KaniTTS-research-team/emolia_filtered_v1

Emolia Filtered v1是一个包含103,521个样本的语音情感数据集，是laion/Emolia数据集经过audio_filter pipeline处理后的子集。该数据集保留了所有原始样本（包括质量好、差和不确定的样本），并添加了过滤结果作为额外列。数据集包含音频文件、文本转录、情感描述、说话人信息等多种内容，适用于语音和情感分析任务。数据处理分为两个阶段：质量检测（使用Dual LogisticRegression模型检测噪声、剪辑等问题）和说话人检测（使用Pyannote ONNX segmentation-3.0检测重叠说话人）。最终只有5.0%的样本通过了双重过滤（质量好且说话人清晰）。数据集还提供了详细的列描述和使用方法，方便用户筛选所需样本。

Emolia Filtered v1 is a speech emotion dataset containing 103,521 samples, which is a subset of the laion/Emolia dataset processed through the audio_filter pipeline. The dataset preserves all original samples (including good, bad, and uncertain ones) with filter results as additional columns. It includes various types of information such as audio files, text transcriptions, emotion descriptions, and speaker information, making it suitable for speech and emotion analysis tasks. The data processing consists of two stages: quality detection (using Dual LogisticRegression models to detect noise, clipping, etc.) and speaker detection (using Pyannote ONNX segmentation-3.0 to detect overlapping speakers). Only 5.0% of samples passed both filters (good quality and clear speaker). The dataset also provides detailed column descriptions and usage methods for easy sample filtering.