The latest articles on DEV Community by Bella YL (@bella_yl_549b907bcbf18441). https://dev.to/bella_yl_549b907bcbf18441 https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4229991%2F7c965fd9-9345-473b-b14b-08e2166d5501.png https://dev.to/bella_yl_549b907bcbf18441 en Bella YL Fri, 06 Nov 2026 08:44:48 +0000 https://dev.to/bella_yl_549b907bcbf18441/how-we-built-the-musicgo-project-3bgh https://dev.to/bella_yl_549b907bcbf18441/how-we-built-the-musicgo-project-3bgh <p>I’m a software and AI engineer, and I built MusicGo with a small startup team. We started with a question: once an AI generates the first version of a song, how do you make it easy to work on the parts you want to change?</p> <p>Imagine you want to write a song about a late-night train ride. You might start with a sparse piano arrangement and a quiet vocal. Once you hear the result, you might like the chorus but want to rewrite the second verse or try a more electronic arrangement. That second step shaped MusicGo as much as the first. We wanted users to be able to return to a song, pick up the lyrics or style description, and generate another take.</p> <p>What goes into a song request</p> <p>The MusicGo website uses Next.js and Tailwind CSS. In the creation flow, users can describe a song in plain language, including its genre, mood, instruments, and lyrical theme. They can also supply their own lyrics or have AI help draft them.</p> <p>Lyrics and style descriptions give users two different ways to revise a track. If the verse feels too generic, they can change the words while reusing the style description. If the lyrics work but the arrangement feels too busy, they can reuse the lyrics and try a different style. The next generation is a new interpretation of those inputs, so the result can vary beyond the specific change the user made.</p> <p>Why the library is part of the editor</p> <p>A saved track gives users something concrete to come back to. They can listen again, open the song, revise its lyrics or style description, and generate a new version. That makes the library part of the songwriting workflow. An earlier take can become the starting point for a later one.</p> <p>For example, a user might listen to the piano version on Monday, rewrite the chorus on Tuesday, and try the revised lyrics in a different genre on Friday. The library lets them return to the song they were working on instead of having to reconstruct the idea from memory.</p> <p>Where stem separation fits</p> <p>We also added vocal and instrument separation for generated songs in the MusicGo library. After choosing a track, users can separate the vocal from the accompaniment or isolate parts such as drums, bass, piano, and guitar.</p> <p>Generation and separation are two distinct steps in the workflow. Generation starts with a description or lyrics and produces a song. Separation starts with a song that already exists and gives users individual audio parts they can listen to. For the train song, that might mean hearing the vocal on its own to get a better sense of how the chorus lands.</p> <p>The product question I’m most interested in is what people want to carry forward from one generation to the next. Lyrics and style are two inputs we can give users direct control over. Once people have a take they like, I’d love to hear what they would want to preserve in the next one. Would it be the melody, the vocal delivery, or something else?</p> <h1> showdev </h1> <p><a href="https://musicgo.org/" rel="noopener noreferrer">MusicGo project</a></p> ai nextjs startup webdev