WEIYANG SOUND · 未央声场
为什么还要做中文慢歌：在无限生成里练习选择
不是为了证明 AI 会写歌，而是想做出一首能陪人开车、失眠、回望，并愿意完整听到最后的中文歌。
创作动机 · 12 分钟
不是为了证明 AI 会写歌，而是想做出一首能陪人开车、失眠、回望，并愿意完整听到最后的中文歌。
This is not an attempt to prove that AI can write a song. It is an attempt to make one worth replaying on a long drive, through a sleepless night, and all the way to its final note.
01 · THE BEGINNING
起点不是“看看 AI 能不能”
我开始这一轮中文慢歌实验时，心里想的不是一首“展示 AI 能力”的歌。我想要的是更朴素、也更困难的东西：一首深情、易唱、旋律悠长的中文歌，一首放进汽车音响以后，不会因为听过两遍就厌倦，反而会让人顺手再按一次循环的歌。
这听起来不宏大，却是一个很硬的目标。工具的新鲜感可以维持几分钟，一首歌要陪人走夜路、收拾房间、想起旧人，靠的是另一套东西：一句话有没有落到心里，旋律有没有自己的去处，歌手有没有让人愿意相信的气息。技术可以把门打开，不能替作品走完这段路。
It Did Not Begin as an AI Demonstration
When I began this run of Chinese slow-song experiments, I was not trying to make a technical demonstration. I wanted something at once more ordinary and more difficult: a deeply felt, singable song with a melody that could breathe. I wanted a song that could enter a car stereo without becoming tiresome after two plays—a song that might make the listener reach for repeat before the silence at the end had fully arrived.
That may not sound like a grand ambition, but it is a demanding one. The novelty of a tool can hold attention for several minutes. A song that accompanies a night drive, a quiet room or the memory of someone absent needs another kind of strength. A lyric has to land. A melody has to know where it is going. A voice has to make its emotion believable. Technology can open the door; it cannot walk the song all the way through it.
02 · SLOW IS HONEST
慢歌不是把节拍放慢
慢歌最诚实。速度一慢，所有空洞都会露出来。歌词如果只是漂亮词语的堆积，旋律如果只是把几个长音拖开，听者很快就会走神；主歌没有方向、副歌没有抬升、第二遍只是复制，都会变得格外明显。
所以我听的不只是第一耳是否惊艳。我会听前奏有没有把门推开，主歌能不能把人带进去，副歌是不是情绪真正到达的地方，第二遍回来时有没有新的光线，尾声又是否给这一段感情留下落脚之处。所谓“慢”，其实是不给作品藏拙的机会。
Slow Songs Have Nowhere to Hide
A slow song is not simply a fast song with fewer beats. Slowness exposes every hollow place. If the lyric is only a pile of attractive phrases, if the melody merely stretches several notes, or if the second verse repeats the first without adding light or weight, the listener notices. There is less motion to conceal a weak decision and more time for each word to prove that it belongs.
So I do not judge only by the first impressive moment. I listen for whether the introduction opens a door, whether the verse leads us somewhere, whether the chorus is the true emotional arrival, and whether returning to it changes what we hear. I listen to the ending, too. A slow song needs somewhere to place the feeling it has raised. Its honesty comes from giving every choice enough time to be heard.
03 · HOOK & HEART
抓住人，和留住人，是两件事
短视频时代最有传播力的旋律，往往能在十几秒内抓住人。我也研究这样的片段，但不是为了照搬某一句旋律，而是想理解它们的情感语法：为什么一句话刚好落在那个音上？为什么副歌还没有唱完，听者已经知道下一句会怎样回来？
真正有效的“钩子”不一定突然大喊。它更像一句似曾相识的话，让人觉得它早就在自己心里。但钩子只能把人请进来，完整的段落、声音和情绪弧线，才决定他愿不愿意留下。传播可以短，作品不能被缩成一个效果。
A Hook Can Invite You In; It Cannot Make You Stay
In the era of short video, a melody with reach often reveals itself within seconds. I study those moments, but not to copy a recognizable phrase. I want to understand their emotional grammar: why does a particular word fall naturally on that note? Why does the listener sense the return of a line before the singer reaches it? Why does a small melodic turn feel as though it has always lived somewhere inside us?
