GTC dataset

该数据集来源于英伟达主题演讲，专注于计算机科学和工程领域，包含了大量的独特术语和产品名称。数据集还包括时长从2秒到35秒不等的音频文件片段，以及一个基于高频词汇和短语识别统计的上下文偏置列表。规模上，该数据集分为3小时的开发集和7小时的测试集。任务旨在为自动语音识别（ASR）进行上下文偏置。This dataset is sourced from NVIDIA's keynote speeches, focusing on the field of computer science and engineering, and contains a large number of unique technical terms and product names. The dataset also includes audio clips with durations ranging from 2 seconds to 35 seconds, as well as a context bias list based on recognition statistics of high-frequency vocabulary and phrases. In terms of scale, the dataset is divided into a 3-hour development set and a 7-hour test set. The task aims to perform context bias for automatic speech recognition (ASR).

搜集汇总

背景与挑战

背景概述

该数据集是用于训练英语流式自动语音识别模型的大规模复合数据集，包含数千小时的公开英语语音数据，如Librispeech、Fisher Corpus、Common Voice等。其特点在于支持缓存感知的FastConformer混合模型训练，并允许模型在不同延迟（如80ms至1040ms）下进行流式ASR应用，通过多前瞻机制优化性能。