LibriSpeech

该数据集名为Librispeech，包含了960小时的朗读英语语音，这些语音均源自有声读物。该数据集被用于在Fast-HuBERT模型中进行预训练和微调，包含了用于训练、开发和测试的各种子集。预训练阶段使用了960小时的数据，而微调阶段则使用了100小时的数据（即train-clean-100子集）。该数据集的任务是针对语音识别进行语音表征学习以及微调。

This dataset is named Librispeech, which contains 960 hours of spoken English speech all sourced from audiobooks. It is utilized for pre-training and fine-tuning the Fast-HuBERT model, and includes multiple subsets for training, development and testing purposes. For the pre-training phase, 960 hours of data are employed, while the fine-tuning phase uses 100 hours of data, specifically the train-clean-100 subset. The tasks of this dataset cover speech representation learning and fine-tuning for speech recognition.

搜集汇总

背景与挑战

背景概述

LibriSpeech是一个包含960小时朗读英语语音的数据集，语音源自有声读物，用于Fast-HUbert模型的预训练和微调。预训练使用全部960小时数据，微调使用100小时的train-clean-100子集，主要任务是语音识别中的语音表征学习和模型优化。