Amazon SageMakerMonitor & secureObservability

Deepgram Enhanced Metrics

Usage and billing metrics the Deepgram container publishes to your own CloudWatch account via Embedded Metric Format.

Beyond the metrics SageMaker itself publishes, the Deepgram container emits its own usage and billing metrics directly into your CloudWatch account. They are written as CloudWatch Embedded Metric Format (EMF) lines on container stdout: SageMaker forwards container output to the endpoint’s CloudWatch log group, and CloudWatch Logs extracts the metrics automatically. No agent, sidecar, or extra IAM permission is required, and the metrics keep flowing even in network-isolated Marketplace deployments.

Two namespaces are published, from two different layers of the container:

NamespaceLayerPurposeCan be disabled?
Deepgram/SageMakerInferenceDeepgram’s SageMaker integrationBilling: consumed units per request, with audio duration and character countsNo — always on
Deepgram/SelfHostedDeepgram API serverUsage: per-method, per-tier, and per-feature utilizationYes — on by default, opt out with an environment variable

Unlike the Prometheus & OpenTelemetry metrics, these are classic CloudWatch metrics: they appear in aws cloudwatch list-metrics, work with get-metric-statistics, dashboards, and alarms, and need nothing enabled on the endpoint configuration.

All dimensions are low-cardinality and contain no PII — never transcripts, TTS input, or per-request identifiers.

Billing metrics: Deepgram/SageMakerInference

One EMF record is emitted per completed request (each streaming session and each pre-recorded or TTS request). These are the same consumed-unit values that drive AWS Marketplace metered billing, so this namespace is the tool for reconciling your AWS bill against actual traffic.

MetricUnitDescription
ConsumedUnitsCountBillable inference units for the request. Sum over a period is the total billed volume.
AudioDurationSecondsSecondsDuration of audio processed (speech-to-text requests).
CharCountCountCharacters synthesized (text-to-speech requests).
BillingLatencyMsMillisecondsInternal latency of the billing pipeline. Useful only for diagnosing billing delays; not a request-latency metric.

Dimensions are published in three sets — [Category], [Category, Model], and [Category, Model, Transport] — so you can query at any granularity:

DimensionValues
Categorystt_streaming, stt_batch, tts
ModelThe deployed model, for example nova-3, flux-general-en, aura-2-thalia-en
Transportws (streaming), http (pre-recorded / batch TTS)

Example — total consumed units per hour, per category:

$aws cloudwatch get-metric-statistics \
>  --namespace Deepgram/SageMakerInference \
>  --metric-name ConsumedUnits \
>  --dimensions Name=Category,Value=stt_streaming \
>  --start-time $(date -u -d '24 hours ago' +%Y-%m-%dT%H:%M:%S) \
>  --end-time $(date -u +%Y-%m-%dT%H:%M:%S) \
>  --period 3600 \
>  --statistics Sum SampleCount \
>  --region YOUR_AWS_REGION

SampleCount is the number of billed requests; Sum is the units consumed.

Usage metrics: Deepgram/SelfHosted

The Deepgram API server emits a second EMF stream with raw usage broken down by method, model tier, and enabled features — independent of billing. Use it to understand how your endpoint is used: which features are enabled, how much audio each tier processes, and how much of the traffic is streaming versus pre-recorded.

MetricUnitDimensionsDescription
AudioMsMillisecondsDeployment, MethodAudio processed per request (speech-to-text).
RequestsCountDeployment, MethodCompleted requests.
TtsCharactersCountDeployment, MethodCharacters synthesized (text-to-speech).
TokensCountDeployment, MethodTokens consumed by intelligence features (summarization, sentiment, and so on).
VoiceAgentMsMillisecondsDeployment, MethodVoice-agent connection time.
TierAudioMsMillisecondsDeployment, TierAudio processed, broken down by model tier (for example nova-3, flux).
FeatureAudioMsMillisecondsDeployment, FeatureAudio processed with a given feature enabled. A request with several features enabled attributes its full duration to each.
FeatureTokensCountDeployment, FeatureTokens consumed per intelligence feature.

Dimension values: Deployment is sagemaker; Method is streaming or sync (pre-recorded); Feature is the request parameter name, for example diarize, smart_format, punctuate, redact, keyterm, interim_results.

Example — how much audio ran with diarization enabled today:

$aws cloudwatch get-metric-statistics \
>  --namespace Deepgram/SelfHosted \
>  --metric-name FeatureAudioMs \
>  --dimensions Name=Deployment,Value=sagemaker Name=Feature,Value=diarize \
>  --start-time $(date -u -d '24 hours ago' +%Y-%m-%dT%H:%M:%S) \
>  --end-time $(date -u +%Y-%m-%dT%H:%M:%S) \
>  --period 3600 \
>  --statistics Sum \
>  --region YOUR_AWS_REGION

Opting out

The usage stream is on by default. To disable it, set an environment variable override in the endpoint configuration’s Environment map (see Configure Amazon SageMaker Deployments):

1"Environment": {
2  "DEEPGRAM_API_01": "emf.enabled=false"
3}

The billing stream (Deepgram/SageMakerInference) cannot be disabled — it is part of the metering pipeline.

Because EMF metrics are extracted from the endpoint’s CloudWatch log group, they stop if the log group is deleted or the container’s logging permissions are removed. Standard CloudWatch metric pricing applies per unique metric/dimension combination; both streams keep dimension cardinality small by design.