Skip to content

Song workflows and media

View Markdown
For agents

Give your agent this page’s Markdown link, or choose a collection.

DAWN can draft a project, prepare local audio and keep a longer song job resumable. Start from an existing project and inspect each command’s help before adding an optional runtime.

dawn make --seed <n> makes a repeatable draft. dawn make --from <chart.json> performs an authored chart. dawn song --list lists the seeded composer’s styles; dawn song writes a project from one of them. A natural-language local-model draft needs a configured local model.

Terminal window
dawn make --help
dawn song --list

Use the resulting TypeScript as editable source, then run dawn check and render a review file. Keep an existing song’s source committed or copied before making a new project in its place.

dawn song --palette <palette.json> plays the composed parts with a palette’s sounds; a <collection>/<entry> installed with dawn collection install works too. The notes stay the composer’s. Build a sound palette explains which sound plays each part and how levels are matched.

The workflow registry lists durable procedures, including original songs and reference studies:

Terminal window
dawn workflow list
dawn workflow read original-house

dawn make --from <chart.json> creates and runs an original-house job; there is no dawn workflow start, which only queued a job that nothing ran. status inspects a record, and cancel and resume mark it cancelled or runnable again. dawn album plans, makes and gates an album with per-stage caching and retained hand edits, and makes a song that fails its gates again. Its help describes the brief and output directory:

Terminal window
dawn album --help

Local-model stages need an inference runtime and model weights. Set those up separately and account for their disk, memory and compute needs.

For local-model words, --words-best <1-5> compares several candidates; the default is one. --words-sample <n> sets the initial sampling seed. With sung vocals, candidates are sung, rendered and transcribed before selection, so additional candidates need more compute and the singing and transcription runtimes.

Stock singing uses the optional local Kokoro and WORLD runtime, and DAWN checks a sung line by listening to it with a local speech recognizer. One command looks for both, and one installs them:

Terminal window
dawn setup # what is here, what is missing and how large it is
dawn setup --install # fetch the missing pieces, about 1.9 GB from nothing
dawn setup --test # sing a phrase and hear it back
dawn vocals doctor
dawn vocals sing --help

dawn setup --install needs nothing installed first: it fetches its own pinned copy of uv when the machine has none, builds the voice’s Python environment and DAWN’s own speech recognizer, and fetches the Kokoro and whisper models. The voice builds a package or more from source, so it needs a C++ compiler (xcode-select --install on a Mac, build-essential on Debian and Ubuntu); setup says so before it starts. An install ends by singing a short phrase and listening to it, and dawn setup --test does that at any time: it exits 1 when the words do not come back. dawn vocals setup prints the same steps as commands to run yourself and installs nothing. dawn album setup is the same command as dawn setup under its first name; use it when your DAWN has no dawn setup yet. Keep generated takes and their records in the song project so they remain available for a later render. Rendering prepared takes does not require regenerating them.

dawn reference prepare accepts a local audio file or a supported video URL and retains source, stems and timing evidence for a study. dawn reference recreate uses the prepared manifest; dawn reference setup describes optional dependencies.

Use recordings you are authorized to process. A URL-based study downloads media and needs network access and the configured helpers. Resulting manifests and derived files stay in the project.

Terminal window
dawn media doctor
dawn media --help
dawn media chop --help

Media commands cover download, stem separation, analysis, audio-to-note extraction, sampling, transcription and chopping. Helpers can use a configured local StemDeck service or tools such as FFmpeg, yt-dlp, Demucs, Basic Pitch and whisper-cli. doctor reports what is available before a longer job.

Chop operations write derived WAVs with sidecar records and retain inputs. Inspect the output and keep the sidecar with the audio. The CLI reference covers checks, exports and measurements after preparation.

Documentation drafted with AI and checked against the product source.