在打下第一个字之前,先决定逐字稿的级别:完整逐字稿保留每一个“嗯”和重新开头,净化逐字稿保留意思而去掉口头语,编辑稿则为了便于阅读而整理语法。选定一套标注停顿、重叠和听不清片段的符号,把它写成文档顶部的说明,并始终如一地标注说话人。手工转录时,Texas Tech University Libraries 的口述历史指南给出的比例是每小时录音六小时打字;Cornell University 的一份课程指南说是四到六小时。自动转录几分钟内就能给你一份初稿;人工的工作从打字转向对照录音检查。
我们用 SozAI 访谈工具背后的引擎处理了 NASA Johnson Space Center 播客 Houston We Have a Podcast 第 180 期“Artemis Mission Management”的前 5 分钟(300 秒),主持人是 Gary Jordan,嘉宾是 Mike Sarafin——这是 NASA 的作品,因此在美国通常不受版权保护。它返回了 778 个词,检测到英语,并发现 2 位说话人。开头几行,按引擎的呈现如下:
[00:00] Speaker A: Houston, we have a podcast. Welcome to the official podcast of the NASA Johnson Space Center, episode 180, Artemis Mission Management. I'm Gary Jordan, and I'll be your host today.
[01:19] Speaker B: T-minus five seconds and counting. Mark.
[01:22] Speaker A: Launch commit lights are correct. There she goes. Houston. We have a podcast. Mike Serafin, thanks for coming on Houston, We Have a Podcast today.
[01:36] Speaker B: Thank you. It's a pleasure to be here to talk about these exciting Artemis missions we have ahead of us. Yeah, right?
[01:43] Speaker A: Returning to the Moon. Not a bad thing to be a part of.
仅这五次发言中,就有三处需要人工过一遍。01:19 的发射倒计时和 01:22 的“Launch commit lights are correct. There she goes.”是剪进播客里的发射存档音频,不是主持人或嘉宾在说话——引擎还是把它们标成了 Speaker B 和 Speaker A。01:36 那次发言末尾的“Yeah, right?”其实是主持人重新接过话头;它属于下一次发言的开头,而不是嘉宾发言的结尾。而嘉宾的姓氏在这里成了 Serafin,尽管本期节目自己的字幕署名拼的是 Sarafin。
稍后的一次发言展示了净化逐字稿的用处。引擎在 02:51 的输出是:“Yes, so you’re, you’re right in that I lived in the operations realm for 22 of my 27 years in, in my NASA career.”净化逐字稿处理会把它变成:“You’re right in that I lived in the operations realm for 22 of my 27 years in my NASA career.”完整逐字稿则会原样保留这些重复;如果说有什么倾向,引擎更倾向于保留它们——它在整个文件中保留了大部分重复和重新开头(“you’re, you’re right”、“in, in my NASA career”、“That’s, that’s quite a number of”),这使它的输出更接近逐字稿,而不是净化逐字稿。如果项目需要净化逐字稿,无论哪个引擎生成初稿,都要为一遍编辑留出预算。