Most side projects begin with a vague idea.
Music is not very different. You may know the atmosphere you want, but not the chords, melody, arrangement, or even the right words to describe it.
I wanted to see how far a single sentence could go:
An electronic pop song about driving through an empty city at 2 a.m.
There is a mood in that sentence, but not much else. No tempo, song structure, instruments, or lyrics. Traditionally, turning it into a complete track would require several separate decisions and probably a few hours inside music production software.
Instead, I used it as the starting point for a song on DeepSong.ai.
The first attempt
DeepSong offers a simple mode where a song can begin as a plain-language description. I entered the original sentence without adding any technical music terms.
The result captured the broad idea, but it also exposed everything the prompt had failed to explain.
“Electronic pop” could mean a bright dance track, a slow synth ballad, or something darker and more cinematic. An “empty city” could feel peaceful, lonely, or threatening. The first generation had to make those choices on its own.
That was useful. Hearing one interpretation made it much easier to describe the version I actually wanted.
A better prompt
For the second attempt, I rewrote the prompt:
A restrained electronic pop song about driving alone through an empty city at 2 a.m. Start with soft analog synths and a distant drum machine. Keep the verses quiet and slightly tense, then open into a wide, melodic chorus. Male vocal, medium tempo, reflective rather than sad.
This version contains several kinds of direction:
- A genre: electronic pop
- A setting: an empty city at night
- An emotional boundary: reflective, not sad
- An arrangement: quiet verses followed by a larger chorus
- Instrument ideas: analog synths and a drum machine
- Vocal and tempo preferences
The prompt is still readable as normal prose. It does not require a list of production parameters or knowledge of music theory.
That is one of the more practical things about generating music from language: the interface between the idea and the software is ordinary writing.
Moving into custom mode
A description is enough for exploration, but lyrics need more deliberate control.
DeepSong’s custom mode separates the title, lyrics, musical style, and other choices. I gave the track a working title, Streetlights After Two, and sketched a short structure:
[Verse]
Traffic lights change for nobody
Storefronts sleep behind the glass
Every road looks unfamiliar
Even ones I've driven past
[Chorus]
Streetlights after two
Turning red to green
The whole city disappears
Somewhere in between
Structure tags such as [Verse] and [Chorus] help communicate how the words should be arranged. They also make revision easier. If the chorus feels too crowded, it can be shortened without rewriting the entire prompt.
This is where the process started to feel less like requesting a finished asset and more like working through a draft.
Generation is most useful as a loop
The first custom version was not the final one.
One line had too many syllables. The transition into the chorus felt abrupt. The vocal delivery was stronger than the quiet mood needed.
Those were not failures so much as feedback. A written lyric can look fine on a screen and still feel awkward once it is sung.
The workflow became:
- Describe the song.
- Generate a version.
- Listen for one specific problem.
- Change the prompt or lyrics.
- Generate again.
Short feedback loops matter in creative work. When producing another draft takes hours, people tend to protect weak ideas because they have already invested time in them. When a draft is quick to test, it becomes easier to discard a line, change direction, or try something strange.
Where this becomes useful
The obvious application is songwriting, but complete songs are only one part of it.
A developer building a game can test different musical directions for a scene before the soundtrack is finalized. A YouTube creator can explore background music shaped around the pace and tone of a video. A podcaster can prototype an intro. A designer can give a product demo its own temporary soundtrack.
It can also help with communication.
“Make it more cinematic” means different things to different people. Generating a quick reference track gives collaborators something concrete to discuss. They can point to the rhythm, instrumentation, or energy instead of debating adjectives.
For early prototypes, that can be more valuable than obtaining a perfect track.
Keeping track of the results
Experiments create files quickly, and filenames such as final-version-3-really-final.mp3 are not much of a system.
DeepSong keeps generated tracks in a personal song library, so versions can be replayed and managed after the initial session. Tracks can also be created privately, while public creations can be shared with the wider community.
That distinction matters. Some generations are finished pieces; others are sketches, references, or ideas that should remain part of a private creative process.
What I learned from the experiment
The main lesson was not that one sentence can magically become a perfect song.
It was that one sentence can become something you can react to.
Before hearing the first generation, “electronic pop in an empty city” seemed specific enough. After listening, it became clear that the idea was missing an emotional direction, arrangement, vocal character, and sense of movement.
The generated track made the gaps visible.
That is where tools like DeepSong are most interesting. They shorten the distance between imagining and evaluating. The human part of the process remains: choosing the idea, recognizing what feels wrong, rewriting it, and deciding when the result finally says what it was meant to say.
The blank page is still there. It just makes sound now.
Top comments (0)