⚙️ Dots TTS Engine

Dots TTS is the speed pick among the newer engines. It covers a wide spread of languages with official auto-detection, and - the useful bit - it ships distilled variants (SOAR and MeanFlow) that generate in far fewer steps than a normal diffusion/flow model, which means faster turnaround without a huge quality hit. At ~6GB it's mid-weight. If you're iterating on multilingual lines and don't want to wait, it's a sensible engine to have wired up. This node configures it and outputs a TTS_ENGINE for 🎤 TTS Text or 📺 TTS SRT.

How it works

Dots is a flow-matching TTS (hence num_steps and guidance_scale). The trick is the distilled checkpoints: dots.tts-soar and dots.tts-mf (MeanFlow) are trained to produce good audio in a handful of steps, where the plain dots.tts-base needs more. That's why the default step count here (10) is so much lower than the 32-ish you see on other flow engines - the distilled variants are built to run short. Language handling is official rather than bolted-on: it auto-detects or you set it explicitly.

The inputs and outputs that matter

• model_variant (dots.tts-soar, dots.tts-mf, dots.tts-base, default SOAR) - SOAR and MF are the fast distilled variants and the reason to pick this engine; base is the full model. Start on the default.
• language (default Auto, 24 languages) - Auto detects from text, or pin one. There's also a None option to skip language conditioning entirely.
• num_steps (1–64, default 10) - generation steps. The low default suits the distilled variants; if you switch to dots.tts-base you'll likely want more steps for clean output.
• guidance_scale (0–5, default 1.2) - prompt/conditioning adherence. Modest by default; nudge up if delivery drifts from the reference.
• speaker_scale (0–5, default 1.5) - how strongly it locks onto the reference speaker identity. Raise it if the clone feels loose, lower it if it sounds forced.

max_generate_length caps output length in tokens. The optionals - template_mode (TTS vs Instruction TTS, for instruction-style prompting), precision, normalize_text, and optimize (a speed switch) - are refinements you can leave at defaults. Output is a single TTS_engine.

Installing it

Part of TTS Audio Suite. Easiest: ComfyUI Manager → search TTS Audio Suite → install → restart, which runs the pack's install.py and handles the dependency conflicts (NumPy, librosa, s3tokenizer) and Python 3.13. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/diodiogod/TTS-Audio-Suite.git, then python install.py from inside the folder with your ComfyUI venv active (run the script, not just a bare requirements install). Linux: portaudio19-dev libsamplerate0-dev first. The Dots checkpoints (~6GB, with tokenizer, vocoder, and speaker encoder) auto-download into ComfyUI/models/TTS/dots_tts/ on first use.

Common issues

• Base variant sounds rough at 10 steps. The low default step count is calibrated for the SOAR/MF distilled models. If you switch to dots.tts-base, raise num_steps accordingly - the distillation is exactly what lets the others run short.
• Weak or drifting clone. Bump speaker_scale (toward 2–2.5) so it holds the reference identity harder, and make sure you're feeding a clean reference via 🎭 Character Voices.
• Wrong language pronunciation. Auto usually gets it, but on short or ambiguous text it can guess wrong - pin language explicitly.
• Not sure it's worth it over F5/ChatterBox. Its edge is speed on multilingual work via the distilled variants. If you're English-only and happy with F5's pace, you may not need it - but for fast iteration across languages it's a nice option.