VoxCPM Streaming TTS

Let's be honest about the name before anything else: "streaming" here means the generation happens in chunks, not that you get a live playback pipe out of ComfyUI. The node runs VoxCPM's streaming generator internally and then concatenates the chunks into one normal AUDIO that comes out at the end, the same as any TTS node. You still wait for the full run. What you're actually getting is the chunked generation path - which handles long text differently than the one-shot call and, in practice, tends to behave better on long-form speech.

VoxCPM Streaming TTS is otherwise the flexible middle child of the pack: it does plain text-to-speech with no reference, or voice cloning if you give it a reference clip and its transcription. So it's the one node you'd reach for when you want to test both modes, or when long text is being unruly on the one-shot path.

The inputs

• model - from the pack's VoxCPM Model Loader.
• text - what to say.
• cfg_value (1–3, default 2) and inference_timesteps (4–30, default 10) - the same diffusion dials as the rest of the family.
• normalize - normalize the output level.
• prompt_audio and prompt_text (optional) - wire both in for voice cloning. They work as a pair: giving one without the other means the clone silently doesn't happen, and you get plain TTS instead.

Output is audio (AUDIO), straight into VoxCPM Save Audio or anywhere else that eats AUDIO.

Installing it

Same pack, same ritual:

cd ComfyUI/custom_nodes/
git clone https://github.com/DekaNear/ComfyUI-DN-VoxCPM.git
cd ComfyUI-DN-VoxCPM
pip install -r requirements.txt

Restart ComfyUI, or search "DN VoxCPM" in ComfyUI Manager. First run downloads the VoxCPM model through the Model Loader, so the very first generation is slow regardless of which TTS node you picked.

When to use it over the siblings

If you're doing straightforward narration, VoxCPM Text-to-Speech is the simpler choice. If you want cloning, VoxCPM Audio Reference is the more direct one. Streaming TTS earns its place when you're feeding it long text and the one-shot path gives you artifacts or gets twitchy - the chunked approach splits the work, and some people find it cleaner on sustained speech. It's also the natural choice if you ever want to experiment with chunking without adding a whole parallel pipeline. For everyone else, it's a capable fallback that costs nothing extra to keep on the menu.