Guide · Symphonies
How to make a symphony with AI
This guide is a translation of the original Spanish article. The three symphony pages it links to are written in Spanish; the music needs no translation, and your browser can translate the text.
It can be done, and it sounds better than I expected. But a symphony with AI does not come from asking for “a symphony”: it comes from deciding what story is being told, in what order, and which idea comes back at the end, and then checking afterwards what the AI did and did not deliver.
This is what I have learned from making three: the Symphony of the Dome, the Symphony of the Orchestra and the Symphony of the Lighthouse. They add up to 13 movements and almost an hour of music (59 min 53 s), and I measured the audio of each movement to compare it with what I asked for.
Can you make a symphony with AI?
Yes, if you separate three things that are also separate in a real orchestra: the idea (what is being told and how), the performance (who plays) and the editing (which take is chosen and in what order it is heard).
In my symphonies the idea and the editing are mine and the performance is done by an AI. That is why I talk about “directing” and not “composing”: I do not write the score note by note, I describe in words what I want each scene to sound like and decide what stays.
A symphony, moreover, is not a long song: it is a work in several movements, each with its own character, that together tell something. Each movement comes from a different generation, so the coherence between them has to come from the plan, not from the machine.
Which tool and which settings I use
I generate the music with Suno (model v6) in instrumental mode. I prepare the instructions with the help of another AI, but I review and decide what is asked for.
I keep the same settings fixed in every movement so they can be compared with each other:
- Weirdness at 20 %, Style Influence at 75 % and Variety at 0 (with Variety above 0 the program rewrites what you write and the test stops measuring my instruction).
- Max Mode on, which is the most expensive generation and the one I use for pieces longer than two minutes.
- A separate “Exclude” list for what I do not want (voices, choir, electronics).
- Instructions per scene in square brackets, with tempo, time signature and instruments.
These are not a winning recipe: they are the settings I fix in order to compare. I have not tried varying them systematically.
The method, step by step
A story with an arc
Something that changes from beginning to end. The Dome tells of a dome that takes generations to build; the Orchestra, how an orchestra is born, from the idea to the premiere; the Lighthouse, a lighthouse on the Atlantic coast, from the first flame to the beam that turns on its own.
An object that returns transformed
This is what makes the ending mean something. In the Dome it is the first low note, which returns at the start of the last movement; in the Orchestra, the first A; in the Lighthouse, the light of the lamp.
A short motif
Four or five notes, described in words and not by the names of existing works, which are born fragile and change colour over the course of the symphony, for example from minor to major.
Movements with different forms
An opening, a fast movement, a slow one and a finale. If all four have the same form, the work becomes a single song repeated.
Instructions per scene
Each scene carries its tempo, its time signature and its instruments. I describe the sound (acoustic symphony orchestra, concert hall, a solo oboe) and I never name a composer: Suno can block the generation if you name real artists, and what I am after is the sound, not an imitation.
Several takes and one choice
Suno returns several versions of each instruction and I keep one per movement. That choice, together with the final order, is one of the most important decisions of the direction.
Measure it and report it
Once a movement is chosen, I measure the audio and publish two things separately: what I asked for and what is actually heard. It is the most useful thing I have taken from the whole process.
What the AI delivers and what it does not: 13 movements measured
This is the comparison between what I asked for and what the automatic audio analysis measures across the 13 movements. They are measurements by a program, not impressions by ear, and the last row explains its limits.
