🎵 ACE-Step 1.5XL TTS-like Voice Prompt

If you've ever stared at ACE-Step's lyrics format - the [verse] / [chorus] section markers, the line-by-line structure - and wished you could just paste a paragraph of narration and be done, this node is for you. TTS-like Voice Prompt takes a plain script and a couple of musical settings, and outputs a fully-formed set of tags and lyrics plus all the musical parameters the text encoder wants. It's the "make this sound like someone talking/singing a script" workflow's brain.

The honest name matters here: this is TTS-like, not TTS. ACE-Step won't read your script back with a strict voice - it reinterprets timing, melody, pronunciation and phrasing, which is exactly what you want for voice-guided narration or spoken-word tracks, and exactly wrong if you need word-perfect speech. Pair it with a reference voice sample (Reference Audio) and it behaves like a loose, musical voice-clone pipeline.

How it works

The node formats, it doesn't generate. It splits your script into lines (on newlines; a single paragraph gets split on sentence-ending punctuation), groups them into blocks of lines_per_section, and wraps each block in a [section_label] marker - [verse], [chorus], [spoken] or [narration]. If your script already has [section] markers, it passes it through untouched, which is the polite thing to do. The default script is a Mandarin Chinese monologue, and the default voice_tags are tuned for "clean lead vocal, lyrics forward, minimal accompaniment" - the exact profile that works with a reference voice.

The inputs and outputs that matter

The few you actually touch:

• script - your raw text. No section markers needed; it formats them for you.
• voice_tags - the style prompt. Swap the Mandarin defaults if you're not doing Chinese.
• section_label and lines_per_section - how the script gets chunked.
• bpm, duration, timesignature, language, keyscale - the musical frame, passed through untouched.

Outputs are tags, lyrics, bpm, duration, timesignature, language, keyscale - and the whole point is that they plug directly into ACE-Step 1.5XL Text Encode's matching inputs. No retyping, no transcription errors between your plan and the encoder.

Install

ComfyUI Manager ("ComfyUI-ACEStep") or:

cd ComfyUI/custom_nodes
git clone https://github.com/starsFriday/ComfyUI-ACEStep.git

then restart. No pip deps of its own; the heavy lifting is the ACE-Step 1.5 XL models (acestep_v1.5_xl_turbo_bf16.safetensors, qwen_0.6b_ace15/qwen_4b_ace15 CLIP, ace_1.5_vae.safetensors) from the ACE-Step repo, and XL wants ~12GB+ VRAM.

Gotchas

Temper your expectations on the vocals themselves. The community consensus on ACE-Step is blunt: instrumentals are genuinely good, vocals and lyrics are the weak suit - so this node gets you a singable script, not a polished singer. Keep your reference clip to a clean vocal (10–30 seconds beats a full mix), keep generate_audio_codes off on the text encoder when the reference voice should lead, and expect to iterate. The other trap is forgetting that duration here just flows through - if it disagrees with the latent length in your sampler, you'll hear the mismatch as truncated output. Let this node's duration output be the single source of truth.