idle() the music of ipe()

No. 3· behind-the-music

How an AI Lo-Fi Song Gets Made

How one idle() AI lo-fi song goes from lyric to release: who hands what to whom, where we check, and how we say plainly that it is made with AI.

한국어로 읽기 →

The album cover: a latte being poured, with the words dear(you); vol.1, dear you and a Korean line meaning to you who made it through today

The cover of dear(you); vol.1. The ten songs on it were made following the order described below.

Is this lo-fi made by AI? Yes, with Suno. Here are the five hand-offs from lyric to release, and where a person steps in. We are sometimes asked how an idle() song is made. The one-line answer is "with AI", but there are more hand-offs and checks inside that line than you might expect. This post walks through the production order we have set for ourselves, from the first lyric to the release. It describes a procedure we have written down, so it does not mean every song moved through it without a single deviation.

First, the sentence that matters most. The music is produced with AI (Suno). Lyrics are AI-drafted and reviewed by a person. The rest of this post is really an explanation of what that sentence looks like as actual work.

Five hand-offs

A diagram of five stages joined by arrows: lyrics, music, visual, review and release, each with the role in charge and what it produces. The review stage is outlined in gold

The five stages one song passes through. The fourth, review, is a gate that cannot be skipped.

A song passes through five stages: lyrics, music, visual, review and release. The next stage starts only when the previous one is done, and the person in charge changes at every step. We call each change a hand-off, and in this post we also count the planning that comes before the first lyric. A hand-off is made of files and statuses, not conversation. When the person in charge of lyrics finishes a song file and sets its status to "composition requested", the person in charge of music reads that file and starts work. Because who handed over which file in which state is always on record, context does not get lost on the way.

The statuses a song moves through are fixed as well. A new lyric is a "draft". While its tone and emotional flow are being shaped it is "polishing". Handing it to music makes it "composition requested", producing the audio makes it "composing", and only after it passes review does it become "released". One look at this board tells you whose hands a song is in right now.

Stage 0: planning, because the album comes before the song

The first step is really about the album, not the song. The concept, the track order and the flow of feelings are drawn first. For dear(you); vol.1, a concept ("ten letters to the you who made it through another hard day") and an arc of ten feelings (empathy, understanding, fear, reassurance, rest, connection, gratitude, acknowledgment, hope, calm) were set first, and each song was handed its place on that arc before it reached the lyric stage. A song starts as one slot in an album rather than as a stand-alone piece, and we think that is part of why our songs resemble each other in structure.

Stage 1: lyrics, one song in two versions

The output of the lyric stage is one song file. It holds the theme, the mood, the story and the lyrics in two versions: one with Korean as the main language and one with English. The Korean side is the original, and the English side is written as an adaptation that carries over tone and feeling rather than words. If an AI's literal translation is sung as it is, the rhythm and the emotion break and the quality of the song drops.

The English-main version is sung in English by a female vocal with Korean whispers, and the Korean-main version is sung in Korean by a male vocal with English whispers. Both end with the whispered line ipe. be still

The rule for the two versions: each version is sung in one language, and the other language appears only as a whisper.

The principle here fits in one line: the language that is sung stays fixed within a version, and translation is delivered as subtitles. It means we do not force a literal translation into a melody. The other language is added only as a short whisper, so the English version carries a Korean word or two and the Korean version carries an English one. Those whispers are the seam that holds the two versions together as one song.

There is one more thing the lyric stage must do: put "ipe. be still" at the very end of the song. It is a one-breath whisper, fixed in English, and every song since dear(you); ends with it. The lyric file counts as finished, and can be handed to music, only after this signature is in.

Stage 2: music, from words to sound

The music side takes the lyric file and builds a production package. The package contains the sung lyrics for each version and a style prompt: a written description of genre, instruments, the character of the voice and the texture of the mix. The first song of the album, for example, is listed as ambient lo-fi with soft piano, warm guitar, gentle pads, vinyl noise and light rain ambience, with a sweet female vocal for the English version and a gentle male vocal for the Korean one.

With that package, the audio and vocals are generated with Suno, an AI music tool. A single generation rarely gives exactly what we want, so several are tried and one is confirmed. The confirmed version becomes the song. When a song gets subtitles for video, the subtitle file has to come from that confirmed version, because timing changes with every generation. That rule is written into our procedure.

This is where translation enters, as subtitles. On YouTube there are two kinds: captions the viewer turns on, and lyrics burned into the picture. Automatic timing is not available for Korean lyrics, so a subtitle file has to be uploaded by hand, and a song whose words we want to emphasize can be made into a lyric video with the text burned in. Each language gets its own cover and its own subtitle file.

Stage 3: visual, giving the sound a face

Once the audio is confirmed, the visual side makes the cover, the YouTube thumbnail and, when needed, a looping background for the video. Its inputs are the concept of the song, the audio, and the tone of the time of day the song belongs to. Season 1 splits one day into four albums (night, dawn, morning and evening), so songs on the same album share the colors of one time of day.

A few visual rules are fixed. The palette centers on deep navy, warm gold, sage green and cream, and the lettering uses a monospaced typeface reminiscent of a code editor. On a cover, English is the lead and Korean is the supporting line. Look at the cover of the first song and you will see the English title set large with the Korean title beneath it.

Stage 4: review, a gate that cannot be skipped

When the lyrics, audio, cover, thumbnail and the release card on this site are all ready, review opens just before release. This gate cannot be skipped. It has two sets of eyes. One looks at quality and risk: is the sound clean, do the language and the voice in the audio match the lyric (a female voice for the English version, a male voice for the Korean one), do the feeling and the track order fit the album plan, are there copyright risks such as AI labeling or plagiarism, and are links and metadata correct. The other looks at the brand: palette, typeface, a tone we describe as quiet before noise, and the language hierarchy on the cover.

A few items are hard stops. Is "ipe. be still" at the very end? Is the language hierarchy on the cover and thumbnail respected? Do the links and languages on the site's release card match? If the signature is missing, the song is sent back, fixed, and reviewed again. The reviewer only points out problems and does not rewrite the lyric. Fixing belongs to whoever made it. Keeping the person who checks apart from the person who made the thing is the heart of this gate.

Stage 5: release, upload, connect, record

After review, the music side uploads to YouTube and, for songs that need it, handles distribution to streaming services. The upload step includes checking the label that says the content was made with AI. Then the song's status becomes "released", the YouTube links are written on the song board, and the release card on this site and the summary documents are updated. A song can look finished and still have one job left: checking that its links agree in every place they appear.

The release method is a hybrid from the start: singles go out one at a time to keep the channel moving, and once a series has enough songs they are gathered into an album. All ten songs of dear(you); vol.1 are on YouTube in an English version and a Korean version. A long full-album video, for listening to the whole album in one sitting, is also on YouTube.

Why we put "made with AI" up front

The question we get most is, in the end, a single one: why show that it was not made by hand? Our answer is simple. There is nothing to hide, and not hiding is the approach that lasts. Producing the vocals with AI is written down as the default option in our working principles, and one reason is that it matches the brand idea of disclosing AI production openly. What people do is set the direction, then read, review and choose from the draft. What the AI does is produce drafts and sound quickly. We don't blur the two together. We write down what each one takes on, and that is our promise.

Listen

The ten songs of dear(you); vol.1, made following the order described above. Each is on YouTube in an English version and a Korean version.

The whole channel is at YouTube ipe-rest.

We hope the next time you listen to a song, you notice a little more. Behind the whispered line at the very end, there are five stages and one gate that cannot be skipped between a lyric and a release. You can listen on the idle() home page.