Improvements
The Mac version runs on macOS 14 Sonoma and later. Liquid Glass styling is used on macOS 26.
The LoRA pack (about 380 MB) is no longer downloaded when the engine starts. WavePrompt asks before downloading it the first time Push, Arranged or a LoRA is used, and the LoRA submenus have a Download LoRA Pack item.
New Features
LoRA support. The Push option strengthens the chosen genre or voice with a LoRA trained on it, for 14 genres and for male and female voices.
The Genre and Voice menus have a LoRA submenu with Dream Pop, Old School Hip-Hop, Male and Female.
Import LoRA in the LoRA submenu adds a YuE2 LoRA from a Hugging Face or download link, or from a .safetensors file. Imported LoRAs can be removed from the same submenu.
Instrumental songs have an Arranged option that plans an intro, verses, choruses and an outro before rendering.
Improvements
Sung vocals in songs stay on the beat: each vocal take is matched to the song tempo before it is placed on the backing.
The LoRA pack (about 380 MB) downloads with the models the first time the engine starts.
New Features
Import Memo in the piano roll turns a hummed or sung recording, such as an iPhone Voice Memo, into a melody of up to 16 bars. The tempo and key are set to match the recording, and the melody can then be rendered as a loop, vocal phrase or full song.
Improvements
Sung vocals in songs stay in time with the backing. Vocal takes are no longer rendered again for lines that start slightly early, which could replace a good take with a worse one.
Songs with lyrics render faster: the lyric check takes a few seconds instead of minutes.
New Features
Songs have an Auto length that fits the song to its lyric lines. Instrumental songs on Auto keep the planned structure instead of being cut to a set duration.
Improvements
The Key setting is now followed: the planned score is moved to the chosen key for loops, vocal phrases and songs.
Sung vocals in songs start on the right bar when the backing begins mid-bar.
The vocal melody follows the chords the backing actually plays, so lines stay in tune when the backing skips or repeats a bar.
Vocal takes that sing a line before its bar are rendered again.
Lyrics in a language other than the chosen one are checked correctly instead of failing every take.
Fixes
Fixed an error when a song's backing came out shorter than planned.
Fixed the vocal melody of songs being placed after the end of the song in MIDI exports.
New Features
Custom score planning with a piano roll. Draw a melody and pick a chord for each bar, or generate them in the chosen key with density, register and chord options. Loops are rendered on your melody and chords, songs repeat it through their sections, and vocal phrases follow its pitches.
The piano roll plays the melody and chords back in a loop, and can be transposed, snapped to the key, repeated, undone and exported as MIDI.
Every result now has a MIDI file of its score next to the audio, with melody, chord, tempo, key and section marker tracks.
Library export to MIDI, and to a folder with the WAV, AAC and MIDI files together.
Improvements
Sung vocals in songs keep a steady level, and quiet backing sections are raised.
Multi-bar loops have more bars to choose the loop from, and short renders are cut from their last full bars.
A description that names a genre but no instruments keeps that genre's instruments.
Tempos typed into the description are ignored in favour of the Tempo setting, and genre spellings such as "dnb" or "lofi" are recognised.
Negative tags such as "no fills" are left out of the prompt, since the model tends to read them as asking for the thing.
Rhythm-only loops no longer contain a melody or stray voices.
Improvements
Loops render faster. Generation stops as soon as a steady groove has played, and score planning stops once it has enough bars for the loop.
Short loops start on a full groove bar instead of the build-up or a fill.
Song lyrics open in a full-size editor next to the library. Hovering over the lyrics line in the sidebar shows them in a popup.
The Words to Sing box grows when you click into it.
New Loop, Vocal Phrase and Song tabs with a highlighted active tab.
15 second song length.
Bug Fixes
Tempo matching no longer makes loops fade out over time.
Synth leads are no longer mistaken for vocals and removed from instrumentals.
The sidebar no longer changes width when switching tabs.
Improvements
The description is now the first part of the style prompt and replaces the genre's default instruments, so it has a stronger effect on the result.
Instrumental loops and songs are checked for singing, and any vocals found are removed.
File menu options to choose the output folder, return to the default folder and show it in Finder.
Models are now stored in ~/Music/WavePrompt/Models and can be moved to another folder or drive from the File menu. Models from 1.0 are moved there on first launch.
Sound names and styles in the library wrap to the window width.
Faster batched decoding. Vocal takes of songs are rendered only when needed.
New app icon.
Bug Fixes
The 808 and hi-hats instrument set no longer asks for trap hi-hats.
Vocal takes of songs are no longer tempo matched a second time.
First release. WavePrompt generates loops, vocal phrases and full songs from a text prompt on Apple Silicon Macs, using the YuE2 music model. Everything runs locally.
Loops
Bar-accurate loops from 1 to 16 bars at 60 to 180 BPM, in any major or minor key and in 4/4, 3/4, 6/8, 5/4 or 7/8.
Sample-accurate loop starts with nudging, so loops line up with a DAW grid.
Vocal Phrases
Short sung phrases from your lyrics in English, Chinese, Japanese, Korean, Spanish, French, German, Italian or Portuguese.
Eleven voice types including breathy, powerful, deep, raspy, falsetto, whisper, choir and rap.
Optional vocal isolation with Demucs.
Songs
Full songs from 30 seconds to 4 minutes, with vocals or instrumental.
Songs end on an outro. Endings are trimmed to whole sections, and the intro is reused when no outro was planned.
Songs with lyrics are rendered as a backing track plus a fitted vocal. Lyrics are checked with Whisper.
Style
34 genres, 14 moods and 11 instrument sets, plus a style strength setting for how closely the audio follows the prompt.
Shuffle and reset buttons for the style, voice, timing and advanced sections.
Fixed seeds to repeat or vary a result.
Library and Export
Every render is kept in the library with its recipe. Reuse in current type carries a sound's seed and style to another tab.
Export to WAV (48 kHz, 24-bit) or AAC.
Tempo matching with the Rubber Band Library.
Move to Trash can be undone.
Engine
Generation runs on the Apple GPU through Metal. No server, no account and no upload.
Batched decoding, with the model kept in memory between renders.
Progress is reported for each stage.
Works offline after the first model download.