LibriCSS

连续语音分离 (CSS) 是一种处理会话音频信号中重叠语音的方法。大多数先前的语音分离算法都在人工混合的预分割语音信号上进行了测试，因此通过隐式假设已经从输入音频中提取重叠区域来绕过重叠检测和说话人计数。 CSS 尝试通过在线处理直接处理连续传入的音频信号。在 [1] 中建立了主要概念，并在真实会议记录中评估了其有效性。由于这些录音是专有的，因此同一个研究小组在 [2] 中准备了一个名为 LibriCSS 的公开数据集。该存储库包含用于 LibriCSS 评估的程序。 LibriCSS 数据集可用于评估离线算法。

Continuous Speech Separation (CSS) is a method for handling overlapping speech in conversational audio signals. Most prior speech separation algorithms have been tested on artificially mixed pre-segmented speech signals, thus bypassing overlap detection and speaker counting by implicitly assuming that overlapping regions have already been extracted from the input audio. CSS attempts to directly process continuously incoming audio signals via online processing. The core concepts were established in [1], and their effectiveness was evaluated on real meeting recordings. As these recordings were proprietary, the same research team prepared a public dataset named LibriCSS in [2]. This repository contains the procedures for LibriCSS evaluation. The LibriCSS dataset can be used to evaluate offline algorithms.

搜集汇总

背景与挑战

背景概述

LibriCSS是一个公开数据集，专为评估连续语音分离（CSS）算法而设计，尤其适用于离线处理场景。该数据集由微软于2020年发布，基于真实会议记录，旨在解决重叠语音检测和说话人计数问题。