ruslan-stressed
收藏资源简介:
RUSLAN with Word Stress Marks(带词重音标记的RUSLAN)是一个基于RUSLAN俄语单说话人语音合成(TTS)语料库的增强版本数据集。其核心特点是在原始RUSLAN语料库的转录文本中,为所有多音节俄语单词以及单音节词(包括代词和介词)添加了明确的词重音标记。该数据集的创建旨在解决训练俄语TTS模型时的一个关键难题:标准俄语正字法不编码词汇重音,导致模型难以从原始文本中学习正确的重音位置,尤其是对于同形异义词(例如‘за́мок’(城堡)与‘замо́к’(锁))。通过提供带有重音标注的文本,该数据集消除了此类歧义,预期能显著提升合成语音的韵律自然度。数据集包含22,200个语音-文本对,总计约31小时的俄语语音。音频部分与原始RUSLAN语料库完全一致,为22 kHz采样率的单声道WAV文件。文本部分则是在原始转录基础上,使用组合锐音符(U+0301)在重读元音后标注了重音(例如‘тре́вожным’)。字母‘ё’因其固有重音属性而不额外标注。数据以WebDataset格式组织,音频和文本文件分别打包成三个tar分片存储,并附带一个以管道符分隔的metadata.csv文件,便于仅使用文本数据的用户。该数据集专为训练或微调需要俄语精确重音信息的TTS模型(如Kokoro、Tacotron、VITS、StyleTTS、XTTS等)而设计。其数据来源于CC BY 4.0许可的RUSLAN语料库,重音标注由Anthropic Claude Opus模型自动完成,并经过了自动验证流程,但标注中可能仍存在极少量的同形异义词错误。
RUSLAN with Word Stress Marks is an enhanced version of the RUSLAN Russian single-speaker text-to-speech (TTS) corpus. Its core feature is the addition of explicit word stress marks to the transcriptions of the original RUSLAN corpus, covering all polysyllabic Russian words as well as monosyllabic words (including pronouns and prepositions). The dataset is designed to address a key challenge in training Russian TTS models: standard Russian orthography does not encode lexical stress, making it difficult for models to learn correct stress positions from raw text, especially for homographs (e.g., за́мок (castle) vs. замо́к (lock)). By providing text with stress annotations, this dataset eliminates such ambiguities and is expected to significantly improve the prosodic naturalness of synthesized speech. The dataset contains 22,200 speech-text pairs, totaling approximately 31 hours of Russian speech. The audio portion is identical to the original RUSLAN corpus, consisting of mono WAV files at a 22 kHz sampling rate. The text portion is based on the original transcriptions, with stress marked using the combining acute accent (U+0301) after stressed vowels (e.g., тре́вожным). The letter ё is not additionally marked due to its inherent stress properties. The data is organized in WebDataset format, with audio and text files packaged into three tar shards, accompanied by a pipe-separated metadata.csv file for users who only need text data. This dataset is specifically designed for training or fine-tuning TTS models that require precise Russian stress information (e.g., Kokoro, Tacotron, VITS, StyleTTS, XTTS). The data is sourced from the RUSLAN corpus under the CC BY 4.0 license, with stress annotations automatically generated by the Anthropic Claude Opus model and subjected to an automated validation process, though minimal homograph errors may still be present in the annotations.




