CosyVoice 3.0 TTS (AIIA)

This is the one node that speaks with the entire CosyVoice family - V1's 300M, V2, and the 3.0 0.5B - and switches its behavior to match whatever the Loader handed it. The author's framing is "版本自适应" (version-agnostic), and it's not hype: the same text field does different things depending on which model you loaded, and the node knows which. That flexibility is the whole point. CosyVoice 3.0 is one of the best open TTS models around - 3-second cloning, natural emotion, real dialect support - and this node is how you drive it without touching the upstream code.

The five generation modes

The README breaks it into five, and they're worth internalizing:

1. Instruct - text plus a style description ("speak very slowly, magnetic tone"). The description controls performance, not identity.
2. Zero-shot clone - text plus a short reference_audio. The clone path; V3 needs only ~3 seconds.
3. Hybrid - reference audio for identity plus a description for emotion/dialect.
4. Cross-lingual - a cloned voice speaking any of the 9 supported languages.
5. SFT / fixed ID - pick a built-in voice via spk_id.

The inputs that matter

• tts_text - what it says. CosyVoice uses a streaming architecture, so long text isn't a problem here (unlike the token-window-limited models in this pack).
• instruct_text - the style/emotion/dialect description. With V3 this is powerful and literal: "在大雨中嘶吼" works. The README's tip: specific scene descriptions ("whispering at 3am") land better than single adjectives.
• reference_audio (optional) - the clone source. For V3, feed the highest-quality clip you have; the model rewards it.
• base_gender - Female/Male, the identity fallback when you're in pure instruct mode with no reference audio.
• dialect / emotion - preset dropdowns (Cantonese, Sichuan, Henan... and Happy, Sad, Robotic, even "Peppa Pig style") that get prepended to your instruction. Handy, but the free-text field does the same thing.
• speed / seed - the usual.
• spk_id - for SFT mode (e.g. pure_1). Leave empty for zero-shot models.

Output: audio (AUDIO).

Version-specific traps

• 300M-Instruct needs English instructions. The README is emphatic: "Sad tone", "Fast speed" work; Chinese instructions get silently ignored. There's also a fixed bug where the model used to read its own instruction aloud.
• 300M-Instruct can't do dialects - architecture limitation, not a setting.
• The 0.6B/0.5B V3 class wants a reference audio for the "identity" layer; without one it falls back to a built-in seed voice based on base_gender, which is fine for testing but not for shipping a character.

Wiring it up

CosyVoice Model Loader → CosyVoice 3.0 TTS → Audio Enhance (Resemble) → Audio Post-Process. That's the "sounds like a real recording" chain, because CosyVoice outputs at 22/24kHz and the enhance step rebuilds the missing top end.

Troubleshooting

• Instructions do nothing - you're on a Base/SFT-style model that doesn't take them, or on 300M-Instruct with Chinese text.
• Clones sound flat - V3's clone quality scales with reference quality; give it a clean, unprocessed clip.
• Crashes (AudioDecoder, KeyError) - the old upstream bugs; the pack ships fixes, but keep your PyTorch off 2.9.x (the README documents the regression).