我的征尘是星辰大海。。。
The dirt and dust from my pilgrimage forms oceans of stars...
-------当记忆的篇章变得零碎，当追忆的图片变得模糊，我们只能求助于数字存储的永恒的回忆
作者:黄教授
手机视频列表
从复杂工程到一键解决
视频
音频
原始脚本
从复杂工程到一键解决，大模型如何简化语音转写校对全流程？最初的困境源于一个实际需求，剪映导出的视频字幕被烧录为 OSD，无独立 SRT 流，无法直接获取可编辑文本，而视频底稿又无法重新下载。 为了保留字幕修改痕迹，只能通过 FFmpeg 提取音频火山引擎 ASR 转写文本的路径还原内容。 但 ASR 的核心原则是忠实记录，基于短帧数据包识别，优先选择概率最高的字词，缺乏长上下文语义校验，导致同音字错误、人名不一致。 逻辑矛盾等问题频发。 比如男性角色被标注为他，人名路克被转写为路科路科。 最初的解决方案设计得相当复杂。 计划先让大模型识别 ASR 文本中的可疑错误，结构化存储为 JSON，标记错误字词位置，再用 FFmpeg 截取对应视频片段。 通过 OCR 识别烧录字幕，最后与可疑错误比对校对，整个流 程涉及音频提取、ASR 转写、错误标记、视频截取、OCR 识别、文本比对等多个环节，工程链路长且易出错。 但实际测试后发现火山引擎最新大模型的能力远超预期。 将 ASR 转写后的整篇文本直接输入，模型通过通读全文能自动校正95%以上的错误。 不仅精准修复同音字，代老、代老、代代，统一人名、路克 科陆科陆科，还能捕捉到短上下文无法发现的逻辑矛盾、它它，甚至修正断句、重复等格式问题。 最终输出的修正文本几乎无需二次校对，原本设计的复杂工程流程被一键输入一键输出的简单操作替代。 这背后的核心感慨是大模型的长上下文理解与语义推理能力正在重构复杂问题的解决路径。 很多时候我们习惯用多环节工程拆解的思维应对问题。 但当模型能够直接跨越识别理解校正的全链路时，复杂流程反而成了冗余。 技术进步的魅力正在于此，不是用更复杂的算法堆砌解决方案，而是用更强的基础能力让问题本身迎刃而解。
修正脚本
从复杂工程到一键解决，大模型如何简化语音转写校对全流程？最初的困境源于一个实际需求，剪映导出的视频字幕被烧录为 OSD，无独立 SRT 流，无法直接获取可编辑文本，而视频底稿又无法重新下载。 为了保留字幕修改痕迹，只能通过 FFmpeg 提取音频，再经火山引擎 ASR 转写文本的路径还原内容。 但 ASR 的核心原则是忠实记录，基于短帧数据包识别，优先选择概率最高的字词，缺乏长上下文语义校验，导致同音字错误、人名不一致、逻辑矛盾等问题频发。 比如男性角色被错转写为她，人名路克被转写为路科。 最初的解决方案设计得相当复杂。 计划先让大模型识别 ASR 文本中的可疑错误，结构化存储为 JSON，标记错误字词位置，再用 FFmpeg 截取对应视频片段。 通过 OCR 识别烧录字幕，最后与可疑错误比对校对，整个流程涉及音频提取、ASR 转写、错误标记、视频截取、OCR 识别、文本比对等多个环节，工程链路长且易出错。 但实际测试后发现火山引擎最新大模型的能力远超预期。 将 ASR 转写后的整篇文本直接输入，模型通过通读全文能自动校正95%以上的错误。 不仅精准修复同音字、代词，统一人名路克，还能捕捉到短上下文无法发现的逻辑矛盾、错字，甚至修正断句、重复等格式问题。 最终输出的修正文本几乎无需二次校对，原本设计的复杂工程流程被一键输入一键输出的简单操作替代。 这背后的核心感慨是大模型的长上下文理解与语义推理能力正在重构复杂问题的解决路径。 很多时候我们习惯用多环节工程拆解的思维应对问题。 但当模型能够直接跨越识别理解校正的全链路时，复杂流程反而成了冗余。 技术进步的魅力正在于此，不是用更复杂的算法堆砌解决方案，而是用更强的基础能力让问题本身迎刃而解。
英文翻译
From complex engineering to one-click solutions: How large models simplify the entire process of speech transcription and proofreading? The initial dilemma stemmed from a practical need: video subtitles exported from CapCut were burned into OSD without an independent SRT stream, making it impossible to directly obtain editable text, while the original video footage could not be re-downloaded. To preserve the editing traces of subtitles, the only option was to extract audio via FFmpeg, then use Volcengine ASR to transcribe the audio into text to restore the content. However, the core principle of ASR is faithful recording, based on short-frame packet recognition, prioritizing the most probable words, and lacking long-context semantic validation. This led to frequent issues such as homophone errors, inconsistent names, and logical contradictions. For example, a male character was mistakenly transcribed as "she," and the name "Luke" was transcribed as "Luke" (with incorrect Chinese characters). The initial solution was designed to be quite complex: first, have the large model identify suspicious errors in the ASR text, store them structurally in JSON, mark the positions of erroneous words, then use FFmpeg to capture the corresponding video clips. Next, OCR was used to recognize the burned-in subtitles, and finally, a comparison was made with the suspicious errors for proofreading. The entire process involved multiple steps—audio extraction, ASR transcription, error marking, video capture, OCR recognition, and text comparison—making the engineering chain long and error-prone. However, actual testing revealed that the latest large model from Volcengine far exceeded expectations. By directly inputting the entire ASR-transcribed text, the model could automatically correct over 95% of errors by reading through the full context. It not only accurately fixed homophones, pronouns, and unified the name "Luke," but also caught logical contradictions, typos, and even corrected formatting issues like sentence breaks and repetitions that short-context approaches could not detect. The final output required almost no secondary proofreading, and the originally designed complex engineering process was replaced by a simple one-click input and output operation. The core insight here is that the long-context understanding and semantic reasoning capabilities of large models are reshaping the approach to solving complex problems. Often, we are accustomed to tackling issues by breaking them down into multi-step engineering processes. But when the model can directly span the entire chain from recognition to understanding to correction, complex workflows become redundant. That is the beauty of technological progress: not piling up more complex algorithms to build a solution, but leveraging stronger foundational capabilities to make problems solve themselves.
back to top