AWS says Deepgram gains 11-70 percent more audio per dollar on Trainium2

Deepgram's Nova-3 speech recognition model processed 11-70 percent more audio per dollar on AWS Trainium2 than on the compute options Deepgram had been running, AWS reported. The evaluation covered offline batch transcription workloads, including contact-center analytics and medical transcription, using instance prices at the time of testing. Deepgram and AWS ported the model from standard PyTorch code, tuned its operations and built a Rust runtime that coordinates eight model replicas on one trn2.3xlarge instance. At the batch size that maximized throughput, the Rust runtime delivered a 7.5% throughput gain over the multiprocessing Python baseline. The results concern throughput-oriented batch processing rather than live voice-agent response times.