Chapter 11. A Voice from Tape第十一章 磁带里的声音
A manga for grown-ups. Read it in English or Chinese.给大人看的漫画。可以用中文或英文读。





























































































This chapter kept the hardest nights for the pictures. Here are the ideas in words.
On page 5 Haoran explains how he made Grandma's voice. In earlier chapters the machines answered "what is this?" or "what comes next?". A generative model does something else. It learns from many examples, then makes new ones that look or sound like them. Sound is a list of numbers, thousands of them every second. A voice program learns the patterns in those numbers and makes new sound a tiny piece at a time. In 2016 a system called WaveNet did this and sounded far more natural than the speech programs before it. It is a little like a language model writing one word at a time. It does not glue pieces of old recordings together, which is why Haoran says every note is new.
On page 6 a picture grows out of grey snow. A picture program is trained by taking real pictures and adding noise to them, step by step, until only static is left, and learning to guess the noise that was added. To make a new picture it starts from pure static and takes a little away at each step. Each start gives a different picture. The words you type are turned into numbers that nudge every step, and a setting decides how strictly the program follows them. Many systems work on a squeezed-down version of the picture to save time. They learn from hundreds of millions to billions of pictures with captions, collected from the web.
On page 7 twelve minutes of tape is enough. By now speech systems use transformers and diffusion-like methods, and they can imitate a particular voice from a short recording. Twelve minutes is the story's number, not a rule. What matters is how little it takes.
On page 13 the machine sings perfectly and Meilan sang flat. That detail belongs to the story, but the idea is real. The program learned what voices and singing in general sound like, from many voices, and learned Meilan's sound from twelve minutes. So it made a voice like hers singing the way people usually sing. It knew what voices sound like and had not learned her. It is wrong, too, to say a picture model never copies. In 2023 Carlini and colleagues showed that diffusion models can reproduce a few of their training pictures almost exactly, mostly pictures that appeared many times.
On pages 14 and 16 the same tools make fakes of real people. Grandma could not have said "Huihui, Mama is proud of you", and a program said it in her voice in two seconds. Fakes of this kind are called deepfakes. They fool people, and they also give liars a new excuse, because anyone can now say that a true recording is fake. Researchers call this the liar's dividend.
On page 16 Mama and Xiaoyu choose a family password. This is a real and simple defence. Families and companies agree on a secret word, and anyone who rings and asks for money has to say it first. A voice on the phone proves nothing any more.
On page 18 the tape gets a label, and Haoran deletes the file. Some cameras and apps now attach a signed record of how a picture or recording was made. Some programs hide a mark in what they make. Some laws now say that made content must be labelled, and that people must be told when they are talking to a machine. Tools that try to spot fakes help, but they work imperfectly. The oldest check still matters most. Where did it come from, and who says so?
In 2014 Ian Goodfellow and colleagues described a way to make pictures with two networks. One made images from random numbers, and the other tried to tell the real ones from the made ones. The paper compares them to counterfeiters and police. By the late 2010s the faces they made looked strikingly real. They were hard to train, and since about 2021 diffusion has largely replaced them for high-quality images.
The snow idea came from physics. Sohl-Dickstein and colleagues had it first, in 2015. In 2020 Ho, Jain and Abbeel made it work really well. Song and Ermon in 2019, and Song and colleagues later, found a related way to see the same thing.
Made pictures often get counting, hands, writing on signs and how things fit together wrong. The model learned what pictures look like, not how the world is built. They are much better every year and not yet perfect. The painters of this manga meet the same trouble, and every picture is checked finger by finger.
WaveNet in 2016 made raw sound one tiny sample at a time. In 2018 Tacotron 2 turned text into a picture of sound, then turned that into a voice. On the other side, recognising speech, Whisper in 2022 learned from 680,000 hours of audio with typed words.
