AISHELL-4

该数据集名为AISHELL-4，提供了120小时的会议录音，每场会议通常有4至8名参与者。录音采用的是8麦克风、5厘米半径的圆形麦克风阵列，通常放置在会议桌的中央位置。该数据集适用于进行说话人识别、语音活动检测以及重叠语音检测等任务。

The dataset is named AISHELL-4, which consists of 120 hours of meeting audio recordings. Each meeting typically involves 4 to 8 participants. The recordings are collected via a circular microphone array with 8 microphones and a 5-centimeter radius, which is generally positioned at the center of a conference table. This dataset is suitable for tasks including speaker recognition, voice activity detection, and overlapping speech detection.

搜集汇总

背景与挑战

背景概述

AISHELL-4是一个多通道中文普通话会议语音开源数据集，于2021年1月发布，包含211场会议、约120小时的8通道录音，每场会议有4至8名说话人。该数据集模拟真实会议场景，涵盖停顿、重叠、噪声等特征，并提供精确的音字转写文本和时间戳，旨在支持语音前端处理、语音识别和说话人分割等多说话人处理任务的研究与开发。