TTS-AGI/podcast-tokenized-bg3.5-enj5-with-speaker-embeddings
收藏资源简介:
该数据集是TTS-AGI/podcast-tokenized-bg3.5-enj5的扩展版本,增加了说话人嵌入、余弦相似度分数、说话人聚类分配和参考匹配标志。每个样本的JSON元数据包含以下字段:target_speaker_embedding(目标音频的L2归一化说话人嵌入,128维)、ref_speaker_embedding(参考音频的L2归一化说话人嵌入,128维)、speaker_cosine_similarity(目标与参考嵌入的余弦相似度)、ref_matches_target(参考和目标是否为同一说话人,基于余弦相似度阈值0.42)、speaker_cluster_id(基于10,000个预计算说话人质心的最近聚类分配)。数据集使用DACVAE模型将潜在表示解码为48kHz音频,并使用Speaker-wavLM-tbr模型从16kHz音频中提取128维说话人嵌入。处理流程包括解码、截断、重采样、嵌入提取、相似度计算和聚类分配。数据集包含1,828,199个样本,分布在481个tar文件中,85.4%的样本参考音频与目标音频来自同一说话人,适用于语音克隆和说话人条件TTS任务。聚类分析基于10,000个质心,分布右偏,中位聚类大小为63个样本。还提供了说话人相似度阈值分析,推荐阈值用于高置信度说话人匹配。
This dataset is an extended version of TTS-AGI/podcast-tokenized-bg3.5-enj5, with added speaker embeddings, cosine similarity scores, speaker cluster assignments, and reference matching flags. The JSON metadata for each sample includes the following fields: target_speaker_embedding (L2-normalized 128-dimensional speaker embedding extracted from the target audio), ref_speaker_embedding (L2-normalized 128-dimensional speaker embedding extracted from the reference audio), speaker_cosine_similarity (cosine similarity score between the target and reference speaker embeddings), ref_matches_target (indicator of whether the reference and target audio originate from the same speaker, determined using a cosine similarity threshold of 0.42), and speaker_cluster_id (nearest cluster assignment based on 10,000 pre-computed speaker centroids). The dataset uses the DACVAE model to decode latent representations into 48 kHz audio, and employs the Speaker-wavLM-tbr model to extract 128-dimensional speaker embeddings from 16 kHz audio. The processing pipeline consists of decoding, truncation, resampling, embedding extraction, similarity calculation, and cluster assignment. This dataset contains 1,828,199 samples spread across 481 tar files; 85.4% of the samples have reference and target audio from the same speaker, and it is suitable for voice cloning and speaker-conditioned TTS tasks. Cluster analysis is based on 10,000 centroids, with a right-skewed cluster size distribution and a median cluster size of 63 samples. Speaker similarity threshold analysis is also provided, with recommended thresholds for high-confidence speaker matching.




