to each other - vary instruct_text or seed rather than the voice.
• It doesn't appear in the Add Node menu - this pack needs ComfyUI >= 0.3.40 (V3 schema).

If you're doing more than a handful of lines in one voice, skip the raw reference path and go straight to the preset. It's the difference between babysitting a reference clip and just writing lines.