Speaker-labeled transcription with WhisperX on SageMaker AI

This article demonstrates how to deploy the AWS WhisperX Deep Learning Container on Amazon SageMaker AI for production speech-to-text workloads with speaker diarization and precise timestamps.

• WhisperX extends OpenAI's Whisper with per-word timestamps, speaker labels, and faster batched inference for accurate transcription
• Deploy as real-time endpoint for short interactive clips (under 60 seconds) or asynchronous endpoint for long audio without time limits
• Real-time endpoint returns transcripts inline; asynchronous endpoint brokers input/output through S3 and supports polling
• Must pin GPU AMI version to al2-ami-sagemaker-inference-gpu-3-1 to avoid container startup failures
• Scale by adding instances with MaxConcurrentInvocationsPerInstance=1, not by increasing concurrency per container
• Supports multiple output formats (JSON, verbose JSON, SRT, VTT) for analytics pipelines and video editors
• Asynchronous endpoints can autoscale to zero for cost savings on bursty batch workloads

WhisperX on SageMaker AI enables accurate, searchable transcripts with speaker identification for contact centers, media, legal, and healthcare applications.