In early 2024 a finance worker at the engineering firm Arup in Hong Kong sent 200 million Hong Kong dollars, about 25.6 million US dollars, after a video call in which the finance chief and other colleagues were all deepfakes. In January 2024 robocalls with an imitation of President Biden's voice told New Hampshire voters not to vote, and the US Federal Communications Commission fined the consultant behind them 6 million dollars.
The labels that show a recording was made by a machine are still new, and not every program adds them. Hidden marks can be damaged, and detectors make mistakes in both directions. The European Union's AI Act says that people must be told when they talk to an AI and that certain made content must be marked, from 2 August 2026.
Much of the data used to teach these programs was collected without asking, and many people feel it should have been their choice. A voice is part of a person. Whether you may copy it is a question of consent, and the law is still being written.
Bridge City, Xiaoyu, Grandpa Chen, Grandma Meilan, Mama, Haoran, Uncle Wu, Haoran's father and Ms Hart are made up. So are Meilan's tapes and her flat singing, Grandpa's song "The Moon on the Lake", the tape player and the radio factory, Haoran's voice program, the twelve minutes, the picture of 1975, Lantern, the flight and the family password. The story is set a few years from now. Grandpa's memory is drawn from no one in particular, and the story makes no medical claim. Whether a made voice of someone who has died helps or hurts is something families weigh, and there is no single right answer. Copying a real person's voice without their say-so raises real questions of consent, and in this story Grandma could not be asked.
Yingying Fu created this manga. She chose the story and its people, directed every page, and edited every word and picture. The words were drafted with Claude, an AI by Anthropic. The pictures are made with AI image generation, painted panel by panel from her direction and checked and corrected by hand.
这一章最难熬的几个夜晚,都留给了画面。这里再用文字把道理讲一讲。
第五页,浩然讲了外婆的声音是怎么做出来的。前面几章里的机器,回答的是“这是什么”或者“下一个是什么”。生成模型做的是另一件事。它从大量例子里学习,然后生成新的东西,看上去或听上去都和例子很像。声音其实是一串数,每秒钟有好几千个。声音程序学会这些数里的规律,再一小段一小段地生成新的声音。2016年,一个叫WaveNet的系统就是这样做的,听起来比以前的语音程序自然得多。这有点像语言模型一个词一个词地往下写。它并不是把旧录音的片段拼起来,所以浩然说,每个音都是新的。
第六页,一幅画从灰色的雪花点里慢慢显出来。画图程序是这样训练的。拿真实的图片,一步步往上加噪声,直到只剩一片雪花点,同时学着猜出加进去的噪声。生成新图时,它从纯噪声出发,每一步去掉一点。起点不同,画出来的图也不同。你输入的文字会被转成数字,在每一步上推它一把。还有一个参数,决定程序要多严格地照着文字来。为了省时间,很多系统先把图片压缩成小得多的版本,再在上面操作。训练用的图片从几亿到几十亿张不等,都是从网上收集来的,每张都带着文字说明。
第七页,十二分钟的磁带就够了。如今的语音系统用上了Transformer和类似扩散模型的方法,凭一段很短的录音就能模仿某个人的声音。十二分钟是故事里的设定,不是硬性规定。要紧的是,需要的录音少得惊人。
第十三页,机器唱得分毫不差,美兰却总跑调。这个细节是故事编的,背后的道理却是真的。程序从许多人的声音里学会了人声和唱歌通常是什么样,再从十二分钟的录音里学会了美兰的音色。所以它做出来的,是一个像她的声音,按大多数人的唱法在唱。它懂人声,却不懂她。不过,说图画模型从不照抄,也不对。2023年,卡利尼和同事发现,扩散模型能把少数训练图片几乎原样复现出来,这些图片大多在训练数据里出现过很多次。
第十四页和第十六页,同样的工具也能伪造真人。外婆从没说过“慧慧,妈妈为你骄傲”,程序却用她的声音,两秒钟就说出来了。这种伪造叫作深度伪造。它能骗人,也给说谎的人添了新借口,因为现在谁都可以说一段真录音是假的。研究者把这叫作“说谎者红利”。
第十六页,妈妈和小雨定了一个家庭暗号。这是个简单实用的防骗办法。家人之间或者公司内部约好一个暗语,谁打电话来要钱,都得先说出暗语。如今光听电话里的声音,已经证明不了什么了。
第十八页,磁带贴上了标签,浩然删掉了文件。现在有些相机和应用会给图片或录音附上一份带数字签名的记录,说明它是怎么生成的。有些程序会在生成的内容里藏一个水印。有些法律已经规定,生成的内容必须标注,人在跟机器对话时也必须被告知。检测伪造的工具也有用,但并不完全可靠。最老的办法依然最管用。它是从哪儿来的?是谁说的?
