TTS-AGI/ears-corrupted-dramabox-roundtrip
收藏资源简介:
EARS Corrupted DramaBox Roundtrip 是一个用于训练语音恢复模型的配对干净/腐败语音数据集。每个样本包含来自EARS数据集的干净录音,以及由DramaBox神经音频VAE生成的三种腐败变体。数据集包括17,227个样本,每个样本有4个音频文件(干净录音、标准VAE往返、噪声潜在往返和增强输入往返)和1个元数据文件,总音频时长约96小时(仅干净录音)或约383小时(所有变体),总大小约43 GB,音频格式为MP3、单声道、48 kHz、256 kbps。腐败变体模拟了神经编解码器伪影、潜在空间噪声和DSP增强效果,用于训练Sidon语音恢复声码器以去除DramaBox伪影并保持语音质量。
Paired clean/corrupted speech dataset for training speech restoration models. Each sample contains a clean recording from the EARS dataset alongside three corrupted variants produced by the DramaBox neural audio VAE. The dataset includes 17,227 samples, each with 4 audio files (clean, roundtrip, noisy roundtrip, augmented roundtrip) and 1 metadata file, total audio duration ~96 hours (clean only) or ~383 hours (across all variants), total size ~43 GB, audio format MP3, mono, 48 kHz, 256 kbps. The corruption variants simulate neural codec artifacts, latent space noise, and DSP augmentation effects, used to train the Sidon speech restoration vocoder for removing DramaBox artifacts while preserving speech quality.




