CSM Text-to-Speech

"Text-to-speech" undersells this. It makes you think of a robot reading a manual, and this is not that. CSM Text-to-Speech runs Sesame's CSM-1B model - the open weights behind the "Maya" voice demo that went around in early 2025 and made everyone ask how TTS could sound that human. No API, no key, no cloud. A couple of gigabytes of weights, running in your own ComfyUI graph.

This is the node you reach for when you want a workflow that ends in spoken audio instead of pixels - narration over a video, dialogue for a scene, a character voice for a comic. It's one of the genuinely impressive local TTS models out there, and it pairs with the context version for voice cloning.

How it works

CSM-1B is small by LLM standards: a Llama-3.2-1B backbone plus a 100M decoder. Instead of emitting text tokens, it emits audio tokens - 32 parallel codebooks from Meta's Mimi codec, which get decoded back into a 24 kHz waveform. The pack's code builds the model with torchtune, loads ckpt.pt in bfloat16, and wraps it in a cached Generator. First generation after a restart is slow because the model has to load (it also pulls the kyutai/mimi codec from HuggingFace once); after that it stays cached in memory, so subsequent runs are fast.

The inputs that matter

• model_path and tokenizer_path - feed these from Load CSM Checkpoint and Load CSM Tokenizer rather than hand-typing paths. The node validates them and throws if they don't point at a real file.
• text - what the voice says. CSM loves natural phrasing, commas, even "um"s; it reads well-written text much better than flat sentences.
• speaker - an id from 0–10 that gets embedded as a speaker token. With no context clip this nudges the timbre rather than cloning anyone. Real voice control comes from the context variant.
• temperature / topk - sampling knobs. Defaults are 0.7 / 100. Lower both and the voice gets steadier and more monotone; raise them and it gets more expressive and more likely to go off the rails.
• max_audio_length_ms - hard cap on output length, default 90 seconds.

One input to not touch: sample_rate. CSM always generates at 24 kHz regardless - this node just stamps the value you pick onto the output. Change it to 48000 and you've mislabeled a 24 kHz file, which plays back wrong in editors. Leave it at 24000.

Outputs

A single audio output in ComfyUI's AUDIO format - a {waveform, sample_rate} dict that wires straight into any audio save/preview node.

How to install

It's a normal custom node pack:

cd ComfyUI/custom_nodes
git clone https://github.com/thezveroboy/ComfyUI-CSM-Nodes
pip install -r ComfyUI-CSM-Nodes/requirements.txt

(Or use ComfyUI Manager and search "ComfyUI-CSM-Nodes".) Then restart ComfyUI. The heavy dependencies are torch, torchaudio, transformers, tokenizers, torchtune, and torchao - the same stack as most local audio work.

The models are the fiddly part:

• Download ckpt.pt from sesame/csm-1b into ComfyUI/models/sesame/ (create the folder - it won't exist).
• Download the Llama-3.2-1B tokenizer files (e.g. from unsloth/Llama-3.2-1B) into ComfyUI/models/sesame_tokenizer/<some-dir>/. The tokenizer loader picks subdirectories, so give it its own folder.

Common issues

• Russian-language errors. The pack's checks throw messages like Текст пуст! (text is empty) or model_path некорректен! (invalid model path). If you see Cyrillic, an input check fired - usually a path node not wired.
• First run is brutally slow. That's the model loading plus the Mimi codec download. Subsequent generations are fast. Don't blame the node.
• bfloat16 on an older GPU. The code loads weights in bf16 unconditionally, which needs an Ampere-or-newer GPU. On older cards you'll hit errors rather than slow speed.
• Long prompts bump into CSM's 2048-token window (input too long error). Keep text short-ish, especially when feeding context.