2014年,古德费洛和同事提出了一种用两个网络生成图片的方法。一个网络从随机数出发生成图片,另一个负责分辨哪些是真图、哪些是生成的。论文把它们比作造假币的人和警察。到2010年代末,它们生成的人脸已经逼真得惊人。不过这种网络很难训练,大约从2021年起,在高质量图片生成上,它基本被扩散模型取代了。
“雪花”的点子来自物理学。2015年,索尔-迪克斯坦和同事最早提出了这个想法。2020年,何、杰恩和阿贝尔让它真正好用起来。2019年的宋和厄尔蒙,以及后来宋和同事的工作,从另一个角度得出了相通的方法。
生成的图片常在这几处出错:数目、手、招牌上的字,还有东西之间怎么连在一起。模型学会的是图片看起来什么样,而不是世界是怎么构成的。它们一年比一年好,但还不完美。画这部漫画时,画师也碰到了同样的麻烦,所以每幅图都要一根手指一根手指地检查。
2016年的WaveNet直接生成原始声音,一次只生成一个采样点。2018年的Tacotron 2先把文字转成一张“声音的图”,也就是频谱图,再把它变成声音。另一头是语音识别,2022年的Whisper用68万小时带文字稿的音频训练而成。
2024年初,工程公司奥雅纳在香港的一名财务人员参加了一场视频会议,会上的首席财务官和其他同事全是深度伪造的。会后他汇出了2亿港元,约合2560万美元。2024年1月,有人用模仿拜登总统声音的自动语音电话,劝新罕布什尔州选民不要投票。美国联邦通信委员会对幕后的顾问罚款600万美元。
用标签表明一段录音是机器生成的,这种做法还很新,也不是每个程序都会加。隐藏的水印可能被破坏,检测工具也会两头出错,既会把真的当成假的,也会把假的当成真的。欧盟《人工智能法》规定,从2026年8月2日起,人们在与AI对话时必须被告知,某些生成内容也必须加以标注。
训练这些程序的数据,很多都是未经同意收集来的,不少人认为这本该由自己做主。声音是一个人的一部分。能不能复制别人的声音,归根到底是同意的问题,相关法律还在制定之中。
桥城、小雨、陈外公、外婆美兰、妈妈、浩然、吴爷爷、浩然的爸爸和哈特女士都是虚构的。美兰的老磁带和她跑调的歌声、外公的歌《湖上的月亮》、那台录音机和无线电厂、浩然的声音程序、那十二分钟、1975年的那幅画、小灯、那趟航班和家庭暗号,也都是编的。故事发生在几年以后。外公的记忆状况不以任何真人为原型,故事也不作任何医学上的判断。用生成的声音让去世的亲人再开口,是安慰还是伤害,要由每个家庭自己掂量,没有唯一正确的答案。未经本人同意就复制一个真人的声音,会引出实实在在的同意问题,而在这个故事里,外婆已经无从问起。
这本漫画由 Yingying Fu 创作。故事和人物由她决定,每一页由她导演,每一句话、每一幅画由她修改。文字是和 Anthropic 的人工智能 Claude 一起起草的。画面由人工智能生成,按她的要求一格一格地画,再逐张检查、修正。
© 2026 Yingying Fu. All rights reserved.© 2026 傅颖盈。保留所有权利。