| What I asked for | What happened | Where you can see it |
|---|---|---|
| Silences inside the piece (one to four bars) | It does not generate them: the only silence is the one at the end | 0 of 13 movements; neither those of Silentium nor the pause in La Prima |
| Tempo changes within a movement | Each movement stays at an almost constant pulse | In the 8 movements where I asked for them, none could be clearly distinguished |
| Asymmetric time signatures (7/8, 5/4) | They come out in 4/4 or cannot be made out | Machina sounds in 4/4 at about 152 beats per minute; the 5/4 of Dialogo cannot be made out |
| Wide volume contrasts | The dynamics are always narrow | Between 9.8 dB (Tavolozza) and 15 dB (Dialogo) across the 13 |
| A long final fade (8–10 seconds) | It usually arrives much shorter | Il Primo La, La Prima, Costa Ciega and Luz Sola: a few seconds or only the last half second |
| Starting key | Almost always the one requested | Exceptions: in El Torrero I asked for F minor and the analysis finds B-flat major; the last stretch of Prove does not reach major |
| Pulse (exact BPM) | Approximate: the detector sometimes counts double or half | Dialogo reads 172–178 for the first 40 s, double the 88 requested |
A music teacher told me that Machina sounded to him like 4/4 rather than 7/8. I measured it and he was right: I asked for 7/8 and the AI played 4/4. I left the difference visible on the page because it is part of the experiment.
What I changed in each symphony based on what I measured
I made them in this order, and each one starts from what I measured in the previous one.
- Symphony of the Dome. The first, and the one where I asked for the most: measured silences, a scherzo in 7/8 and tempo changes inside every movement. When I measured it, almost none of that could be heard.
- Symphony of the Orchestra. I moved the volume contrasts over to instrumentation and density: a solo instrument against the whole orchestra, fast notes against long ones. I kept asking for tempo changes, which also could not be distinguished.
- Symphony of the Lighthouse. I stopped asking for what the AI did not deliver: no internal silences, no tempo changes, no unusual time signatures. A single pulse per movement, and the first one with no fixed pulse. The measurement matches what I asked for more closely: the requested key appears in three of the four movements. I am not saying it is better music, only that I asked for things the AI does know how to do.
Tips for asking Suno for a symphony
These are conclusions from my measurements, not laws: another model or another version may behave differently.
- One tempo per movement. If you need two, generate two scenes separately and join them afterwards.
- Contrast with instrumentation and density, not with volume. The dynamics always come out between 10 and 15 dB.
- Silences afterwards. If silence is part of the story, make it in an audio editor with a fade-out, a pause and an entry on a strong beat.
- Do not depend on a 7/8 or a 5/4 if it matters that it is heard. It may come out in 4/4.
- Make the long final fade yourself if it is important for the ending.
- Allow for more duration than you think. Instrumentals with fast scenes tend to run longer than the sum of the scenes suggests.
- Describe the sound, do not cite anyone. It is safer and avoids resemblance to specific works.
How I measure what you hear
I analyse the audio of each movement with librosa, an audio analysis library: pulse, grouping into bars, silences, level and dynamic range, and the final fade.
It is reliable for silences, level and fade, and for the pulse when it is very marked. It is only indicative in soft movements, and unreliable for the key with an orchestra, where several pieces end up in “B-flat major”, probably as an artefact of the analysis. That is why each page warns when a measurement should not be trusted.
The three symphonies
They can be heard in full, movement by movement or in order, without signing up. Each page includes the story, the motif, the styles requested and the measurement of each movement. The pages are in Spanish. See the English summary of the three symphonies.
Frequently asked questions
Can an AI compose a symphony?
It can generate the sound of an orchestra and produce coherent music within each movement. What it does not do on its own is decide what to tell, in what order and with what idea that returns: that direction is human. And it does not faithfully follow what it is asked for in silences, tempo changes or asymmetric time signatures, as the measurements above show.
Which tool are they made with?
With Suno (model v6), in instrumental mode. I prepare the instructions with the help of another AI, and I choose, order and publish the result myself.
How long is each symphony?
The Dome lasts 17 min 43 s (4 movements), the Orchestra 24 min 5 s (5 movements) and the Lighthouse 18 min 5 s (4 movements). Each movement lasts between about 4 and 5½ minutes.
Are they works by a famous composer?
No. They are works generated with AI and directed by their creator, and authorship is always credited to the station. Each page shows stylistic affinities with composers as a reference to orient the listener, but they are not imitations or works by them.
Where can I listen to them?
On their own page of the blog, free and without signing up, with the “Escuchar en orden” (listen in order) button or movement by movement. You can also open the sound studio to try the equaliser and surround sound.
Last updated: