Trelis/libritts-snac-tokens

libritts-snac-tokens是一个基于LibriTTS-R音频数据集编码的语音表示数据集，使用hubertsiuzdak/snac_24khz模型进行分层残差向量量化（hierarchical RVQ）处理。该模型包含三个级别（12/24/48 fps），每个级别有4096个条目。数据采用Orpheus风格的交错格式，每音频帧包含7个令牌，速率为84 fps。词汇表偏移为12,288，分为三个级别范围：L0在[0, 4096)，L1在[4096, 8192)，L2在[8192, 12288)。解码时使用公式：level = token // 4096，code = token % 4096。数据集镜像了源LibriTTS-R的分割（经过parler-tts过滤），总时长约538小时，包含train.clean.100、train.clean.360、train.other.500和dev.clean等分割，各分割之间说话人不重叠。每个样本（每行）包含以下字段：id（源话语ID，格式为LibriTTS的speaker_chapter_segment）、speaker（说话人ID）、duration（音频时长，秒）、text（标准化文本）和codes（编解码器令牌列表）。数据集用于语音AI服务，提供加载方法，并有多个伴生数据集（如libritts-bpe-tokens等）。许可证为CC-BY-4.0，SNAC模型权重为MIT。

libritts-snac-tokens is a speech representation dataset encoded from the LibriTTS-R audio dataset using the hubertsiuzdak/snac_24khz model, which employs hierarchical residual vector quantization (RVQ) with three levels at 12/24/48 fps, each with 4,096 entries. The data uses an Orpheus-style interleave format, with 7 tokens per audio frame at a flat rate of 84 fps. The vocabulary offset is 12,288, divided into three level ranges: L0 in [0, 4096), L1 in [4096, 8192), and L2 in [8192, 12288). Decoding is performed with the formulas: level = token // 4096, code = token % 4096. The dataset mirrors the source LibriTTS-R splits (filtered by parler-tts), totaling approximately 538 hours, and includes splits such as train.clean.100, train.clean.360, train.other.500, and dev.clean, with no speaker overlap between train and dev splits. Each sample (row) contains the following columns: id (source utterance ID in LibriTTS speaker_chapter_segment format), speaker (speaker ID), duration (audio duration in seconds), text (normalized text from source), and codes (codec tokens list). It is designed for voice AI services, provides loading instructions, and has companion datasets (e.g., libritts-bpe-tokens). The license is CC-BY-4.0, with SNAC model weights under MIT.