Deepgram 用 AWS Trainium2 优化语音 AI 推理
原标题：How Deepgram optimized voice AI at scale with AWS Trainium
AI 摘要
Deepgram 与 AWS 合作，将其 Nova-3 语音转文字模型从标准 PyTorch 代码移植到 AWS Trainium2 芯片，用于离线批量工作负载（如呼叫中心分析和医疗转录）。通过 TorchNeuron 编译、NKI 内核调优和 Rust 数据并行运行时，最终在单台 trn2.3xlarge 实例上运行八个模型副本，每美元处理的音频量比原有方案提升 11%–70%。这一优化帮助 Deepgram 以更低成本扩展语音 AI 服务，并惠及在 Amazon SageMaker AI 上自托管其模型的客户。
正文节选
How Deepgram optimized voice AI at scale with AWS Trainium As voice becomes a first-class modality for AI, Deepgram’s models span speech-to-text, text-to-speech, and voice agents. They now process quadrillions of audio tokens a year for more than 200,000 developers and 1,400 organizations. Customers use them through the Deepgram API or run them in their own environments, including on Amazon SageMaker AI. For Deepgram, serving efficiency determines how much audio the system can process for a give