mesolitica/Malaysian-Speech-Instructions

该数据集是一个马来西亚语音指令数据集，它筛选了适合转换为语音提示的马来西亚短提示，并使用mesolitica/Malaysian-F5-TTS-v2生成了合成语音。数据集还通过Force Alignment进行了验证，以确保发音几乎正确。该数据集遵循Qwen/Qwen2-Audio-7B-Instruct的聊天模板，并包含来自多个子集的数据，包括短编码、评论家马来西亚、马来西亚通用问答等。对于超过20秒的音频（除了tatabahasa和MaLLM-Bench任务）以及马来西亚通用问答和马来西亚Alpaca中的短答案进行了过滤。

This dataset is a Malaysian speech instruction dataset that filters Malaysian short prompts suitable for conversion to voice prompts and generates synthetic speech using mesolitica/Malaysian-F5-TTS-v2. The dataset is also verified with Force Alignment to ensure that the pronunciation is almost correct. The dataset follows the chat template from Qwen/Qwen2-Audio-7B-Instruct and contains data from multiple subsets, including Short coding, Critics Malaysian, Malaysian General QA, etc. It filters audio longer than 20 seconds (except for the tatabahasa and MaLLM-Bench tasks) and short answers in Malaysian General QA and Malaysian Alpaca.