Inspiration

I wanted hobbyists, storytellers, and brands to turn ideas into full-fledged songs without needing a studio, session musicians, or years of theory lessons. Advances in generative audio and large-language models finally made that vision realistic—MakeSong.com is the result.

What it does

AI music generator——text to music or lyrics to song • Turn any text prompt, mood board, or reference track into a complete song—melody, harmony, arrangement, and lyrics. Export high-quality WAV/MP3 stems for vocals, drums, bass, and backing so creators can remix or master elsewhere.

How I built it

Data & Model • Curated a 1.8 M-track dataset of licensed multitrack stems and aligned lyrics. • Fine-tuned a diffusion-based audio synthesizer and a transformer lyric model that share a joint latent space, allowing melody–lyric coherence. Backend • Microservices (Python/FastAPI) on AWS Fargate + GPU-backed ECS tasks; Redis queues for job orchestration. • On-the-fly audio post-processing pipeline using SoX and FFmpeg. Frontend • Next.js + React + Tailwind for a responsive UI. • WebSocket layer for live render status, waveform previews, and collaborative edits.

Challenges I ran into

As an independent developer, ensuring that the music generated for my users is high-quality poses a significant challenge. To address this, I have optimized the functionality and introduced new techniques and methods for crafting effective song prompts.

Accomplishments that I'm proud of

Passed 50k registered users in the first 90 days, with a 27 % weekly retention rate.

What I learned

What's next for Make Song-AI Song Generator

Add the new features: • Stem separation for user-uploaded tracks, enabling seamless remixing. A song cover feature will be added later.

Built With

Share this project:

Updates