A hook does not have to shout. It can be a sentence that sounds familiar before we know why. But a hook can only invite someone into the room. The complete verse, voice, arrangement and emotional arc determine whether that person stays. A song may travel through a short clip, yet it cannot be reduced to an effect without losing the very thing worth sharing.
04 · THE CAR TEST
“车里会不会再听一遍”
我的选择标准很生活化。我会问：这句能不能不看歌词也跟着唱？开车时听到副歌，会不会自然地把音量调高一点？歌曲结束以后，安静的两秒里，会不会想再听一遍？
未央声场的歌曲档案甚至长期保留“车载循环”这一项判断，因为音乐最终不是在参数表里完成的。它是在某个人独处、赶路、回忆或失眠的时候完成的。能进入日常，比只在第一次播放时显得厉害，更接近我心里真正的好歌。
The Car Test
My most useful standard is deliberately unglamorous. Can I sing along without looking at the lyrics? When the chorus arrives in the car, do I turn the volume up a little? After the last note, do the next two seconds of quiet make me want to hear the song again? These questions tell me more than a long list of production adjectives.
For that reason, the Weiyang Sound archive has long kept a judgment called ‘car-loop potential.’ Music is not completed in a parameter sheet. It is completed when someone hears it while driving, cooking, grieving, remembering or failing to sleep. Entering ordinary life matters more to me than looking impressive during a first playback. A song earns its place by becoming useful to feeling.
05 · ABUNDANCE
无限生成，没有取消创作
Suno 带来的最大变化，是把“可能性”一下子放大。不同速度、音色、情绪、语言和段落结构，可以很快变成听得见的版本。以前最难的是让一个想法发出声音，现在更难的是在声音的洪流里知道自己要什么。
机器可以在很短的时间里给出很多版本，却不会替我决定哪一个值得留下。创作并没有因为工具变快而变轻松，它只是把劳动从“做不出来”转移到了“如何选择”。生成是开始，比较、否定、等待、重做和最后那个不肯妥协的“就这一版”，才慢慢构成作品。
Infinite Generation Did Not Eliminate Creation
Suno changed the scale of possibility. Tempos, timbres, languages, moods and structures can become audible versions with extraordinary speed. In the past, the hardest step was often getting an idea to make sound at all. Now the harder task is knowing what one wants while standing in a flood of sound.
A system can produce many alternatives in a short time; it cannot decide which one deserves to remain. Creation did not vanish when generation became easier. The labour moved from ‘Can this be made?’ to ‘Why this version?’ Comparison, rejection, waiting, returning, remaking and finally refusing to compromise on one particular take—those decisions slowly turn output into a work. Generation is the opening move, not the authorship of the whole game.
06 · MANY VERSIONS
同一首歌，为什么还要做多个版本
同一首歌词放进不同的声线和编曲，不只是换一件衣服。有的声音在亲密的中文慢歌里成立，到了宽银幕管弦里反而失去人物；有的歌手在安全的抒情容器里显得普通，回到适合自己的语言、力度和场景，身份才突然清楚。
因此多个版本不是为了把数量堆高，而是在做判断实验。所谓“最好”也不是抽象的排行榜：它必须回答谁在唱、唱给谁、发生在哪里。只看平台热度，很容易把最熟悉的声音误当成最合适的声音。
Why One Song Needs More Than One Version
Placing the same lyric in another voice or arrangement is not merely changing its clothes. A voice that feels intimate in a Chinese ballad may lose its character inside a wide-screen orchestral setting. Another singer may sound ordinary in safe pop production, then become unmistakable when returned to the language, register and dramatic space that suit the voice.
Multiple versions are therefore experiments in judgment, not a way to inflate a catalogue. ‘Best’ is never an abstract ranking. It must answer: who is singing, to whom, and in what world? Popularity can reveal what is familiar, but the most familiar voice is not always the right one. Sometimes a song needs several failed identities before its true singer steps forward.
