Music Was Always a Language Model

Why I built i·CANTO — an app that uses music to teach language the way music always has.

Before I ever wrote a line of code, I learned most of my Spanish from Manu Chao.

Not from a textbook. Not from Duolingo. From his albums — Clandestino, Próxima Estación: Esperanza — on loop during a summer in Texas. I didn't study it. I absorbed it. The same way you absorb anything you love enough to repeat.

Years later, I was in Barcelona at a coding bootcamp, leaving class one afternoon with headphones in, and I walked straight into Manu Chao on the street. I'm pretty sure it was the happiest moment of my life.

That encounter made something click. I'd been thinking about building a language app. Bumping into the person who'd taught me Spanish through music made the idea feel less like a startup and more like a personal obligation.

The Observation

Language learning has a strange problem: the methods that work don't feel like studying.

Nobody sits down and studies their native language. They hear it. They imitate it. They repeat it in context, with emotion, over time. The inputs are imperfect — song lyrics, slang, half-understood conversations — but the brain fills in the gaps.

Chomsky's theory of the Language Acquisition Device suggests we're born wired for language — that we don't learn it so much as unlock it. All we need is enough good input.

Music is extraordinarily good input.

A song repeats the same phrases across multiple verses at an emotional peak. You memorize it without trying. You feel the meaning before you can define it. That's not a gimmick — it's basically how language acquisition works in childhood. We just stop exploiting it once we hit a classroom.

What I Built

i·CANTO is an attempt to build around that observation.

The core idea: take a song, fetch the lyrics, and translate them in a way that preserves the emotional weight rather than just the literal meaning. Not "I write your name on the sand" as a dictionary exercise — but as a phrase you actually feel, in context, against the melody.

The translation pipeline feeds GPT-4o not just the line but its context — artist, mood, surrounding lyrics, tone. The goal isn't accuracy in the Duolingo sense. The goal is that the translated line feels like it could be in the song.

"Translate this lyric into natural, poetic English. Keep the emotional tone, not just the meaning."

Under that is a lyrics pipeline with multi-source fallback — Deezer, LRCLIB, Lyrics.ovh — Redis caching server-side, IndexedDB locally. The UI is built for the learning loop: hear the line, translate it, replay it, revisit it across sessions.

Singing is spaced repetition in disguise. You hear a phrase at the chorus, then again in the second verse, then again as the song ends. It sticks because it's music, not because you tried to memorize it.

Why the Music Industry Doesn't Scare Me

Paul Graham once wrote that the two startup verticals founders tend to regret are hardware and music. Hardware is intrinsically hard. The music industry is mafia.

That warning assumes you're selling music. I'm not.

i·CANTO doesn't license catalogs or compete with Spotify. Music is the input method, not the product. It's a pedagogical tool — the most powerful one we have for language acquisition — and the app exists to exploit that. I'm not a music company that teaches language. I'm a language app that uses music the way music has always been used: as the medium through which culture travels.

The licensing question dissolves when you change what you're building.

What I'm Still Learning

The app works. The translations feel right. The loop holds.

But the more I use it — in French, in Spanish, in Italian — the more I notice what's missing. The social dimension. The moment you play a song for someone else and your understanding of it changes. The comment in the margin. The argument about what a lyric actually means.

Language has always been social first, solitary second. Music as a language model is a strong thesis. But building the social side of it — that's the part I'm still figuring out.

Is the best language learning that ever happened already going on quietly in people's headphones?

Stack: Next.js 14, TypeScript, Vercel AI SDK, GPT-4o with GPT-3.5 fallback, Redis + IndexedDB, Deezer / LRCLIB / Lyrics.ovh.