Up until the last chapter,
I had been writing about the period
when I was producing instrumental tracks one after another.
By then,
I already had enough songs.
The overall world and atmosphere were becoming clear.
That was exactly why
I felt it was time
to try adding vocals.
The first track I worked on
was “花は散り”.
I’ve been composing music for a long time,
so I don’t have a strong fear of writing lyrics.
Still, this time was different.
The music itself had been created by AI.
So first,
I asked ChatGPT to come up with lyrics.
They weren’t bad.
But something didn’t quite click.
I decided to set the AI’s draft aside
and write lyrics freely,
just following what came into my head.
I fed those lyrics into Suno
and tried adding vocals using the cover feature.
That’s when
the outline of the melody began to appear,
faintly at first.
I adjusted the lyrics to fit that melody.
Then, almost in reverse,
wanting to draw out the melody more,
I stretched certain lines
or repeated words deliberately.
As I went back and forth
between lyrics and melody,
the shape of the song slowly emerged.
At one point,
I felt the overall mood might be too dark.
So I took the near-final lyrics
and put them back into ChatGPT,
this time asking for objective feedback.
Among the suggested revisions,
there were a few moments where I thought,
“Yes—this works.”
I incorporated those,
and for the time being,
the lyrics were finished.
Once the lyrics were locked in,
it was just a matter of repeated generation.
I re-generated the vocals
again and again
until it felt right.
My credits
disappeared faster than I expected.
One surprisingly tricky part
was how often kanji were mispronounced.
I changed some words to hiragana,
or intentionally used katakana
to control the accent and phrasing.
After many small adjustments,
the song was finally complete.
When I listened to the finished version of “花は散り,”
I was honestly surprised.
The performance was solid,
and the arrangement was impressively sophisticated.
When I tried analyzing it more closely,
I noticed something interesting.
The rhythm subtly swings
with sixteenth notes—
just slightly off-center.
That instability
creates a floating, suspended feeling.
The chord progression is simple at its core,
but includes passing diminished chords
and effective on-chord bass movement.
And the vocals fit the song
almost too well.
There was a sense of melancholy in the tone—
to the point where I found myself thinking,
“Is this really an AI singing?”
I even tried singing the song myself.
That’s when I realized, physically,
“This is actually a difficult song.”
There was a lot to learn
just from listening to the AI’s vocal performance.
Compared to the instrumental version,
the world of the song felt far clearer.
I couldn’t remember the last time
I genuinely thought,
“A piece of work is finished.”
To be honest,
I was a little moved.
Riding that momentum,
I added vocals in the same way
to “Blue Carpet”
and “Last Memories.”
I gradually expanded the lyrics,
and whenever I got stuck,
I consulted ChatGPT.
That back-and-forth itself
became part of the creative process.
Even while working a full-time job,
making music became a daily routine again.
During commutes,
or in short moments at night,
I would immerse myself in music.
Before I realized it,
music had returned
to the center of my life.
Next time,
I want to write about
how these songs came together
and took shape as an album.
🎬 花は散り
🎬 Blue Carpet
🎬 Last Memories
