shangeth/libritts-r-mimi-codes

LibriTTS-R Mimi Codes是一个预提取的神经编解码器标记数据集，基于LibriTTS-R构建，专为TTS研究设计。与LibriSpeech相比，LibriTTS-R具有24 kHz的采样率（与Mimi原生匹配，无需重新采样）、句子级分段、保留标点符号以及应用了语音恢复的音频质量。数据集包含ID、文本、说话者ID、编解码器标记、帧数和编解码书数量等列。提取细节包括使用24 kHz、12.5 fps的kyutai/mimi编解码器，所有8个编解码书均已提取。数据集分割包括train_clean_100、train_clean_360、train_other_500、dev_clean、dev_other、test_clean和test_other。

LibriTTS-R Mimi Codes is a pre-extracted neural-codec tokens dataset based on LibriTTS-R, specifically designed for TTS research. Compared to LibriSpeech, LibriTTS-R features a 24 kHz sample rate (native to Mimi, no resampling needed), sentence-level segmentation, preserved punctuation, and applied speech restoration for audio quality. The dataset includes columns such as ID, text, speaker ID, codec tokens, number of frames, and number of codebooks. Extraction details involve using the kyutai/mimi codec at 24 kHz, 12.5 fps, with all 8 codebooks extracted. Dataset splits include train_clean_100, train_clean_360, train_other_500, dev_clean, dev_other, test_clean, and test_other.