📣 Generate Audio

Generate Audio is Mana Nodes' text-to-speech node: type some text, get a spoken .wav back. Under the hood it runs Bark, Suno's generative TTS model, which is what gives it a party trick most TTS nodes don't have - it can laugh, sigh, sing, switch languages mid-sentence, and do sound effects, not just read text flatly.

The honest take first

The author labels this node experimental, and lists "Improve Text to Speech" on the pack's to-do list. Take that at face value. Bark is genuinely fun and expressive, but it's also slow, non-deterministic (the same text can come out sounding different run to run), and prone to going off-script - swallowing words, adding noises you didn't ask for, or drifting speakers. The ComfyUI audio space has since filled up with other TTS options (Whisper Speech, ParlerTTS, and newer voice models), so if you need reliable, clean narration this may not be the node you settle on. If you want characterful audio and don't mind rerolling, it's a good time.

How Bark makes it interesting

Bark treats speech and sound as one continuous thing, so you steer it with inline tags rather than sliders. Some of what it understands:

• Non-speech cues in brackets: [laughter], [sighs], [gasps], [clears throat], [music].
• ♪ ... ♪ around lyrics to nudge it toward singing.
• CAPS for emphasis on a word.
• MAN: / WOMAN: / NARRATOR: speaker hints (respected sometimes, not always).
• Code-switching - mix languages in one prompt and it'll attempt each with a fitting accent.

So a prompt like [clears throat] Hello - and, uh, I like pizza. [laughs] produces something with actual personality.

The inputs and output

There are only two inputs, which tells you how much of the behavior lives in the text itself:

• text - multiline. Your script, plus any Bark tags above.
• filename_prefix - where the .wav is written (default audio\audio).

The output is audio_file - a string path to the generated WAV - which you can hand to Combine Video as the audio track, so text-to-speech narration can drive a captioned clip end to end.

Installing it

It's in the Mana Nodes pack. ComfyUI Manager → search Mana Nodes → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/ForeignGods/ComfyUI-Mana-Nodes.git

Install requirements (pip install -r requirements.txt; the python_embed variant on Windows portable) and restart. Big heads-up: Bark downloads several gigabytes of model weights on first use from HuggingFace. The first generation will sit there for a while pulling models - that's the download, not a freeze - and you'll want the disk space and a decent chunk of VRAM for it to run comfortably.

Common issues

Set expectations on speed and consistency. Bark is not fast, and it's not repeatable - if a take comes out garbled, mispronounced, or with a random noise, the normal move is to run it again rather than hunt for a setting, because there aren't many settings to change. Long text is where it wanders most; feed it a sentence or two at a time and concatenate, rather than a whole paragraph.

If the first run seems hung, it's almost certainly the multi-GB model download - let it finish. And because this is the experimental corner of the pack, don't be surprised if it's the node that's rough around the edges; the text-animation side is where Mana Nodes shines, with TTS clearly a bonus feature the author is still working on.