Deepgram enhanced its voice AI capabilities by optimizing its Nova-3 speech-to-text model on AWS Trainium2, achieving a 11–70 percent increase in audio tokens processed per dollar. This marks the company's first major hardware optimization update since its initial deployment on Amazon SageMaker AI.
The model's performance was measured using instance pricing from the evaluation period, showing significant cost savings on offline batch workloads such as contact-center analytics and medical transcription. These results help Deepgram scale efficiently and deliver better models at lower prices to its customers.
The Nova-3 model is built on AWS Trainium2 and targets high-volume batch processing for organizations and developers. Availability for this optimization began with the deployment of the model on Amazon EC2 trn2.3xlarge instances, initially for customers using the Deepgram API or running models in their own environments.
"We measured this using instance pricing at the time of the evaluation," said Deepgram's engineering team. The optimization effort involved porting Nova-3 to the AWS Neuron software stack, tuning it for the new hardware, and improving throughput and cost efficiency.
The announcement follows Deepgram's earlier work on deploying its models on Amazon SageMaker AI. The company emphasized that the optimization helps customers self-host models in their VPCs while improving ease of setup, reliability, and observability.
Deepgram did not say how the optimization will affect real-time processing, and the company raised the question of whether further improvements can be made by combining Trainium2 with other AWS AI chips. The next step is to continue improving the model's performance and expanding its use cases.
Source: awsml
