MAESTROGEN: Building a Full-Stack AI Music Generation SaaS Platform

Abstract

Music compositionrequired training and equipment which most people usually do not have. The fastemerging progressingenerative artificial lintelligencehasopenednewdoorsforcreativecontent generation, but still accessible end-to-end AI music creation platforms remain rare. This paper depicts MAESTROGEN, a full-stack Software-as-a-Service (SaaS) platform that allows users to generatesongsincludingaudio,lyrics,andalbumartworkfromnaturallanguagedescriptionsalone. The system has a combination that integrates three self-hosted AI models: the Ace Step diffusion transformer for audio synthesis, the Qwen 2 large language model (LLM) for automated lyric and prompt engineering, and SDXL Turbo for visual thumbnail generation. While akey architectural contribution is the resolution of the serverless timeout problem inherent to long-running GPU inference through durable event-driven orchestration via Inngest. The platform is built and developed on a modern stack comprising Next.js, TypeScript, Tailwind CSS, Neon serverless PostgreSQL, and AWS S3, deployed on Vercel with Modal providing on-demand GPU compute. Experimental evaluation illustrates high generation quality along three user-controlledmodes,zero timeout failures under concurrent load, and a 97% infrastructure cost reduction through scale-to- zero GPU provisioning. MAESTROGEN operates the text-to-music vision as a deployable, user- accessiblesystem, closing the gap between research-level generative audio model sand production- ready applications.

Introduction

MAESTROGEN is a full-stack AI-powered music generation platform that enables users to create complete songs from natural language prompts. Unlike existing AI music systems that focus on only one aspect of music creation (such as lyrics, melodies, or audio), MAESTROGEN integrates lyric generation, audio synthesis, and album artwork creation into a single production-ready SaaS platform.

The system addresses traditional barriers to music creation—such as the need for musical expertise, expensive software, and significant time investment—by leveraging recent advances in large language models (LLMs) and diffusion models. Research and market growth indicate increasing adoption of AI music generation, but existing solutions remain fragmented.

Key Features

• Generates lyrics, music, and album artwork from a single text prompt.
• Supports Simple Mode (prompt-based generation) and Custom Mode (user-defined lyrics, genre, and style).
• Includes authentication, credit-based monetization, song sharing, and cloud storage.

System Architecture

MAESTROGEN consists of four scalable layers:

1. Interface Layer: Next.js frontend for user interaction.
2. Orchestration Layer: Inngest manages asynchronous workflows and prevents serverless timeout issues.
3. Compute Layer: Modal serverless GPUs run AI models efficiently using scale-to-zero infrastructure.
4. Data Layer: Neon PostgreSQL stores metadata, while AWS S3 stores generated media.

Generation Pipeline

• Qwen 2 interprets user prompts, generates structured music style tags, and writes lyrics.
• Ace Step uses a diffusion transformer to generate playable WAV audio from the style tags and lyrics.
• SDXL Turbo creates matching album artwork simultaneously to reduce latency.
• Inngest coordinates the entire workflow asynchronously, improving scalability and reliability.

Experimental Results

The evaluation focused on generation quality, prompt accuracy, system reliability, infrastructure cost, and user experience:

• Manual lyrics mode achieved the highest audio quality.
• LLM-based prompt expansion improved genre and mood recognition.
• The asynchronous architecture handled concurrent users with zero timeout errors.
• Serverless GPUs reduced infrastructure costs by over 97% compared to always-on GPU deployments.
• Users reported high satisfaction and ease of use, with generation latency (2–3 minutes) identified as the primary limitation.

Conclusion

This paper presents MAESTROGEN, an all-in-one AI music generation SaaS (software as a services) platform thatcanturnauser\'snaturallanguagedescriptionintoafullsonglyricsthatcanplayaudio,andalbumartthrough a synchronised three-model, pipelines are used like any other AI model that\'s theoneof the most important thing inthisproject and.That\'showtheyall worktogethertoperformtheMastrogenQwen2,isaboutfiguringouthow to plan prompts and understand language through Ace Step, uses a diffusion transformer that is conditioned on both style tags and lyrics to put the audio together. SDXL Turbo matches the visual art, all of which iscoordinated by Inngest\'s durable event bus, which solves the serverless timeout problem with long GPU. The experimental results show that the generation quality is in line with the text-conditioned audio paradigm tested.

References

[1] M.Aghaei,L.Zhang,andF.Gu,\"AI-EnabledText-to-MusicGeneration:AComprehensiveReview,\" Electronics, vol. 14, no. 6, p. 1197, Mar. 2025. doi: 10.3390/electronics14061197. [2] AINewsHub,\"AIMusicGeneratorsin2025,\"2025.[Online].Available:https://www.ainewshub.org/post/ai- [3] music-generators-in-2025 [4] A.A.Abdullah,H.Veisi,andT.Rashid,\"AdvancingDeepLearningforExpressiveMusicComposition,\" Scientific Reports, vol. 15, Jul. 2025. doi: 10.1038/s41598-025-13064-6. [5] J.Copet,F.Kreuk,I.Gat,T.Remez,D.Kant,G.Synnaeve,Y.Adi,andA.Defossez,\"SimpleandControllable [6] MusicGeneration,\"NeurIPS,vol.36,2023,pp.47704–47720. arXiv:2306.05284. [7] S.Dingetal.,\"SongComposer:ALargeLanguageModelforLyricandMelodyGeneration,\" arXiv:2402.17645, Feb. 2024. [8] A.Agostinellietal.,\"MusicLM:GeneratingMusicFromText,\"arXiv:2301.11325,Jan.2023. [9] Z.Ju,P.Lu,X.Tanetal.,\"TeleMelody:Lyric-to-MelodyGenerationwithaTemplate-BasedTwo-Stage Method,\" EMNLP 2022, pp. 5426–5437. [10] S.ForsgrenandH.Martiros,\"Riffusion,\"2022.[Online].Available:https://riffusion.com/about [11] T.Guo,B.Chen,Y.Shengetal.,\"AI-BasedAffectiveMusicGenerationSystems,\"ACMComputingSurveys, [12] vol.57,no.1,2024.doi:10.1145/3672554. [13] B.-A.Cretuetal.,\"MusicGenerationwithMachineLearningandDeepNeuralNetworks,\"Procedia Computer Science, vol. 246, pp. 1855–1864, 2024. [14] J.Ho, A. Jain, and P. Abbeel, \"Denoising Diffusion Probabilistic Models,\" NeurIPS, vol. 33, pp. 6840–6851, 2020. [15] H.Liuetal.,\"AudioLDM:Text-to-AudioGenerationwithLatentDiffusionModels,\"ICML,2023,pp. 21450–21474. [16] E.Jonasetal.,\"OccupytheCloud:DistributedComputingforthe99%,\"ACMSoCC,2017,pp.445–451. [17] A.Vaswanietal.,\"AttentionIsAllYouNeed,\"NeurIPS,vol.30,2017. [18] T.Brownetal.,\"LanguageModelsareFew-ShotLearners,\"NeurIPS,vol.33,pp.1877–1901,2020.

Copyright

Copyright © 2026 Dr. Juthuka Aruna Devi, Budagatla Sai Vinuthna , Pendyala Madhu Sudhana Rao, Pedagadi Teja, Thammina Harika. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.