laion/dramabox-voice-acting-data-annotated
收藏资源简介:
DramaBox配音数据-标注版是laion/dramabox-voice-acting-data的后处理版本。处理流程包括:应用RE-USE语音增强技术对非唱歌样本进行降噪;使用LavaSR超分辨率将音频扩展到48kHz;通过Whisper Turbo ASR生成带词级时间戳的完整转录;在CUT TO:过渡处将音频分割为两部分(第1部分和第2部分);以及使用VoiceCLAP Large生成每部分的3584维语音嵌入。每个DramaBox样本包含来自同一说话者的两个情感场景,由CUT TO:分隔。通过分割成部分,创建了具有相同语音身份但不同情感的配对片段,从而支持训练配音系统,以在情感转换中保持说话者一致性。数据格式包括256kbps单声道48kHz的MP3文件,每个样本的JSON注释包含ASR、时间戳、嵌入向量和元数据,并打包为与源数据集结构匹配的tar文件。每个样本的文件包括:{prompt_id}_seed{NN}_part1.mp3(场景1音频)、{prompt_id}_seed{NN}_part2.mp3(场景2音频)和{prompt_id}_seed{NN}.json(完整注释)。使用的模型包括nvidia/RE-USE(语音增强)、YatharthS/LavaSR(超分辨率)、openai/whisper turbo(ASR)、laion/voiceclap-large(语音嵌入)和ResembleAI/DramaBox(源TTS模型)。
The annotated DramaBox Voice Acting Dataset is a post-processed version of laion/dramabox-voice-acting-data. The post-processing pipeline includes: denoising non-singing audio samples using the RE-USE speech enhancement technique; resampling audio to 48 kHz via LavaSR super-resolution; generating full transcripts with word-level timestamps using Whisper Turbo ASR; splitting audio into two segments (Segment 1 and Segment 2) at the "CUT TO:" transition; and generating 3584-dimensional speech embeddings for each segment using VoiceCLAP Large. Each DramaBox sample contains two emotional scenes from the same speaker, separated by the "CUT TO:" transition. By splitting into segments, paired segments with identical speaker identity but distinct emotions are created, which supports the training of voice acting systems to maintain speaker consistency during emotional conversion. The dataset uses MP3 files with 256 kbps, mono channel, and 48 kHz sampling rate. The JSON annotation for each sample contains ASR results, timestamps, embedding vectors, and metadata, and the dataset is packaged as tar files matching the structure of the source dataset. Files for each sample include: {prompt_id}_seed{NN}_part1.mp3 (audio for Scene 1), {prompt_id}_seed{NN}_part2.mp3 (audio for Scene 2), and {prompt_id}_seed{NN}.json (full annotation). The models utilized are: nvidia/RE-USE (speech enhancement), YatharthS/LavaSR (super-resolution), openai/whisper-turbo (ASR), laion/voiceclap-large (speech embedding), and ResembleAI/DramaBox (source TTS model).




