⚙️ Echo-TTS Engine

Echo-TTS is the diffusion-based option in the suite - a DiT model that does English voice cloning from a reference clip. Two things make it its own animal: it's built to shine on short chunks (roughly 30 seconds or less per generation), and it has an explicit control for a problem that plagues chunked TTS - the voice slowly drifting into a different person across a long passage. That control is "Force Speaker KV," and it's the whole selling point.

Standard drill: this is a config node. It outputs a TTS_engine you feed into 🎤 TTS Text or 📺 TTS SRT.

How it works

Being diffusion, Echo generates over a number of denoising num_steps rather than autoregressively, and it uses classifier-free guidance with two separate scales - one steering toward the text, one toward the speaker's identity. When you stitch several chunks together for longer text, later chunks can lose the reference speaker's character; force_speaker_kv pins the speaker's key/value attention state so identity holds across the seams. There's a stack of presets in the node precisely because this model has a lot of dials, and the presets encode known-good combinations so you don't have to guess.

The inputs and outputs that matter

If you only touch a few things, touch these:

• preset - start here. Custom exposes everything; the Independent-*-CFG presets are ready-made recipes (high/low CFG, with or without the "Flat" schedule). Try a preset before hand-tuning.
• num_steps (default 40) - diffusion steps. More = potentially cleaner but slower; 40 is a sane middle.
• cfg_scale_text (3) and cfg_scale_speaker (8) - how hard it pushes toward the words vs. toward the target voice. If cloning is weak, the speaker scale is your lever.
• force_speaker_kv (off by default) - flip it on for multi-chunk / longer text to stop the voice from wandering. speaker_kv_scale and its companions tune how aggressively.

The rest - truncation_factor, rescale_k, rescale_sigma, cfg_min_t/cfg_max_t, sequence_length - are exactly the kind of internals the presets exist to set for you. Leave them alone until you have a specific reason.

Output is TTS_engine → TTS Text or TTS SRT.

Installing it

Ships in the pack. ComfyUI Manager → "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

First generation auto-downloads ~7.1GB (base model + DAC codec) to ComfyUI/models/TTS/echo-tts-base/.

Common issues & troubleshooting

Check the license before you ship anything. This is the big one. Echo-TTS weights are CC-BY-NC-SA - non-commercial. The suite's own code is MIT, but each engine keeps its own terms, and Echo's are not commercial-friendly. If you're making something you'll sell or monetize, use a different engine (ChatterBox and the Chatterbox Multilingual model are MIT, for instance).

It's painfully slow. Echo runs but crawls on CPU. Use CUDA for any real work - the README calls CPU "very slow for real workloads," and that's not an exaggeration for a diffusion model.

The voice changes partway through long text. That's the drift Echo is designed to fight. Turn on force_speaker_kv and, if needed, raise speaker_kv_scale. Also just keep chunks short - Echo is happiest at ~30s or less per generation, so let the unified chunking split long scripts rather than forcing giant single passes.

English only. Don't expect other languages to come out right - Echo is an English model. For multilingual cloning, look at Higgs v3, CosyVoice, or the Chatterbox Multilingual model.