ace-studio-workflows

Files (1)

• SKILL.md

SKILL.md (first lines)

# ACE Studio workflows

# Song generation

When the user wants a whole song generated, use **Inspire Me**.

## 1. Generate

See `ace-studio-features` for what Inspire Me takes and what it costs.

## 2. Align the project tempo to the result

Do this before any editing — an unaligned project makes every later edit fight
the grid. Which way you go depends on what else is on the timeline, because
**applying a tempo replaces the project's tempo and meter wholesale and moves
existing content.**

- **The generated audio is the only thing on the timeline.** Nothing can be
  disturbed, so apply its tempo to the project directly.

- **The timeline already has other content.** Ask the user first: confirm which
  version of the music they are keeping, and that it is the one the tempo should
  follow. Then apply that one's tempo. Everything already on the timeline will
  shift, so say so before you do it.

## 3. Offer to keep going

Ask whether they want to edit the song. If they do, the first workflow below is
the route in — split it into stems and transcribe the parts they want to change.

---

# Edit a song from audio

When the user hands over a complete piece of music and wants to edit it, work
through it like this.

## 1. Align the project to the audio

Imported audio triggers beat analysis automatically. Apply that analysis so the
project tempo matches the audio's tempo — the operation also snaps the audio's
first downbeat forward to the nearest bar line, which shifts the audio later.

Tell the user the audio will move before you apply, and remember the whole apply
replaces the project's tempo and meter wholesale.

## 2. Split the music into stems

Use the Stem Splitter. Pick the mode from what the user actually needs — see
`ace-studio-features` for what each mode separates and what it costs.

## 3. Turn the melodic stems into MIDI

### 3.1 The lead vocal

Use the built-in Vocal to MIDI. It gives you the melody **and** the lyrics, which
is why it beats a generic transcriber here. It needs a monophonic vocal.

### 3.2 The other instrument stems — Basic Pitch, outside Studio

There is no instrument transcriber on ACE Studio's surface. Use
[Basic Pitch](https://github.com/spotify/basic-pitch), which runs locally and
free, then bring the MIDI back in.

**The stems live on tracks, and Basic Pitch reads files** — so the chain is:
export each stem track as audio → run Basic Pitch on the file → import the
resulting MIDI onto a track.

**Setting it up.** Check for a working environment first; do not install over the
user's existing one. Basic Pitch wants an older interpreter than the newest
system Python, so give it its own virtual environment — Python 3.11 is a
compatible starting point:

```
python3.11 -m venv work/transcribe-env
work/transcribe-env/bin/python -m pip install basic-pitch librosa pretty_midi mido
```

On Windows the interpreter is `work/transcribe-env/Scripts/python.exe`.

```
basic-pitch <output-dir> <stem.wav>
```

**Basic Pitch is a candidate generator, not a transcription.** It is built for
single-instrument material, which is exactly what a stem is — but the output
still needs cleaning before it is worth importing. Running it several times and
getting the same answer is not evidence; it is the same model twice.

### 3.3 Clean the MIDI before importing

Remove what the recogniser invented:

- **Harmonic stacking** — phantom notes an octave or a fifth above the real one.
- **Ultra-short noise notes** and **tail fragments** left by decay.
- **Duplicate onsets** on the same pitch at the same moment.

Keep what is musical: ornaments, syncopation, and rests are content, not noise.
**Judge against the actual tempo and phrasing — do not delete every short note
with one fixed millisecond threshold**, or you will flatten the performance.

Then check per instrument:

- **Bass.** Octave errors are the common failure, and slides come back as dense
  runs of semitones. Cross-check the root against the kick and the chord changes.
  Keep it monophonic by default: bleed from guitar or keys must not become bass
  chords.
- **Guitar.** Distorted guitar transcribes badly. Check short phrases against the
  spectrogram, the onsets, and the chord underneath.
- **Keys.** Sustain pedal and tail resonance come back as masses of held notes —
  do not transcribe them literally. Reconstruct the two-hand division from the
  bass and the harmony.
- **Drums. Do not use Basic Pitch.** It is a pitch model; percussion is out of
  scope for it.

### 3.4 Load sound sources

Load a suitable sound source onto whichever tracks the user wants to edit. See
`ace-studio-features` for choosing between AI Instruments and plugins.

From BeatMagic/acestudio_agent_plugin, MIT.