VoiceNet/emolia

emolia-balanced-5M-subset数据集是一个用于高质量音频-文本对比训练的语料库，经过重新打包处理。音频被重新编码为单声道FLAC格式（48 kHz，PCM 16位），并以WebDataset格式存储，每个样本包含配对的.flac音频文件和.json元数据文件。JSON元数据文件携带完整的注释信息：包括原始元数据（如id、text、duration、speaker、language、dnsmos）、基于情感标注标量生成的自由文本emotion_caption、54个数值型emotion_annotation标量（涵盖情感、音质、录音质量和人口统计信息），以及由MOSS-Audio-8B-Instruct模型生成的18个语音维度组（共59个短代码字段）。数据集包含约500万个样本，1052个分片，总大小约1.6 TB，支持德语、英语、法语、西班牙语和中文。主要用途是对比音频-文本训练（CLAP风格），其中text字段是训练目标，辅助注释可用于更细粒度的分析。许可证为CC-BY-4.0。

This is the emolia-balanced-5M-subset corpus repackaged for high-quality audio–text contrastive training. Audio is re-encoded as mono FLAC at 48 kHz (PCM 16-bit) and stored as a WebDataset of paired .flac and .json samples. The JSON sidecar carries the full annotation stack: original metadata (id, text, duration, speaker, language, dnsmos), a free-text emotion_caption derived from emotion-annotation scalars, 54 numeric emotion_annotation scalars covering emotion, voice quality, recording quality, and demographics, and 18 MOSS-Audio voice-dimension groups (59 short-code fields total) generated by MOSS-Audio-8B-Instruct. The dataset contains ~5 million samples, 1052 shards, total size ~1.6 TB, and supports languages: de, en, fr, es, zh. Intended use is contrastive audio–text training (CLAP-style) where the text field is the training target and auxiliary annotations are available for finer-grained probing. License: CC-BY-4.0.