07 · THE VIRTUAL CAST
虚拟歌手不是一排预设音色
未央声场现在整理出十二位虚拟歌手。灵婉、Mary、林居怡、赛博，以及 Male Slow、宝莱坞男声、好莱坞大片女声、气声女音、芭蕉女、地中海男声、一直到永远和傣族阿妹，并不是十二个临时标签，而是十二条需要长期验证的声音路线。
赛博是清亮男高音歌手，不是“赛博风格”；摇滚是一种风格，也不是某个歌手。把歌手、声部和风格分开，作品才可能长出人物。一个虚拟歌手真正成立，不在于名字取得多漂亮，而在于听众经过几首歌以后，能不能认出他的天光、她的雨夜、她的厚度或他的远方。
A Virtual Singer Is Not a Preset
Weiyang Sound currently organizes twelve lead virtual singers: Lingwan, Mary, Lin Juyi, Cyber, Male Slow, Bollywood Male, Hollywood Female, Breathy Female, Banana-leaf Woman, Mediterranean Male, Forever and Dai Amei. They are not temporary labels attached to twelve presets. Each is a continuing vocal route whose identity must be tested across songs, registers, languages and arrangements.
Cyber is a clear male tenor singer, not a ‘cyber style.’ Rock is a style, not a singer. Keeping singer, vocal range and genre separate allows character to emerge. A virtual singer becomes real in the artistic sense only when, after several songs, listeners begin to recognize the brightness, rain, grain, distance or gravity of that voice. A name may start the fiction; repeated, coherent musical choices make the performer.
08 · OLD WORDS, NEW EARS
古诗不该只躺在课本里
做中文慢歌，很自然会走回唐诗、宋词、元曲和《红楼梦》。这些文字在成为经典以前，首先也是人的声音：思乡、离别、相爱、失意，看见江水、荷花、青山和落花时突然生出的感触。既然这些情绪今天仍然存在，它们就不该只停留在注释和考试里。
我不想只给古诗加一层“古风”包装。真正的问题是：一个今天戴着耳机、坐在车里的人，能不能先被旋律打动，然后愿意回头看看这句话从哪里来。古典感不是堆满古筝、烟雾和旧字体，而是让文字仍有分寸，让浓烈的情感不失去骨骼。
Ancient Poetry Should Not Live Only in Textbooks
Writing Chinese slow songs naturally led me back to Tang poetry, Song lyrics, Yuan songs and Dream of the Red Chamber. Before these works became classics, they were human voices: homesickness, separation, love, disappointment, and the sudden feelings stirred by a river, lotus fragrance, green mountains or falling flowers. Those emotions have not expired, so the words should not survive only as annotations and examination material.
I do not want to wrap an old poem in decorative ‘ancient style’ and call the work finished. The real question is whether someone wearing headphones today can first be moved by the melody, then become curious about where the words came from. Classical feeling does not require a warehouse of traditional instruments, mist and antique typefaces. It requires proportion: language with a skeleton, and intense emotion that does not collapse into excess.
09 · CULTURES IN MOTION
中西交融，不是贴上几个异国标签
这些实验也不断向外走。英文、法文、世界语，中英混合和多语言交错；宝莱坞、地中海、安达卢西亚、世界音乐和少数民族风；叶芝的诗、《红楼梦》的人物、世界文明主题和当代中文慢歌，都可以在同一个声场里相遇。
但跨文化不是把一种乐器、一个地名贴在歌曲表面。每种语言有自己的重音、口型和情感温度，每种音乐传统也有自己的结构。真正的交融需要让双方都保留来处，同时找到一条普通听众能走进去的路。陌生感不必被抹平，它可以成为继续听下去的理由。
Cultural Fusion Is More Than an Exotic Label
The experiments keep moving outward: English, French and Esperanto; Chinese-English songs and multilingual crossings; Bollywood, Mediterranean, Andalusian, world-music and minority traditions; Yeats beside Red Chamber characters, civilizational themes beside contemporary Chinese ballads. They can meet in one sound field, but a meeting is not the same as a mixture.
