cody.soy

Fluency, but Make It Sound Good

How I've built i·CANTO — a multilingual lyric learning engine that turns music into memory, powered by GPT-4o and a little poetic engineering.

6 min read

TL;DR
I built i·CANTO — a multilingual lyric-learning app that uses GPT-4o to turn music into a poetic language tutor. Lyrics are translated line-by-line with contextual prompts, cached locally, and enhanced via a hybrid API system with GPT fallback. Think educational karaoke meets AI engineering. Source: GitHub

🎧 Music Is the Original Language Model

Before I ever wrote a line of code for i·CANTO, I wrote lyrics. Before I cared about LLMs, I memorized verses. And before I could conjugate verbs in Spanish, I could belt the chorus of “Ahora Te Puedes Marchar” like I meant it.

The idea behind i·CANTO was never “what if AI could teach language?”
It was: “What if the way we already learn music… was the best way to learn language?”

So I started building. Not a lesson. Not a module.
An instrument.

🧠 Building the Learning Instrument

i·CANTO started as a UI experiment. A kind of “Educational Karaoke.”
It became a real-time, multilingual, lyric-learning pipeline — powered by:

  • GPT-4o translation streams
  • A resilient lyrics API with Redis + multi-source fallback
  • Dynamic UI feedback for learning state & translation health
  • IndexedDB caching for blazing-fast recall
  • Smart prompt chaining for followups (quizzes, conjugation, journaling)

The frontend sings. The backend improvises.
And every translation is tuned like a line of poetry.

📜 Poetic Engineering with GPT-4o

Literal translation? That's easy.
i·CANTO does something harder: it makes translations feel like lyrics.

Our prompt pipeline feeds GPT-4o not just the line, but its context — previous/next lyrics, mood, artist, tone. We strip formatting. We adapt idioms. We never explain — we rephrase.

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

The result is something no dictionary can offer:
Translation that sings.

We’re not giving you “the meaning” — we’re giving you the emotional resonance.
The turn of phrase you might actually say, or better yet, feel.

Whether it's a breakup line from Rosalía or a whispered confession in Japanese, our model doesn't just translate — it interprets. With context. With style. With rhythm.

🧠 AI Built for Lyricism

Most translation APIs treat language like a math problem.
i·CANTO treats it like a poem.

Our API stack is hybrid by design:

  • translate-lyric (GPT-4o) — context-rich, poetic, idiomatic
  • fallback-translate (GPT-3.5) — resilience under rate limits
  • autocorrect-song — fixes malformed inputs
  • gpt-prompt — suggests follow-up prompts
  • lyrics — hybrid search pipeline w/ cache, enrichments, and error handling

Each endpoint plays a role in the full lyrical loop:
✨ Discover → 🧠 Understand → 🔁 Repeat → 🔥 Create

We’ve optimized for developer flow too.
Every part of the system is testable, observable, and swappable.
Need a new fallback LLM or localization model? Swap it.
Want to bring your own prompt library? Plug it in.

🔮 The Future Sounds Like Fluency

Our roadmap isn’t just about tech — it’s about connection.

We’re building:

  • 🧠 Persistent chat memory, so GPT remembers your vibe
  • ✍️ Real-time journaling and reflection from lyrics
  • 🎙️ Karaoke-style pronunciation scoring
  • 💾 Prompt bookmarking and history
  • 🎧 Artist-based learning journeys

Because we believe learning a language isn't about drilling vocab.
It’s about feeling something in a new tongue — and then saying it back.

i·CANTO is how you speak what you’ve felt all along.
Fluency, by way of feeling.
And it sounds damn good.

🔁 Singing Is Studying

Here’s the secret: learning a song is spaced repetition in disguise.

That’s why i·CANTO turns lyrics into a musical interface for memory:

  • 🔠 Line-by-line lyric display with tap-to-translate
  • 📀 Local caching for blazing-fast revisits
  • 🎭 GPT followups: quizzes, journaling, verb drills
  • 🎛️ Shimmer loaders and retry controls for failed lines
  • 🎨 Animated transitions with Framer Motion
  • 🌍 Language selector with persistent IndexedDB state

You’re not tapping through flashcards.
You’re listening, translating, repeating — on beat.

🌍 Designed for Polyglots. Built for Builders.

We support 15+ languages. But more importantly,
we support the ways people actually learn.

That means:

  • Contextual followups like “learn more from this artist”
  • GPT prompts for slang, grammar, and journaling
  • A prompt grid that asks: “Want to study French trap? Hindi ballads?”

And under the hood?

  • /api/lyrics pulls from LRCLIB, Deezer, Lyrics.ovh — with normalization, caching, and smart fallback
  • /api/translate-lyric serves poetic, idiomatic GPT-4o translations
  • IndexedDB tracks each line's history and translation status
  • GPT-3.5 fallback ensures learning doesn’t stall

This isn’t an app.
It’s a melodic language engine.

🛠️ From Karaoke to Knowledge

It began as a “fun little AI karaoke.”
But what we built is real-time language intelligence — embedded in music.

i·CANTO now does all this:

  • Turns messy user input into cleaned artist/title data
  • Fetches and normalizes multilingual lyrics
  • Translates emotionally, not literally
  • Teaches through interaction, not instruction
  • Scales from casual singalong → immersive linguistic deep dive

We didn’t write a curriculum.
We built a loop:

🗣 Hear the line → 💡 Translate → 🔁 Repeat → ✍️ Reflect → 🚀 Continue learning

✨ Music Is Memory

i·CANTO works because music already knows how to teach.

It repeats.
It moves you.
It rewards you for feeling, not just remembering.

We wrapped that process in AI.
We made it multilingual.
We made it sing.

This is poetic engineering.
It’s how i·CANTO turns raw lyrics into expressive, idiomatic language that sticks — like the best hooks always do.

When GPT fails?
We fall back gracefully — shimmer loaders, retry buttons, health indicators.
Because resilience is part of learning too.

Want to try it?
👉 https://icanto.tech

Want to build with it?
👉 github.com/wodydoc/icanto-ai-chatbot

Made with ❤️ in Paris — by @yosoycody

⚙️ Stack Notes for the Curious

  • Built with: Next.js 14, TypeScript, Vercel AI SDK, Tailwind, shadcn/ui
  • AI: GPT-4o + GPT-3.5 fallback, streaming with /translate-lyric, /gpt-prompt
  • Caching: IndexedDB (local) + Redis (server) for translations and lyrics
  • Lyric Sources: Deezer, LRCLIB, Lyrics.ovh, OpenAI fallback (auto-generated)
  • UI: Framer Motion + hover/tap interactions
  • Deployment: Vercel, serverless

i·CANTO — fluent by music.