Trelis/libritts-mimi-tokens

libritts-mimi-tokens 是一个基于 LibriTTS-R 数据集编码的音频数据集，使用 kyutai/mimi RVQ 编解码器（包含 8 个码本，每个码本有 2,048 个条目，帧率为 12.5 帧/秒）进行编码。数据集包含两个流：codes_semantic（仅码本 0，基于 WavLM 蒸馏，内容对齐，帧率为 12.5 帧/秒，词汇表大小为 2,048）和 codes_all_flat（所有 8 个码本交错排列，帧率为 100 帧/秒，使用偏移词汇表，总词汇表大小为 16,384）。数据来源于 parler-tts/libritts_r_filtered，并保留了原始 LibriTTS-R 的分割（包括训练集和开发集，总时长约 538 小时）。每个话语包含以下列：id（话语标识符）、speaker（说话者标识符）、duration（音频时长，秒）、text（标准化文本）、codes_semantic（语义编码列表）和 codes_all_flat（所有码本扁平化编码列表）。数据集遵循 CC-BY-4.0 许可证。

libritts-mimi-tokens is an audio dataset based on the LibriTTS-R dataset encoded with the kyutai/mimi RVQ codec (with 8 codebooks of 2,048 entries each, at 12.5 frames/sec). It includes two streams: codes_semantic (codebook 0 only, WavLM-distilled and content-aligned, at 12.5 fps with a vocabulary of 2,048) and codes_all_flat (all 8 codebooks interleaved per frame, at 100 fps with an offset vocabulary of 16,384). The data is sourced from parler-tts/libritts_r_filtered and mirrors the original LibriTTS-R splits (including training and development sets, totaling approximately 538 hours). Each utterance contains columns: id (utterance identifier), speaker (speaker identifier), duration (audio duration in seconds), text (normalized text), codes_semantic (list of semantic codes), and codes_all_flat (list of all codebooks flattened). The dataset is licensed under CC-BY-4.0.