About 40% of songs created through Magical Song are gifts. Weddings, birthdays, anniversaries. Someone sits down, types out a story about the person they love, and a few minutes later they have a full song with real vocals they can share. We built the pipeline at Inithouse, and the process behind it is more interesting than it looks from the outside.
Here is how a wedding story becomes a finished track.
Step 1: The story
The user writes what matters. Not lyrics, not rhymes. Just the story. How they met, the first trip together, inside jokes, the moment they knew. Most people write 100 to 300 words. Some write three sentences. Both work.
The input is deliberately unstructured. We found early on that asking people to "write lyrics" produced stiff, self-conscious text. Asking them to "tell the story" produced raw material that actually sounds personal when turned into a song.
Step 2: Lyrics generation
The story gets processed into song lyrics. This is where the structure appears: verses, chorus, bridge. The system pulls specific details from the story (names, places, moments) and weaves them into a lyrical structure that fits the chosen genre.
A country ballad and an R&B track need different phrasing, different syllable counts, different emotional beats. The lyrics engine handles this per genre. We support 20+ genres, from acoustic folk to hip-hop to orchestral cinematic pieces, and each one has its own structural rules.
What does not happen here: generic filler. Lines like "you light up my world" or "together forever" get filtered. The whole point is that these lyrics could only be about one specific couple, one specific story.
Step 3: Vocal production
This is the part most people ask about, so here it is plainly: these are real vocals, not text-to-speech.
The difference matters. TTS (text-to-speech) engines convert text to audio by synthesizing speech patterns. The result sounds like someone reading aloud. It works for audiobooks and navigation apps. It does not work for singing.
Real vocal production means the output has pitch variation, breath, vibrato, dynamic range. A belted chorus sounds different from a whispered verse. The vocal sits in the mix the way a recorded performance would, not layered on top like a voiceover.
We spent a lot of time on this distinction because early prototypes with TTS vocals got a consistent reaction: "cool tech, but I would not actually send this to someone." The moment we switched to real vocal rendering, the reaction flipped. People started sharing the songs unprompted.
Step 4: Production and mixing
The vocal track gets placed into a full arrangement. Drums, bass, keys, guitars, strings, whatever the genre calls for. The mix is balanced for casual listening (phone speakers, car audio, laptop) because that is how most of these songs get played: someone opens a link at a dinner table or in a group chat.
The master comes out at studio quality. Not "good enough for AI," but genuinely at a level where recipients do not realize the song was generated until they are told.
Step 5: The shareable link
The finished song lives at a unique URL. No app download, no account creation for the recipient. The person who ordered the song gets a link, sends it however they want (text, email, printed QR code on a card), and the recipient just clicks and listens.
We made this deliberate choice because the moment of receiving a custom song is supposed to be frictionless. Nobody wants to download an app to hear their wedding gift.
What we learned building this
Three things stood out after running Magical Song for a while:
| Observation | What it means |
|---|---|
| ~40% of songs are gifts | People do not make custom songs for themselves. They make them for someone else. The product is closer to a greeting card than a music tool. |
| Wedding and birthday spike together | Seasonal patterns track celebration calendars, not music consumption patterns. |
| Story length does not correlate with song quality | A three-sentence story about a specific moment produces better lyrics than a 500-word life summary. Specificity beats volume. |
The pipeline from story to finished song takes minutes. The user does not see any of the intermediate steps (lyrics draft, vocal rendering, mixing). They write a story, pick a genre, and get a song. Everything in between is infrastructure.
Why this matters for the category
Most "AI music" tools right now hand you stems, MIDI files, or instrumental loops. They are built for musicians who want raw material to work with.
Magical Song is built for people who have never opened a DAW in their life and never will. The entire interface is: tell a story, pick a style, get a song. No timeline editor, no mixing board, no export settings.
That is a fundamentally different product for a fundamentally different user. We think the custom song category will split along this line: tools for creators vs. finished products for everyone else. We are building for the second group.
We build and run AI products at Inithouse. Magical Song is one of them. Try it at magicalsong.com.
Top comments (0)