AWS Announces WhisperX for Accurate Speaker-Labeled Transcription

AWS has announced the release of WhisperX, a Deep Learning Container (DLC) for Amazon SageMaker AI, which builds upon OpenAI’s Whisper model to provide accurate, speaker-labeled transcriptions. WhisperX adds per-word timestamps and speaker diarization capabilities to the Whisper model, addressing common issues in standard speech-to-text services such as imprecise timestamps and missing speaker labels.

WhisperX is designed to be deployed on Amazon SageMaker AI, offering both real-time and asynchronous endpoints. The real-time endpoint is suitable for short, interactive clips, while the asynchronous endpoint is recommended for long audio files and high-volume batch processing. The container serves on port 8080 and supports output formats such as JSON, SRT, and VTT, making it versatile for various use cases including contact center analytics, meeting notes, and media captioning.

The AWS WhisperX DLC is a maintained, GPU-ready image that includes Whisper, alignment models, and diarization weights, eliminating the need for a Hugging Face token. It adheres to the standard Amazon SageMaker AI serving contract, ensuring seamless deployment and integration with existing workflows.

In conclusion, WhisperX provides a robust solution for teams requiring precise, structured transcripts from spoken audio, enhancing capabilities for compliance review, sentiment analysis, and content management.

Source: aws-ml