Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

AI and Generative Tools

Maolan includes tools for generating audio and MIDI from prompts, along with neural analysis and resynthesis for pitch correction. Generated material is created inside an open session and added to the arrangement as a new track.

Audio generation

The audio generator offers two model families:

  • HeartMuLa provides the happy-new-year and RL model options. It generates music from a text or lyrics prompt, with optional style tags.
  • ACE-Step 1.5 provides acestep-turbo and acestep-sft options for prompt-driven instrumental generation. The turbo option also lets you choose a language-model planner size.

The generator can condition output on tempo, key, and time signature. You can also adjust duration and generation settings such as guidance scale and step count. Generated audio is imported into the session as a new audio track.

Generation runs on the CPU or Vulkan backend. Model weights are not bundled with Maolan and may need to be downloaded on first use; model availability and hardware requirements vary by model and backend.

MIDI generation

Maolan offers two prompt-to-MIDI approaches:

  • Text to MIDI interprets prompt details such as tempo, key, time signature, instrument, and style, then creates a MIDI sequence deterministically without loading a neural model.
  • MIDI-LLM uses a Llama 3.2 1B checkpoint fine-tuned with a music-token vocabulary to generate MIDI from a text prompt.

Both options can use tempo, key, and time-signature settings. MIDI-LLM also provides token limit and sampling controls; Text to MIDI provides sequence length and seed settings. Generated MIDI is imported as a new MIDI track.

Neural pitch correction

Pitch correction can analyze a vocal clip with the FCPE pitch detector. In resynthesis mode, Maolan uses the PC-NSF-HiFiGAN vocoder to render audio using the edited pitch contour while retaining the source timbre. Model weights are downloaded when needed. This mode requires a working Vulkan GPU; there is no CPU fallback for the neural pitch models.

Maolan also includes non-neural pitch correction modes. See Keyboard Shortcuts and Gestures for how to open and edit pitch correction in the application.