Legal
Client work, AI and training data
Can you deliver it to clients? Independent MIDI: yes, like a licensed MIDI pack. Audio: for sketches and mockups unless your client agrees. Plus how the AI works and what it learned from.
Version 1.0 · Last updated 29 September 2026 · Powered by Stability AI
Can I use it in client work?
Yes. Composer Assistant gives you three kinds of take, and each has its place:
| What you get | Sketches, mockups, demos | Your own releases | Finished client work |
|---|---|---|---|
| Independent MIDI | Yes | Yes | Yes — like a licensed MIDI pack |
| Audio takes | Yes | Yes | Only if your client agrees in writing |
| MIDI of the audio | Yes | Yes | Only if your client agrees in writing |
- Independent MIDI isn't AI. A rule-based engine builds it from patterns in our own MIDI packs, written by composers, fitted to your chords, key and rhythm. Edit it, run it through your own libraries and deliver it in finished client work, as you would a licensed MIDI pack. It's licensed to you, not owned — what you add is yours. If a client contract asks you to warrant that everything is entirely your own original work, check that clause, as you would with any MIDI pack or loop. If Independent MIDI ever starts using AI, we'll say so here first.
- Audio takes and "MIDI of the audio" are AI-generated. ("MIDI of the audio" is read from the AI audio by an AI model.) Many clients, studios, music libraries and distributors don't accept AI-generated material yet, or require you to disclose it, and the law is still settling. Use them to sketch and mock up fast, and deliver them to a client only if they've agreed in writing.
So you get both: fast mockups from the audio, and finished work from the Independent MIDI.
In plain English
- The audio takes are made by AI. Our AI models generate them on our servers, following your idea and your settings.
- Independent MIDI is not AI. A rule-based engine writes it from your settings, using patterns from our own MIDI packs. "MIDI of the audio" is read back from the generated take by a small AI model we trained.
- How a take is made. You choose one of our instruments — chords, motors, ostinatos or percussion — and give it your idea. Its model is Stability AI's Stable Audio 3 Medium, which provides the underlying audio engine, plus the layers we trained only on our own music: Cinematic Composing's MIDI packs, rendered with our own sample libraries. Nothing scraped from the internet.
- Not an open text-to-music tool. Every request is for a part from one of our instruments. The optional description only adds words to that request.
- What the base model learned from. Stability AI pre-trained Stable Audio 3 Medium on about 1.28 million recordings, licensed from AudioSparx and shared on Freesound under Creative Commons licenses.
- Never your music. We never use your files, recordings, MIDI, descriptions or takes to train our models. The training we do uses only music written by professional composers and carefully curated for it. If that ever changed, we would ask your permission first.
- What this means for your music:
- Your own music — yes. Use takes in your mockups, your demos and your own releases, including music you sell.
- Client work: Independent MIDI yes; audio takes and "MIDI of the audio" only if your client agrees in writing. See Can I use it in client work? above.
- Copyright: parts of a take made by AI may not be protected by copyright, and takes aren't exclusive. The music you write yourself stays yours.
- Always be honest about AI. Never present a take as played by real musicians, and tell clients, distributors and registries when AI-generated takes are in your work.
How Composer Assistant uses AI
| Part | What it does | Is it AI? |
|---|---|---|
| Instrument models (chords, motors, ostinatos, percussion) | Generate the audio takes from your idea and settings | Yes. Stability AI's Stable Audio 3 Medium as the audio engine, plus layers we trained on our own music, one set per instrument |
| MIDI of the audio | Reads the notes back from a generated take | Yes. A small model we trained |
| Independent MIDI | Writes MIDI from your settings | No. A rule-based engine working from Cinematic Composing's MIDI packs |
| Input analysis | Reads the notes, pitch and rhythm of audio you drop in or record | Yes. Published transcription models by others (listed on our Third-party notices page); we didn't train them |
All of it runs on our servers. Nothing AI runs on your computer. In Advanced, Style/Color sets how much of the character we trained is applied (1.0 = as trained).
Training data documentation
This section is our training-data documentation under California's AB 2013 (California Civil Code §3111). It covers the models Composer Assistant uses to make your takes. We update it before we release a new or retrained model.
Models covered
| Model | Trained | Available to the public |
|---|---|---|
| Chords | June–July 2026 | 17 September 2026 |
| Motors | July 2026 | 17 September 2026 |
| Ostinatos | July 2026 | 17 September 2026 |
| Percussion | July 2026 | 17 September 2026 |
| MIDI of the audio (the MIDI reader) | September 2026 | 27 September 2026 |
1. Sources and owners of the data
- Our fine-tuning data is Cinematic Composing's own. The notes come from its MIDI packs, composed for Cinematic Composing:
- Cinematic Chord Progressions for Strings (Vol. 1 and 2);
- Fast String Motors;
- Medium Tension Motors;
- Cinematic Ostinatos for Strings;
- Cinematic Percussion (Vol. 1 and 2).
