Cadence GitHub
Roadmap

Built in phases. Each one works before the next begins.

Windows first, with macOS and Linux now shipping under testing. A phase isn't done when the code compiles - it's done when the feature runs, end to end, on a real machine.

DONE Shipped · Jul 2026

The full loop, on Windows

Everything from a blank machine to a finished song works today.

  • Describe a song, or paste your own lyrics - sung in a voice you trained, a generic male/female voice (the music model's own synthesized singer, not a real person), or as an instrumental
  • Songs up to 5 minutes, or an Auto length that matches the lyrics; regenerate just one section of an instrumental and keep the rest
  • Lyrics by Ollama (local, free), Claude, OpenAI, or Gemini - keys in the OS keychain
  • Every song in its own dated folder under Music\Cadence, or any folder you choose, with optional separated stems and WAV/MP3 export
  • Voices you can rename, and delete for real - removing one clears its recordings and its trained model, not just the entry
  • Autotune and smoothing to shape a vocal sung in your own voice, for when the conversion comes out raspy or off-pitch
  • A Library worth returning to: search by prompt, lyrics, or voice, favorite the keepers, and remix any track back into Generate, with a player that keeps going as you move around the app
  • One .msi installer that runs its own engine, checks for updates, and walks you through setup on first launch
  • Download the installer · full changelog
NOW In testing

macOS and Linux

Builds ship for all three platforms. macOS and Linux are under testing.

  • Every release includes a .dmg (macOS), a .deb (Linux), and the .msi (Windows), each built and smoke-tested per operating system in CI
  • A shell setup-backends.sh installs the AI backends on macOS and Linux
  • Real generation on those platforms is still being verified: Linux needs an NVIDIA GPU with CUDA, and Apple Silicon support is unconfirmed. Windows stays the fully tested path
NEXT Planned

Verified, and keeping itself current

Taking macOS and Linux off their training wheels, and taking the work out of staying up to date.

  • macOS and Linux verified on real hardware, generating real songs, so they can drop the "under testing" label
  • Updates that install themselves - today Cadence only tells you a newer version exists and leaves the rest to you
NEXT Planned

More say in how a song turns out

Today you write a prompt and take what comes back. These give you a say in the result.

  • Multiple voices in one song - tag sections with who sings them ([Verse: Anna])
  • Regenerate a section of a song that has vocals, not just an instrumental
  • More shaping for a cloned vocal: brightness, and how closely it holds to your trained voice, alongside today's autotune and smoothing
  • Ask for a tempo and a key instead of hoping the prompt lands on one
LATER Exploring

On the bench

Being explored, not yet promised.

  • Code signing for the Windows and macOS installers, so the operating system stops telling people to think twice - needs a paid certificate, hence "exploring" rather than "planned"
  • Batch generation - queue several prompts and walk away
  • Export stems for tracks already in the Library (today stems save only at generation time)
  • Deleting a track clears its files too, the way deleting a voice already does
  • Reclaim training data left behind by voices deleted before that was fixed
  • A microphone check before voice training, so a bad take is caught at the source rather than in the finished song
Progress, in public

Every phase lands as a commit you can read

Follow the repository to see each milestone as it ships.

View on GitHub Run it from source