🎙️ Pocket TTS Generate

Need a voice for a narration, a placeholder dialogue track, or a video that's waiting for its soundtrack - and need it without waking up the GPU or signing up for anything? Pocket TTS Generate is the lowest-friction text-to-speech node in the ComfyUI-Pocket-TTS pack. It wraps Kyutai Labs' Pocket TTS, a 100M-parameter model designed to run on plain CPU, and the whole pitch is in the name: fast, small, and local. No API key, no account, no cloud round-trip.

The eight voices

The built-in voices are named after Les Misérables characters, which is either charming or baffling depending on your tolerance for musical theater. They're speaker archetypes more than an actor roster:

alba (default) - the all-round warm narrator
javert - deep and authoritative
jean - mature, wise
marius - the young guy
fantine - soft and emotional
cosette - young and bright
eponine - energetic
azelma - playful

You're not getting star performances here - think "solid generic voices with distinct registers," not an ElevenLabs audition. For quick narration and prototyping, that's plenty.

How it works

Mechanically it's almost trivial. You give it text and a voice name; the node calls the model's get_state_for_audio_prompt() with the voice name to build the voice state, then generate_audio() renders the speech and hands you back a ComfyUI AUDIO dict (waveform plus sample rate). The model loads once and stays cached, so the first run is the slow one - it also pulls down the ~100M checkpoint on first use - and everything after is warm. The author's own numbers, tested on a MacBook Air M4: roughly 200ms to the first chunk and about 6x real-time on CPU. Two CPU cores and ~400MB of RAM is all it wants.

Inputs and output

Only two things to set, which is the node's entire appeal:

text (STRING, multiline) - what gets spoken. Defaults to "Hello world, this is a test."
voice (dropdown) - pick one of the eight built-ins.

The single output, audio (AUDIO), wires straight into a preview or save node - ComfyUI core has a Preview Audio node, and Video Helper Suite covers saving to a file. From there the audio can go anywhere a voice track goes: into a lip-sync workflow, under a video, or into a batch of different voices run in sequence.

Installing it

Same story as the pack's clone node:

cd ComfyUI/custom_nodes
git clone https://github.com/ai-joe-git/ComfyUI-Pocket-TTS
cd ComfyUI-Pocket-TTS
pip install -r requirements.txt

Or via ComfyUI Manager: search "Simple Pocket TTS", install, restart. The real dependencies are pocket-tts, torch>=2.5, and scipy; torch is almost certainly already in your install. The checkpoint downloads itself on the first generation.

Things worth knowing before you blame the node

"No module named 'pocket_tts'" means the dependency didn't make it into your ComfyUI venv - pip install pocket-tts inside it and restart.
The README mentions precision settings (float32/float16/bfloat16) and a "Model Loader" node. Neither the settings nor the node ship in the actual code - this pack only contains Generate and Clone. Don't go hunting for them.
"Handles infinitely long inputs" is marketing. Feed it reasonable paragraphs and it's happy; don't expect it to narrate a whole audiobook in one shot.
Quality ceiling is the CPU corner. This sits next to Kokoro at the "best voice per watt" end of open TTS, not the Chatterbox-class end that genuinely competes with paid services. If you need cloned or emotionally expressive speech, the pack's Clone node or a heavier TTS suite is the better stop.

It won't replace your ElevenLabs subscription for production work. What it will do is give you a decent voice on whatever laptop you're carrying, in seconds, with zero setup beyond the install. For that job, it's the one to grab.