matrixdose/WaxalNLP

Waxal NLP Datasets 是一个多语言音频数据集，专注于非洲语言，支持自动语音识别（ASR）和文本到语音（TTS）任务。数据集包含多种语言，如阿坎语（ach、aka）、阿姆哈拉语（amh）、埃维语（ewe）、富拉语（ful）、豪萨语（hau）、伊博语（ibo）、斯瓦希里语（swa）等超过30种语言变体。数据来源于UGSpeechData、DigitalUmuganda/AfriVoice和原始收集，经过人类标注和众包处理。每个语言配置通常包括训练、验证、测试和未标注分割，特征包括音频文件、转录文本、说话者ID、语言代码和性别信息。数据集采用CC BY-SA 4.0和CC BY 4.0许可证，适用于语音技术研究和开发，旨在促进低资源语言的NLP进展。Waxal NLP Datasets is a multilingual audio dataset focused on African languages, supporting automatic speech recognition (ASR) and text-to-speech (TTS) tasks. The dataset includes multiple languages such as Akan (ach, aka), Amharic (amh), Ewe (ewe), Fulfulde (ful), Hausa (hau), Igbo (ibo), Swahili (swa), and over 30 language variants. Data is sourced from UGSpeechData, DigitalUmuganda/AfriVoice, and original collections, with human-annotated and crowdsourced processing. Each language configuration typically includes train, validation, test, and unlabeled splits, with features such as audio files, transcriptions, speaker IDs, language codes, and gender information. The dataset is licensed under CC BY-SA 4.0 and CC BY 4.0, suitable for speech technology research and development, aiming to advance NLP for low-resource languages.