- The audio was rendered from that MIDI by Cinematic Composing, in Cubase, with its own sample libraries. Those libraries are built from orchestral recordings licensed with AI-training rights and from our own recording sessions.
- The text descriptions paired with each clip were generated by our own templates.
- The base model was pre-trained by Stability AI. Its model card states its sources: 806,284 recordings licensed from AudioSparx and 472,618 recordings from Freesound under CC0, CC-BY or CC Sampling+ licenses. Stability AI lists the Freesound attributions at info.stability.ai/attributions.
2. How the data serves the purpose
The MIDI packs give each instrument family its notes, rhythms and phrasing. The rendered audio teaches each model how that music sounds played by an orchestra. The MIDI paired with each clip teaches the MIDI reader to read notes back from audio. The descriptions and musical labels let the model follow your settings.
3. How much data
| Model | Training clips |
|---|---|
| Chords | 10,000–20,000 |
| Motors | 20,000–30,000 |
| Ostinatos | 5,000–10,000 |
| Percussion | 1,000–5,000 |
| MIDI reader | 30,000–40,000 (the same clips as the instrument models) |
| Stability AI's base model | 1,278,902 recordings (per its model card) |
4. What the data looks like
Stereo audio clips, 44.1 kHz, 24-bit, from about 4 to 40 seconds long. Each clip comes with its MIDI, a text description and musical labels: instrument, articulation, mood, key and mode, chord names, tempo, time signature, register and layer.
5. Copyright, trademarks and patents
The compositions and sound recordings in our data are protected by copyright. They're owned by Cinematic Composing or used under license. They are not in the public domain. No material protected by trademark or patent was intentionally included. For the base model, Stability AI states that its data is licensed or under Creative Commons licenses, and that identified copyrighted music was removed from the Freesound portion.
6. Purchased or licensed
The MIDI packs and the rendered audio are Cinematic Composing's own. The recordings behind our sample libraries are licensed for AI training or come from our own sessions. We bought no datasets. For the base model, Stability AI licensed its data from AudioSparx and used Creative Commons audio from Freesound.
7. Personal information
To our knowledge, none. The data is instrument recordings, MIDI and musical labels — no voices, no names, no customer data.
8. Aggregate consumer information
None.
9. Cleaning and processing
- The MIDI was transposed to every key and rendered at several tempos. Tempos come from re-rendering, not time-stretching.
- The dynamics were tuned by hand; some octaves were corrected; percussion patterns were extended and mapped to the instruments.
- The audio was cut into clips, with short fades and trimmed tails.
- Each clip was padded to a fixed length and converted into the model's compressed audio format.
- Purpose: cover every key and tempo, and give the models clean, consistent examples.
10. When the data was collected
The audio was rendered between 24 June and 22 July 2026, from MIDI packs composed earlier for Cinematic Composing. Collection isn't ongoing. When we add instruments, we'll add data and update this page before release.
11. When the data was first used
The chords data was first used in June 2026. The motors, ostinatos and percussion data were first used in June–July 2026. The MIDI reader's training in September 2026 reused the same clips.
12. Synthetic data
Yes, in this sense: all our training audio was rendered from MIDI through sample libraries, rather than recorded live for these compositions. The key and tempo variations were made by re-rendering, and the text descriptions were generated by templates. No audio made by other AI models was used.
Your data and our training
We never use your files, recordings, MIDI captured from your DAW, descriptions, the notes, chords and tempo we read from them, or your takes to train any AI model — ours or anyone else's — and we never sell or license them for AI training. The training we do uses only music written by professional composers and carefully curated for it — a slow, deliberate process. Training on anyone's music would first need their permission, and we have never asked for it. If that ever changed, we would ask your permission first, through a separate opt-in that is off by default. This promise is part of our Terms of use, and anyone who takes over Composer Assistant must keep it.
AI-generated output and the EU AI Act
- Everything Composer Assistant generates is AI-generated.
- Marking. The EU AI Act requires providers like us to mark AI-generated audio in a machine-readable way. We haven't built this marking yet; takes you download today carry no such marking from us. It is in development, and we'll update this page, with the date, when it's live.
- If you publish in the EU. If a take imitates a real person, recording or event so that it could be mistaken for genuine, you may have to tell your audience it was AI-generated. For artistic work, a credit line or note is usually enough.
Questions
Write to hi@cinematiccomposing.com. Composer Assistant is made by Cinematic Composing LLC, Lake Forest, California, USA.