FL Chatterbox Turbo TTS

The standard FL Chatterbox TTS will read your line competently. This one will laugh at it. FL Chatterbox Turbo TTS is the variant in filliptm's pack that swaps in a faster, GPT2-based model with paralinguistic tags - drop [laugh] or [sigh] into your text and the speech actually expresses it. It's the node you reach for when a flat robot reading won't do, and it's the one the README demo is built around.

Same Chatterbox family, same zero-shot voice cloning idea: feed it a reference clip and it speaks in that voice. But the Turbo model (t3_turbo_v1) is built for speed and for acting, not for raw quality. Where the standard node leans on classifier-free guidance and an exaggeration dial, Turbo samples tokens directly with classic language-model knobs - top_k, top_p, temperature, repetition_penalty. Fewer parameters, faster generation, and the trade-off is that it's English-only, and the emotive tags don't carry across languages.

How it works. Same three-part architecture as the rest of the pack - a text-to-token transformer, the S3 tokenizer, and an S3Gen vocoder producing 24 kHz audio. Turbo's tokenizer understands the paralinguistic tags and turns them into the acoustic details that make a laugh a laugh instead of a vocalized "ha ha". Models download automatically on first run from the ResembleAI/chatterbox-turbo HuggingFace repo into ComfyUI/models/chatterbox/chatterbox_turbo/, so budget some time for that first generation.

The inputs that matter. The tag support lives in text, so that's the fun one:

Hello there! [laugh] Isn't this amazing? [sigh] I just love text to speech.

Supported tags: [laugh], [sigh], [gasp], [chuckle], [cough], [sniff], [groan], [shush], [clear throat]. Don't over-stuff a line - one or two tags reads natural, five reads like a cartoon.

The rest of the required inputs are sampling dials: temperature (0.05–2.0) for randomness, top_k (1–5000, default 1000) for how many candidate tokens, top_p (0.1–1.0, default 0.95) nucleus sampling, repetition_penalty (1.0–3.0, default 1.2) to stop the model chewing on a word, and seed for reproducible output. Defaults are sensible; you'll mainly touch temperature and repetition_penalty if output gets weird.

audio_prompt (optional) is your clone reference - and Turbo is pickier than the standard model: the code warns you if your clip is under 5 seconds. use_cpu and keep_model_loaded round out the optional inputs.

Outputs. audio (AUDIO) and message (STRING) - the same status log as the rest of the pack, telling you the device, cache state and any errors.

Install. Identical to the whole pack - ComfyUI Manager, search "FL ChatterBox", or:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI_Fill-ChatterBox
cd ComfyUI_Fill-ChatterBox
pip install -r requirements.txt

No extra model download beyond what the pack already pulls; Turbo's weights land in ComfyUI/models/chatterbox/ on first use. Dependencies are the pack's shared set (librosa, transformers, diffusers, s3tokenizer, soundfile, …), and the optional resemble-perth watermarking is best skipped on Python 3.12+.

Gotchas. Tags are English-only - fine, since Turbo is English-only anyway. Keep keep_model_loaded = True once you're iterating on delivery, because loading the model is the slow part, not the generation. And remember the 40-second-per-generation cap applies here too; script accordingly.