CosyVoiceInference

This is the node that does the actual talking in the CosyVoice half of the pack. Feed it the TTS_MODEL from CosyVoiceModelLoader, a short reference clip of a voice, and the text you want spoken, and you get back synthesized speech in that voice. The whole point of CosyVoice is speaker similarity from almost no reference - a few seconds is enough, and unlike many TTS pipelines it doesn't need you to re-train or fine-tune anything.

The engine behind it is CosyVoice2 (Alibaba FunAudioLLM), which the community knows as the pick when the reference voice's language and identity matter more than English-first polish. The wider ComfyUI TTS landscape leans on Chatterbox and Kokoro, but CosyVoice's Chinese strength and cross-lingual cloning are exactly the niche this pack fills. (CosyVoice 3 is out in the wild since late 2025, but this pack pins 2.)

The three modes

inference_type is an enum, default zero_shot:

• zero_shot - the classic clone path. You provide a reference clip (prompt_audio) and its transcript (prompt_text), plus the text to speak.
• cross_lingual - same voice, different language. It uses the reference clip but not its transcript, so you can have an English speaker say Chinese (or vice versa). The pack's example runs zero_shot, but this is the mode to try when the reference and target languages differ.
• instruct - pass an instruction (instruct_text) that shapes delivery - whisper it, sound excited, read like a news anchor. CosyVoice calls this its instruct2 path and it's the closest you get to emotion control without training.

Inputs that matter

• prompt_audio - AUDIO, the reference voice. Any ComfyUI audio works; the sample uses a LoadAudio node. The node squeezes it to mono and writes it to a temp wav internally.
• prompt_text - the transcript of that clip. For zero_shot, accuracy here matters more than you'd think; a wrong transcript visibly degrades the clone.
• tts_text - what you want synthesized.
• speed - FLOAT 0.5–1.5, step 0.1, default 1.0. 0.8 is a noticeably calmer read, 1.3 starts to sound rushed.
• instruct_text - only used in instruct mode.

Output is tts_speech (AUDIO) at the model's native rate (CosyVoice2's vocoder outputs 22.05 kHz), ready for SaveAudioAdvanced or any audio sink. IS_CHANGED returns a fresh timestamp on every call, so the node never caches - each queue run re-synthesizes, which means your preview is always current.

The is_japanese trap

The is_japanese optional toggle is a stub in the shipped code: tick it on and the node returns a silent audio clip instead of speech. There's no Japanese mode hiding behind it - it's an unfinished branch. Leave it unchecked. (Amusingly, the FishSpeech half of this same pack has the polarity flipped: that node returns silence when the toggle is off. Different packs, different footguns.)

Where it fits

The pack's example workflow is worth studying: LoadAudio → CosyVoiceModelLoader → CosyVoiceInference (zero_shot) → ConditionalBranchAudio → SaveAudioAdvanced, sharing the reference audio and text with a parallel FishSpeech chain so you can A/B the two engines through one output. Install the pack first - clone with --recursive into custom_nodes/, Python 3.12 env, the CUDA 12.8 constraints - then grab CosyVoice2-0.5B into models/tts/. If the loader works but you get empty output, the toggle above is your first suspect.