资源简介:
该数据集是原始laion/gemini-2.5-pro-tts-voice-profiles数据集的增强版本,专门用于文本到语音(TTS)任务,特别关注语音表演、情感表达和语音配置文件生成。核心改进在于为每个样本新增了两个由google/gemma-3-12b-it模型生成的DramaBox风格表演提示词,这些提示词基于样本原有的标注(如BUD-E Whisper转录、Empathic-Insight情感分数、Gemini词级转录/描述和发声行为描述)生成,避免了重新运行自动语音识别(ASR)或Whisper模型的需求。每个样本的JSON文件在保留原始所有字段的基础上,新增了以下字段:1) `dramabox_prompt_full`:完整的DramaBox表演提示词,包含说话者的人口统计信息(如年龄、性别、音色)、情感舞台指示、对话,并将发声行为(如呼吸声、叹息)融入舞台指示中;2) `dramabox_prompt_focused`:聚焦的DramaBox提示词,不包含人口统计信息,仅关注音质、情感表达和对话,适用于需要参考音频或无偏TTS的场景;3) `dramabox_model`:生成提示词的模型ID;4) `dramabox_meta.full_starter` / `focused_starter`:用于增加提示词开头多样性的随机起始词或短语。提示词遵循固定格式:首先是一个以...delivering this high-quality studio voice recording with no background noise.结尾的说话者/声音描述段落,随后是交替出现的`(舞台指示)`和`"对话"`。示例展示了多样化的说话者角色、复杂的情感状态(如优越感/悲伤、愠怒/嫉妒、羞怯/厌恶、愉悦/满足、勇气/失望等)以及丰富的非语言发声行为(如咳嗽、啜泣等)。该数据集旨在为高质量、富有表现力的语音合成提供丰富的、带有详细表演指导的文本输入,适用于需要控制语音情感、风格和表演细节的TTS研究和应用。
This dataset is an enhanced version of the original laion/gemini-2.5-pro-tts-voice-profiles dataset, specifically designed for text-to-speech (TTS) tasks, with a focus on voice performance, emotional expression, and voice profile generation. The core improvement involves adding two DramaBox-style performance prompts generated by the google/gemma-3-12b-it model for each sample. These prompts are based on the samples original annotations (such as BUD-E Whisper transcriptions, Empathic-Insight emotion scores, Gemini word-level transcriptions/descriptions, and vocal behavior descriptions), eliminating the need to rerun automatic speech recognition (ASR) or Whisper models. Each samples JSON file retains all original fields and includes new fields: 1) `dramabox_prompt_full`: a complete DramaBox performance prompt containing speaker demographic information (e.g., age, gender, timbre), emotional stage directions, dialogue, and integrating vocal bursts (e.g., breathing, sighs) into the stage directions; 2) `dramabox_prompt_focused`: a focused DramaBox prompt without demographic information, emphasizing only voice quality, emotional expression, and dialogue, suitable for scenarios requiring reference audio or unbiased TTS; 3) `dramabox_model`: the model ID used to generate the prompts; 4) `dramabox_meta.full_starter` / `focused_starter`: random starting words or phrases to increase prompt diversity. Prompts follow a fixed format: starting with a speaker/voice description paragraph ending with ...delivering this high-quality studio voice recording with no background noise., followed by alternating `(stage directions)` and `"dialogue"`. Examples showcase diverse speaker roles, complex emotional states (e.g., superiority/sadness, resentment/jealousy, shyness/disgust, joy/satisfaction, courage/disappointment), and rich non-verbal vocal behaviors (e.g., coughing, sobbing). The dataset aims to provide rich, detailed performance-guided text input for high-quality, expressive speech synthesis, applicable to TTS research and applications that require control over voice emotion, style, and performance details.