Transformation of Music Generation with AWS Trainium and Amazon SageMaker HyperPod at Splash Music

Generative artificial intelligence is revolutionizing the music industry, enabling creators of all skill levels to produce studio-quality tracks through models that personalize compositions in real-time. In this context, Splash Music, in collaboration with AWS, has developed a music generation system that facilitates professional creation, making this experience accessible to millions of people.

The company has set a new standard with its HummingLM model, designed in partnership with the AWS Generative AI Innovation Center. During the 2024 edition of the AWS Generative AI Accelerator, Splash Music worked alongside AWS Startups to accelerate innovation and enhance its music production model.

With over 600 million global streams, the platform has empowered a new era of creators, continuously adapting to users’ evolving tastes and making music production more accessible. However, its development has not been without challenges, including issues of complexity, scalability, and speed in the evolving industry. Previously, the company relied on external GPU clusters, resulting in unpredictable latencies and management complications.

To address these challenges, HummingLM was developed—a generative model that combines various modalities to interpret and create music. This model employs Descript-Audio-Codec audio encoding, producing compressed audio representations. Its architecture is based on an advanced language model and a specialized encoder, allowing users to transform sung melodies into high-quality instrumental performances.

The collaboration with AWS, along with the use of AWS Trainium EC2 instances, has enabled Splash Music to accelerate its development. With the automation and scalability of SageMaker HyperPod, the company has optimized its operations, reducing training costs by over 54% and decreasing training times by nearly 50%. HummingLM stands out not only for its sound quality but also for its ability to adapt to new instrument presets without requiring additional training.

Looking ahead, Splash Music plans to expand its training dataset by up to ten times, explore multimodal audio and video generation, and continue collaborating with the AWS Innovation Center on research and development projects. With this focus on innovation and a robust infrastructure, Splash Music is redefining the musical creative process, allowing anyone to generate custom tracks that connect with a broad audience.