Generative Music Revolution: Splash Music Scales AI with AWS Trainium
Splash Music, a leader in generative AI for music, partnered with AWS to democratize professional music creation through its advanced foundation models. The company’s flagship innovation, HummingLM, is a multi-billion-parameter, multi-modal model designed to convert human humming into high-fidelity instrumental tracks. HummingLM achieves this by fusing a transformer-based large language model with a specialized music encoder upsampler, utilizing Descript-Audio-Codec for efficient audio representation.
Prior to this collaboration, Splash Music grappled with significant challenges. The sheer complexity and scale of HummingLM demanded immense computational resources. The rapid pace of AI advancement necessitated continuous model training and deployment. Crucially, their reliance on externally managed GPU clusters led to unpredictable costs, latency, frequent interruptions, and burdensome manual management, severely impeding their ability to innovate and scale.
To address these issues, Splash Music adopted AWS Trainium chips on Amazon SageMaker HyperPod. This strategic move yielded substantial benefits: operational downtime was virtually eliminated, facilitating weekly model refreshes and faster feature deployment through automated, resilient, and scalable training orchestration. Training costs were slashed by over 54% compared to previous GPU-based solutions, and training speed more than doubled, accompanied by an 8% throughput improvement and increased batch sizes (70 to 512). These efficiencies enabled Splash Music to train larger models and accelerate its innovation pipeline.
HummingLM’s performance also saw marked improvements, achieving 57.93% better SI-SDR over baselines and demonstrating robust zero-shot generalization to unseen instruments. The robust AWS infrastructure, including Amazon EKS, FSx for Lustre (managing over 2 PB of data), and AWS Inferentia for inference, liberated Splash Music’s teams to concentrate on model design and music-specific research rather than infrastructure maintenance. Looking forward, Splash Music plans to expand its training datasets tenfold and explore multimodal audio/video generation, continuing its transformative collaboration with AWS.
