VoiceBank+DEMAND

该数据集名为VoiceBank+DEMAND，包含了11,572个语音片段。其中，28位发言人的语音片段用于训练，另外2位发言人的语音片段用于测试。在训练阶段，共有10种类型的噪声与干净的语音混合，信噪比（SNR）在0至15分贝之间。而在测试阶段，5种类型的噪声与语音混合，信噪比在2.5至17.5分贝之间。该数据集涵盖了多种训练用的噪声类型，并已用于评估多种语音增强模型。整个数据集规模达到11,572个语音片段，任务重点在于语音增强。

The dataset named VoiceBank+DEMAND contains 11,572 speech segments in total. Speech segments from 28 speakers are utilized for training, while those from another 2 speakers are reserved for testing. During the training phase, clean speech is mixed with 10 types of noise, with the signal-to-noise ratio (SNR) ranging from 0 to 15 decibels (dB). In the testing phase, speech is mixed with 5 types of noise, and the SNR falls within the range of 2.5 to 17.5 decibels. This dataset covers a variety of training noise types and has been used to evaluate multiple speech enhancement models. With 11,572 speech segments overall, this dataset is focused on the task of speech enhancement.

搜集汇总

背景与挑战

背景概述

VoiceBank+DEMAND是一个专为语音增强和TTS模型训练设计的噪声语音数据库，包含干净和带噪的48kHz语音数据，来源于VCTK Corpus并整合了DEMAND数据库的噪声。