Scaling Resemble AI’s Real-Time Voice and Deepfake Models with Cerebrium

Introduction

Resemble AI is trying to answer one of the hardest questions on the internet in under 300 milliseconds: is this real?

Its deepfake detection models analyze audio, images, and video inside live workflows such as phone calls and identity verification in real-time, where identifying a fake after the interaction ends can already be too late. Resemble's researchers don't only study synthetic media from the outside. They also build it: the company's open-source Chatterbox voice models have been downloaded more than 17 million times.

That gives Resemble AI an unusual view into the arms race between generation and detection. Its researchers have shown, for example, that something as simple as playing synthetic speech through a physical speaker and re-recording it can significantly degrade conventional deepfake detectors. As new attack techniques emerge, the team has to turn those discoveries into production models quickly.

Doing that requires infrastructure capable of serving a very different collection of models, hardware requirements, and traffic patterns - at real-time latency and across regions.

That's where Cerebrium comes in.

The Challenge

Resemble AI's models sit directly inside production workflows, from live voice calls to identity verification and fraud detection. Infrastructure performance isn't simply an optimization; it directly affects the product customers experience.

The team had previously been running workloads with another infrastructure provider but began running into limitations as it scaled:

1. Cold starts weren't fast enough for real-time use cases. For detection running during a live interaction, waiting for infrastructure to become available can make the product unusable.

2. Reliability needed to match the importance of the workloads. Resemble AI serves enterprises and government organizations where an unavailable detection endpoint can interrupt a critical customer workflow.

3. Different models required different infrastructure. A voice-generation model like Chatterbox doesn't necessarily have the same hardware profile or autoscaling behavior as a deepfake detection model. Resemble AI needed to deploy multiple architectures without designing a separate infrastructure stack for each one.

4. Customers increasingly required regional deployment and data residency. Supporting enterprises in new markets meant being able to run workloads in specific regions rather than routing every customer through the same infrastructure footprint.

Resemble AI needed a platform that could solve all of these problems without turning its research team into an infrastructure team.

Why Cerebrium

After hearing positive feedback from other real-time companies like Tavus, SuperDial, and Livekit - Resemble AI decided to explore an infrastructure partnership with Cerebrium.

Serverless GPU infrastructure allows capacity to expand with demand rather than requiring Resemble AI to permanently provision for peak traffic. That matters when a large verification batch, customer test, or security incident can cause traffic to increase suddenly.

Fast cold starts make the same infrastructure suitable for latency-sensitive, real-time endpoints. Cerebrium's multi-region infrastructure and failover capabilities also gave Resemble AI the reliability and geographic flexibility it needed. Workloads can be deployed closer to customers, while regional deployment options help Resemble AI satisfy data-residency requirements and pursue deals in markets that would otherwise be difficult to serve.

Importantly, Resemble AI gets those capabilities without sacrificing developer experience. Researchers can take a model from experimentation to a production endpoint without managing the underlying GPU fleet, autoscaling infrastructure, or regional deployment architecture themselves. The integration itself was straightforward: Resemble AI could bring its existing Dockerfile, define the deployment configuration in code, and fit Cerebrium into its existing CI/CD pipelines with minimal additional lift. From there, they can deploy their diverse model portfolio while configuring the hardware and scaling behavior around the needs of each workload, without being married to a single cloud provider as Cerebrium integrates their platform with cloud networks around the world.

The Cerebrium team also works directly with Resemble AI as new infrastructure requirements emerge.

“We talk directly with people who understand our platform needs and infrastructure deeply, rather than routing through a support tier. For a team our size, that is effectively a key extension of our own engineering capacity.”

— Tedi Papajorgji, CTO at Resemble AI

The Results

Since moving its workloads to Cerebrium, Resemble AI has been able to spend more of its engineering effort on the models themselves rather than the infrastructure surrounding them.

That has helped the team build an unusually fast research-to-production loop. Resemble AI now releases updated detection models weekly and sometimes multiple times per week. The company estimates that this is 30–100x more frequent than the quarterly or annual model refresh cycles common among other detection vendors.

That speed matters in an adversarial field. Every major improvement in generative media creates new artifacts for Resemble AI's researchers to study. The faster the team can identify those signals, retrain its models, and deploy them into customer-facing APIs, the faster customers can be protected against the newest generation of synthetic content.

Cerebrium has also given Resemble AI a common infrastructure layer across very different workloads, from open-source voice generation models like Chatterbox to latency-sensitive deepfake detection systems, without forcing the team into a single hardware or scaling configuration.

Multi-region deployment also gives Resemble AI the ability to support customers across the US, UK, and EU, helping it serve the data-residency requirements of the different markets in which it operates.

As Resemble AI expands deepfake detection across more customers and regions, its infrastructure needs to keep pace with a growing range of models, deployment environments, and latency requirements. Cerebrium gives the team the flexibility to serve those models reliably across regions while continuing to iterate quickly.

Resemble AI is building the intelligence layer. Cerebrium gives it the infrastructure to ship it.