NVIDIA Unveils Open-Weight Speech Model and Training Pipeline

内容摘要

NVIDIA has released a new open-weight model designed to handle multilingual transcription, speaker attribution, translation, and audio question answering within a single system. The tool is now available on Hugging Face, allowing developers to process recorded speech using natural language prompts instead of maintaining separate pipelines for each task. This unified approach supports transcribing 40 languages, translating 31 of them to or from English, and identifying speakers across 13 languages.

本文根据公开信息编译与整理，提供主题摘要及词汇辅助阅读。