akan-tts-collect
收藏资源简介:
该数据集是一个多任务语音与翻译集合,包含三个子集:自动语音识别(ASR)、文本转语音(TTS)和机器翻译(MT)。数据通过一个本地离线优先的应用程序收集,并同步到 Hugging Face Hub,最后同步时间为 2026-08-10。数据集采用 Apache-2.0 许可证。ASR 子集包含 30 条已朗读的语音及其转写文本,可用于训练 ASR 模型;TTS 子集包含 30 条录音室级别的朗读语音,可用于训练 TTS 模型;MT 子集当前为 0 条样本,计划包含源文本及人工翻译。支持的语言包括 tw、ak、fat、ee、gaa、dag、gur、nzi、dga、kus、ha、en 等非洲语言及英语。音频为 16 位单声道 WAV 格式,录制设备原生采样率通常为 48 kHz,建议在训练时重采样以避免有损转换。每条记录包含字段:text(ASR/TTS 的朗读提示或 MT 的源句子)、target_text(仅 MT 子集,翻译结果)和 contributor(贡献者假名)。数据仅在被贡献者勾选同意框时写入。使用时可通过 huggingface_hub 和 datasets 库加载,ASR 和 TTS 子集以 audiofolder 格式加载,MT 子集以 JSON 格式加载。
This dataset is a multi-task speech and translation collection, comprising three subsets: Automatic Speech Recognition (ASR), Text-to-Speech (TTS), and Machine Translation (MT). The data is collected via a local offline-first application and synced to Hugging Face Hub, with the last sync date being 2026-08-10. The dataset is licensed under Apache-2.0. The ASR subset contains 30 read-aloud speech recordings with their transcriptions, suitable for training ASR models; the TTS subset contains 30 studio-quality read-aloud speech recordings, suitable for training TTS models; the MT subset currently has 0 samples, intended to include source texts and human translations. Supported languages include tw, ak, fat, ee, gaa, dag, gur, nzi, dga, kus, ha, en, and other African languages as well as English. Audio is in 16-bit mono WAV format, with native sampling rates typically at 48 kHz; resampling is recommended during training to avoid lossy conversion. Each record contains fields: text (the prompt for ASR/TTS or the source sentence for MT), target_text (only for MT subset, the translation), and contributor (contributor pseudonym). Data is written only when the contributor checks the consent box. It can be loaded using the huggingface_hub and datasets libraries, with ASR and TTS subsets loaded as audiofolder format, and the MT subset loaded as JSON format.




