新一代Kaldi技术细节揭秘：K2是核心部分

{"type":"doc","content":[{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"2020年11月15日，由北京希尔贝壳科技有限公司、中国计算机学会语音对话与听觉专业组、AISHELL基金会主办，小米科技、昆山杜克大学、西北工业大学音频语音与语言处理研究组、中国科学技术大学共同协办的第五届Kaldi技术交流会在北京举办。"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"受疫情影响，本次技术交流会采用全天线上直播&下午（北京·小米科技园）线下技术交流会的形式。值得注意的是，本次线下交流活动，Kaldi之父Daniel Povey博士首次亲临现场，与来自北京各大互联网公司、知名高校的开发者们深入交流下一代Kaldi社区未来的发展。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"加入小米一年，Daniel Povey设计并开发出了新一代Kaldi。新一代Kaldi分成三个部分，包括核心算法部分，训练数据准备部分、示例脚本集合部分。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"其中，Lhotse（训练数据准备部分）将替代以前Kaldi中所有数据准备相关的工作，操作各种音频和文本的元数据。Lhotse除了Kaldi本身，也适用于其他应用。而且Lhotse纯Python代码，方便易用。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"Icefall（示例脚本集合部分）将代替Kaldi中的示例脚本集合，并独立成为一个单独的子项目。之所以要把示例脚本集合与核心算法分开，是考虑到示例脚本可能会非常庞大，且经常变动。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"新一代Kaldi的核心部分叫“k2”。k2可以让开发者很容易在PyTorch\/TensorFlow中实现各种语音识别相关算法，比如CTC、LF—MMI、RNN—T、2nd—pass语言模型等，消除以往语音识别算法中训练跟解码不匹配的问题。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"同时，通过k2可以非常容易实现（置信度逐渐提高的）多轮解码过程，这在以往是很难做到的。相较于其他一些语音识别库的优势，k2速度更快，通用性强（可以用来建模多种语音识别算法）。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"现场，据Daniel Povey博士透露，k2核心代码已完成。约41000行代码（主要是C++），本周刚发布0.1版本。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"Daniel Povey博士目前担任小米集团语音首席科学家，由他开发和维护Kaldi 集成了多种语音识别模型，公认是业界语音识别框架的基石。他在本次线下活动中强调：“今天有太多人依托Kaldi在做自己的事业，有很多人为Kaldi社区一直在做贡献，Kaldi会始终坚持开源。”"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"小米集团副总裁、技术委员会主席崔宝秋出席活动并指出，拥抱开源是小米工程文化的重要组成部分，我们要共同努力实现“四赢”：让Kaldi项目和Daniel赢，小米语音赢，全球的Kaldi社区赢，所有跟Kaldi相关的创业公司赢。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"最后，Daniel Povey博士感谢了所有人对社区的贡献，他欢迎更多中国和全球的工程师来社区贡献代码，共同为Kaldi和全球语音行业的发展做贡献。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}}]}