Deepgram Announces New Observability Features for Amazon SageMaker AI
Deepgram has introduced new observability solutions with CloudWatch-integrated Enhanced Metrics and support for Prometheus and OpenTelemetry for Amazon SageMaker AI users.
AI company Deepgram announced two new observability innovations for speech-to-text (STT) and text-to-speech (TTS) models running on Amazon SageMaker AI. These innovations provide users with usage, billing, and engine-level visibility while maintaining data security within their AWS accounts.
Easy Tracking with Enhanced Metrics
The first innovation introduced by Deepgram, Enhanced Metrics, streams usage and billing data directly to Amazon CloudWatch without the need for any intermediary software, sidecars, or additional IAM permissions. This system operates in compliance with AWS Marketplace network isolation by utilizing CloudWatch Embedded Metric Format (EMF) logs written to the container standard output.
Under this feature, the "Deepgram/SageMakerInference" namespace makes it easier to verify AWS Marketplace invoices with metrics such as consumed units, audio duration, and character count. Meanwhile, the "Deepgram/SelfHosted" namespace allows for monitoring the usage of product-level features such as diarization, redaction, and smart formatting.
Prometheus and OpenTelemetry Support
The second innovation, Prometheus and OpenTelemetry support, provides engine-level and per-GPU granular visibility. The AWS-managed OpenTelemetry Collector running on SageMaker AI gathers metrics directly from the Deepgram container, GPU, and host exporters.
This collected data is ingested into CloudWatch's OTel-compliant metrics store, allowing it to be analyzed via PromQL queries using CloudWatch, Grafana, or any Prometheus-compatible tool. As a result, users can perform in-depth performance tracking without compromising the security of audio and transcript data within their own AWS accounts.