Case Study: Resemble AI achieves 2x faster model training with Google Cloud AI Hypercomputer

Resemble AI cuts deepfake detection fine-tuning time by 99% with Google Cloud

Resemble AI, a company specializing in advanced voice and audio generative AI models, faced significant infrastructure challenges as their datasets grew beyond 60 terabytes. Their engineering team was spending up to 90% of their time on data preparation and pipeline management, which slowed model development cycles and created version control issues. To address this, they turned to vendor Apigee and Google Cloud for a modernization effort using AI-optimized tools including components of the AI Hypercomputer like A3 VMs and Hyperdisk ML.

Google Cloud implemented a solution combining Hyperdisk ML for high-throughput storage, A3 VMs for training, and Vertex AI for fine-tuning. This architecture, supported by Compute Engine and GKE, dramatically accelerated Resemble AI's workflows. The results included a 2x improvement in epoch cycle speed, a 99% reduction in deepfake detection fine-tuning time (from 7 days to 1 hour), and the ability to sustain over 100 inference requests per second with sub-250ms response times. This allowed their team to shift focus from data prep back to modeling, doubling research velocity.

Resemble AI