TTS-AGI/dramabox-tuning-data
收藏资源简介:
DramaBox Tuning Data是一个用于微调DramaBox的配对训练数据集,包含12,330个样本,结合了情感语音(Emolia)和播客数据,采用紧凑的两部分格式。数据集支持英语、德语、西班牙语和法语,使用WebDataset格式(.tar分片),包含125个分片。每个样本包括元数据(.json文件)、张量数据(.pth文件)以及目标音频和参考音频(.mp3文件)。张量数据包含潜在表示(latent_part1和latent_part2)、条件嵌入(cond_part1_embeds和cond_part2_embeds)以及说话人嵌入(wavlm_speaker_emb和orange_speaker_emb)。数据集采用双向配对设计,每个音频对存在正向和反向两个方向,以增加训练数据,总共有24,660个训练对。提示生成使用基于LLM的方法(如Gemma-4-E4B和Gemini 3.5 Flash),并通过Gemma-3-12B-IT模型重新编码为条件嵌入。数据集与其他标准DramaBox数据集不同,使用未填充的条件嵌入、单.pth文件以及两部分结构,适用于配对/连续训练。
DramaBox Tuning Data is a paired training dataset for fine-tuning DramaBox, containing 12,330 samples that combine emotional speech (Emolia) and podcast data, and adopts a compact two-part format. This dataset supports English, German, Spanish and French, uses the WebDataset format (.tar shards), and consists of 125 shards. Each sample includes metadata (.json file), tensor data (.pth file), as well as target audio and reference audio (.mp3 files). The tensor data contains latent representations (latent_part1 and latent_part2), conditional embeddings (cond_part1_embeds and cond_part2_embeds), and speaker embeddings (wavlm_speaker_emb and orange_speaker_emb). The dataset adopts a bidirectional paired design: each audio pair has both forward and reverse directions to expand the training data, resulting in a total of 24,660 training pairs. Prompt generation uses LLM-based methods (e.g., Gemma-4-E4B and Gemini 3.5 Flash), and the conditional embeddings are re-encoded via the Gemma-3-12B-IT model. Unlike other standard DramaBox datasets, this one uses unpadded conditional embeddings, a single .pth file, and a two-part structure, making it suitable for paired/sequential training.




