Controlled change
Arrangement, vocal delivery, and mix density vary independently while lyrics, tempo, and the rest of the brief stay fixed.
Cardona CT Lab / Living product evidence
SoundCraft
A local song pipeline built around reviewable decisions.
SoundCraft turns an original song brief into controlled generation takes, vocal direction, listening notes, section candidates, estimated stems, and a portable remix package. This page shows the pipeline, the interface, the outputs, and the boundaries that still require human judgment.
Current signal path
- 01Brief + lyricsOriginal intent
- 02Controlled takesOne variable
- 03Blind reviewHuman promotion
- 04Remix packagePortable assets
Local runtime / Review required
Research snapshot / 2026-10-09
The song stays a set of decisions.
A generated mix can hide how it was made. SoundCraft keeps prompts, lyric versions, vocal roles, take ratings, section requests, and export metadata visible so a creator can compare, reject, revisit, or remix.
Human promotion
Blind labels reduce attachment to generation order. Ratings and notes record why one take advances.
Portable evidence
Remix Pack 2.0 carries the mix, markers, loops, review records, profiles, candidates, and available stems.
Pipeline walkthrough
From sentence to session
Each stage produces something a person can inspect before the next stage earns attention.
- 01
Intent
Write the brief and version the lyrics
Define the musical promise, tempo, structure, vocal roles, and an original lyric draft. The lyrics workspace saves revisions and runs deterministic line, syllable, rhyme, repetition, and cue analysis before generation.
- Input
- Concept + sectioned lyrics
- Output
- Versioned writing package
- Authority
- Writer
- 02
Direction
Set a reusable vocal profile
Vocal Director captures register, tone, aggression, clarity, layer count, and backing-role notes. A profile can be saved with the project and applied consistently to later takes or section requests.
- Input
- Performance intent
- Output
- Project vocal profile
- Limit
- Direction is not identity control
- 03
Generation
Queue a controlled take set
Take Lab creates two or three sequential variants for one selected dimension. Seed changes are recorded, GPU work remains serialized, and the accepted prompt decisions stay fixed.
- Engine
- ACE-Step 1.5
- Output
- Mixed song takes
- Rule
- One changed dimension
- 04
Evaluation
Listen blind, rate, annotate, promote
Variants receive neutral labels before review. A person listens, records a rating and note, then promotes the preferred take. SoundCraft stores that decision in the session instead of treating the latest render as the winner.
- Evidence
- Rating + review note
- Output
- Promoted take
- Authority
- Listener
- 05
Revision
Generate section candidates beside the original
Mark a section, describe the change, set its intensity and transition intent, and generate a candidate as a separate take. Align it on an alternate playlist and promote only after the edit is heard in context.
- Input
- Range + local instruction
- Output
- Separate candidate take
- Limit
- Manual alignment required
- 06
Separation
Create estimated working stems
A selected mix can be separated into vocals, drums, bass, and other with TorchAudio Hybrid Demucs. These are practical remix sources, labeled as estimates because source separation can leave bleed and artifacts.
- Method
- Hybrid Demucs
- Output
- Four estimated stems
- Limit
- Not multitrack originals
- 07
Handoff
Export Remix Pack 2.0
The export groups the mix, available stems, selection loops, bar markers, review records, vocal profiles, section candidate requests, and a machine-readable manifest for the next editing pass.
- Format
- WAV + JSON manifest
- Output
- Portable remix folder
- Goal
- Future edits stay possible
Interface evidence
The review loop is visible in the tool.
These are current application captures from the implemented workflow.
Capture 01
Controlled variants and blind review
Take Lab keeps the active dimension explicit, queues a bounded set, and separates listening notes from generation settings.
- Two or three takes per set
- Sequential local GPU jobs
- Blind labels, rating, note, and promotion
Capture 02
Vocal direction stored with the project
The profile records the intended performance shape. It guides future prompts without claiming exact singer or identity control.
- Register, tone, aggression, and clarity
- Layer count and backing-role notes
- Save, recall, and apply across requests
Capture 03
Section work remains a candidate
A local revision request creates an alternate take and a saved record. It does not silently overwrite the accepted timeline.
- Marked section range
- Intensity and transition controls
- Playlist alignment before promotion
Listening lab / Original demos
Three briefs, three production problems
Each 45-second example was generated locally from an original concept and lyric package. The controls play one track at a time.
01
Alternative post-hardcore / 184 BPM
Cut the Current
Interlocking rhythm guitars, a distinct lead riff, moving bass, double-kick bursts, and a half-time breakdown. The performance brief pairs a melodic, raspy female lead with brief low male response shouts.
45 sec256 kbps MP3Original lyrics
02
Abrasive noise-pop / 138 BPM
Candy Glass
Octave fuzz, controlled feedback, blown-speaker bass, bit-crushed percussion, abrupt mutes, and a hard tape stop push a bright pop hook into a damaged texture field.
45 sec256 kbps MP3Original lyrics
03
Dark electro / 130 BPM
Borrowed Face
An analog bass sequence, distorted arpeggiators, gated drums, glassy counter-melody, synth brass, and sparse guitar stabs frame a low female verse and a forceful chorus.
45 sec256 kbps MP3Original lyrics
No example is playing.
Remix Pack 2.0
The export is a continuation point.
The song leaves SoundCraft with enough context for a later session to understand what was generated, what was selected, and which assets are derived.
Read the technical workflow on GitHubsoundcraft-remix/
- mix.wavPromoted full mix
- manifest.jsonSchema 2 decisions and provenance
- Stems/Available source and estimated stems
- Loops/Selection and marker-based WAVs
- Reviews/Take ratings and notes
- Profiles/Vocal direction records
- Candidates/Section revision requests
Operating boundaries
What the current pipeline does not prove
Clear limits keep the useful parts credible and define the next engineering work.
Vocal identity
Role tags and profiles guide delivery. They do not guarantee a specific singer, consistent identity, or reliable assignment of every backing response.
True source stems
Demucs outputs are estimated from a stereo mix. Bleed and artifacts can remain, so the files stay labeled as estimated.
Interval replacement
The ACE-Step bridge generates a separate candidate. It does not patch a time range inside an accepted mixed file.
Musical judgment
Signal checks can verify duration, format, peak, and file integrity. They cannot decide whether a hook, voice, transition, or mix feels right.
Release evidence
Verified before publication
The current sprint was exercised as software and as a packaged desktop application.
- Repository CI
- PassedFormat, clippy, native tests, doctests, assets, dependency layers, and WASM
- Audio bridge tests
- 12 passedPython bridge and deterministic lyric-analysis coverage
- Remix workflow tests
- 5 passedReviews, profiles, section candidates, and export behavior
- Packaged application
- Smoke testedInspect, profile save/query, and candidate save/query commands
- Demo provenance
- RecordedACE-Step engine and model revisions, settings, lyric packages, and hashes