#

Key takeaways

• AWS published a technical guide for WhisperX on SageMaker AI.

• The container combines Whisper, wav2vec2 forced alignment, and speaker diarization.

• It supports structured transcripts with timestamps and speaker labels.

• AWS describes both real-time and asynchronous deployment options.

• The post gives deployment details, prerequisites, and cost controls.

AWS publishes a deployment guide

AWS published “Speaker-labeled transcription with WhisperX on SageMaker AI” on September 24, 2026. The post is a technical deployment guide. It is not a claim that WhisperX was created by AWS or newly released.

The guide describes an AWS Deep Learning Container for deployment to Amazon SageMaker AI. The container packages Whisper, wav2vec2 forced alignment, and speaker diarization in a GPU-ready image. It is built for speech transcription workflows that need more than plain text output.

What WhisperX adds

WhisperX builds on OpenAI’s open-source Whisper automatic speech recognition model. AWS says it adds batched inference, word-level timestamps, and speaker labels. That combination allows audio to be returned as structured transcripts.

AWS lists four output formats: JSON, verbose JSON, SRT, and VTT. These formats support different downstream uses. The guide presents them as output options, not as measured performance claims.

How the container works

The container follows SageMaker’s serving contract. It listens on port 8080, accepts inference requests at `/invocations`, and exposes `/ping` for health checks. Requests use multipart form data.

The guide says requests may include language, diarization, and response-format fields. That makes the deployment configurable at inference time. The post also includes deployment prerequisites and points to an AWS Samples notebook.

Endpoint options and scaling

AWS compares synchronous real-time endpoints with asynchronous endpoints. Real-time endpoints suit short interactive clips. The guide says real-time responses must fit SageMaker’s 60-second limit.

Asynchronous inference is presented for long audio or higher-volume batch jobs. It submits and retrieves audio through S3 and can scale to zero when idle. The post also details GPU instance selection, a required GPU AMI pin, single-request-per-container scaling, and cost controls.

Example workloads in the guide

AWS names several example workloads where timestamps and attribution can matter. These include contact-center calls, meetings, podcasts, depositions, broadcast media, and regulated-domain review. These are examples only. The post does not present them as customer results.

The guide focuses on deployment mechanics rather than business outcomes. It explains how to package the model, how to send requests, and how to choose an endpoint type. That makes it useful for teams evaluating speech transcription pipelines.

Governance and operational considerations

The source highlights operational choices that affect reliability and cost. Endpoint type matters because short clips and long audio have different needs. GPU selection and the required AMI pin also affect deployment setup.

The guide’s mention of speaker labels and word-level timestamps suggests a need for careful transcript handling. That is an operational consideration, not a legal or regulatory claim. The post does not add country-specific compliance guidance.

Morocco relevance

The source reports no Morocco-specific facts, availability, or adoption. A conditional global lesson is that teams evaluating transcription systems should check endpoint limits, output formats, and scaling behavior before deployment.

Bottom line

AWS’s guide shows how WhisperX can be deployed on SageMaker AI as a GPU-ready container. It combines transcription, alignment, and diarization into one workflow. The result is a structured transcription setup with multiple output formats and two endpoint patterns.

For readers comparing deployment options, the main value is practical. The post explains how to serve audio, how to format requests, and how to choose between real-time and asynchronous inference. It stays focused on implementation details rather than product claims.