⚙️ Higgs Audio v3 Engine

Higgs Audio v3 is the "just type what you want it to do" engine. Instead of wiring up separate emotion nodes and reference clips, you drop official tags straight into your text - <|emotion:amusement|>, <|style:whispering|>, <|sfx:laughter|> - and the model performs them inline. It's a 4B multilingual model with zero-shot voice cloning and support for 100+ languages, and it's built to run natively on the suite's modern Transformers 5 main environment rather than in an isolated legacy runtime like some of its cousins.

This is a config node. It outputs a tts_engine; the text (with your tags in it) goes into 🎤 TTS Text or 📺 TTS SRT, which do the generating.

How it works

The magic isn't in this node's knobs - it's in the tags you write in your prompt. The suite accepts the official pipe-bracket form (<|style:whispering|>) and also a friendlier <style:whispering> alias that it normalizes internally, so you don't have to type the pipes if you don't want to. Because v3 infers language from the text (and from reference context when cloning), there's deliberately no language parameter here - you don't set a language, you just write in the language you want. Voice cloning is zero-shot from reference audio, and it works in both the Text and SRT nodes.

The inputs and outputs that matter

This is one of the simplest engine nodes in the pack - no optional inputs at all, just a short list of generation settings:

• model - one option, higgs-audio-v3-tts-4b. Nothing to pick.
• temperature - the expressiveness dial (default 1.0). With inline tags doing the heavy lifting on emotion, you rarely need to crank this.
• top_p / top_k - sampling breadth; defaults are sensible.
• max_new_tokens - caps how much audio a segment can produce.

device, dtype, and attention are the runtime settings - leave them on auto unless you're deliberately forcing a specific attention backend (sdpa, flash_attention, sageattention) for speed.

The single output is tts_engine → into TTS Text or TTS SRT. The real work is choosing your tags in the prompt, which the 🏷️ Multiline TTS Tag Editor can help you assemble in its dedicated "Higgs Audio v3" mode.

Installing it

Part of the pack. ComfyUI Manager → search "TTS Audio Suite" → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/diodiogod/TTS-Audio-Suite.git
cd TTS-Audio-Suite
python install.py

The v3 4B model (~8GB) auto-downloads to ComfyUI/models/TTS/higgs_audio_v3/ on first generation. Because v3 runs in the main environment on Transformers 5, make sure your suite install is current - this is one of the engines that benefits from the modern stack rather than a legacy runtime.

Common issues & troubleshooting

Your tags come out spoken as words. If the model literally says "emotion amusement" instead of performing it, your tag syntax is off. Use the official <|emotion:amusement|> form or the <emotion:amusement> alias exactly - a stray space or bracket breaks it. The tag editor's v3 mode inserts valid tags for you.

A tag does nothing. v3 supports emotion, style, prosody, and sfx categories; a made-up value inside a real category may just be ignored. Stick to documented values (amusement, whispering, laughter, and so on) rather than inventing your own.

Wrong language out. There's no language dropdown by design - v3 reads the language from your text. If it drifts, make the prompt unambiguously in your target language, and when cloning, give it a reference clip in that language to reinforce it.

First run is slow / stalls. It's pulling ~8GB once. Watch the console; subsequent runs use the cache.