AWS has published a deployment guide for the WhisperX Deep Learning Container (DLC), a GPU-ready image that bundles OpenAI's Whisper speech-to-text model with wav2vec2 forced alignment and speaker diarization. The container follows SageMaker AI's standard serving contract, so teams deploy it to a real-time or asynchronous endpoint without assembling the pipeline themselves or supplying a Hugging Face token.
The core problem WhisperX addresses: generic transcription gives utterance-level timestamps (off by several seconds) and no speaker labels. That breaks compliance review, captioning, and redaction workflows. WhisperX adds per-word timestamps and speaker tags, producing output in json, verbose_json, srt, or vtt formats.
AWS frames two deployment patterns. Real-time endpoints suit short, interactive clips that finish within SageMaker's 60-second response cap, billing while the endpoint is up. Asynchronous endpoints broker input and output through S3, remove the time cap, and can autoscale to zero when idle, aimed at long recordings and high-volume batch processing.
Production details AWS flags as essential: GPU variants require pinning InferenceAmiVersion to al2-ami-sagemaker-inference-gpu-3-1, since the default host AMI's drivers fail to start the CUDA 12.8 image; inference is serialized to one request per container, so throughput scales by adding instances, not concurrency; and instance choice runs from ml.g4dn.xlarge for cost to ml.g5.2xlarge for headroom.
AWS's own listed use cases include contact-center talk-time measurement, script adherence and sentiment analysis, searchable meeting notes, captioning for media and e-learning libraries, and speaker-labeled transcripts for audits and legal discovery in healthcare, legal, and finance.