Cultural fusion cannot be achieved by attaching an unusual instrument or a foreign place-name to the surface of a song. Every language carries its own stress, mouth-shape and emotional temperature; every musical tradition carries structural memory. A genuine crossing lets both sides retain a visible origin while finding a path that an ordinary listener can enter. Strangeness does not always need to be smoothed away. Sometimes it is the reason to keep listening.
10 · WORDS THAT CAN BE SUNG
正式歌词和“发音保险稿”必须分开
AI 歌曲里有一类不容易被听众看见的工作：为了让多音字、外语或古典文字被正确唱出，有时需要准备发音更安全的输入稿。但那只是制作工具，不是应该公开的正式歌词。发布、字幕和文章引用，必须回到经过确认的规范文字。
这件小事提醒我，技术方便不等于可以含糊。一个字唱错、字幕和声音对不上、同名版本连错，都会伤害作品的身份。尊重听者，有时就体现在这些没有掌声的核对里。
The Text We Generate With Is Not Always the Text We Publish
AI music contains a layer of work that audiences rarely see. To help a system pronounce a polyphonic Chinese character, a foreign word or a classical phrase correctly, production may require a phonetic safety draft. That draft is a tool, not the official lyric. Published lyrics, subtitles and quotations must return to the verified and properly written text.
This small distinction carries a larger principle. Convenience does not excuse ambiguity. One mistaken character, a subtitle out of sync with the voice, or a link attached to the wrong version can damage a work's identity. Respect for listeners often appears in checks that win no applause: matching the title, confirming the singer, hearing the entire file and making sure the words on screen are the words being sung.
11 · THE SECOND CREATION
MV 是第二次创作
歌曲完成以后，影像不是随手加上的包装。人物、时代、色彩、镜头节奏、字幕出现的时刻，都要重新从歌曲的完整情绪弧线出发。特别是慢歌，画面不必每秒发生事情；让一个眼神、一片落花、一扇窗多停留一会儿，往往比密集剪辑更接近音乐。
竖屏也不只是平台妥协。它像一扇窄而高的窗，让听者离人物更近。好的 MV 应该让旋律从时间变成可以凝视的空间，同时守住完整、高质量的母版，再为不同渠道做合适的传播版本。
A Music Video Is a Second Act of Composition
After a song is complete, the moving image is not packaging added at the last minute. Character, period, colour, camera rhythm and the timing of subtitles must be composed again from the song's full emotional arc. Slow music in particular does not require a new event every second. A gaze, a falling flower or a window held for a little longer can live closer to the music than restless editing.
Vertical video is not merely a platform compromise, either. It can become a narrow, high window that brings us closer to a solitary figure. The music video turns melody from time into a space we can look at. My ideal process protects a complete, high-quality master first, then creates channel-appropriate versions for circulation. The format may adapt; the identity and emotional duration of the work should not be casually cut away.
12 · THE 45-SECOND LESSON
那次四十五秒，是一记很具体的提醒
网站最初把一段四十五秒音频放进播放器，听者刚被旋律带进去，歌就停了。那当然扫兴，也暴露出一个根本问题：入口被误当成了作品。短片可以用来让人遇见一首歌，但不能冒充一首歌已经被完整呈现。
现在未央声场的站内播放器只把完整 MP3 和完整 MV 当作可播放作品；没有核实的媒体地址，就诚实地不放按钮。尾声不是可有可无的剩余时间，它是情绪落地的地方。让一首歌从第一句走到最后一个尾音，是网站对听者最基本的尊重。
What Forty-Five Seconds Taught Me
An early version of the website placed a forty-five-second audio file inside a player. Just as the melody carried the listener into the song, it stopped. The disappointment was justified. The mistake exposed a basic confusion: an entrance had been presented as the work itself. A short excerpt can help someone encounter a song, but it cannot pretend to have presented the song in full.
The rule is now simple. An on-site player is for a complete, verified MP3 or music video. If a full media address has not been confirmed, the site should honestly offer no play button and instead lead to a verified work page. An ending is not leftover time. It is where emotion finds the ground. Letting a song travel from its first breath to its final tone is the minimum respect a music site owes the listener.
13 · WHEN AUTOMATION MUST STOP
自动化必须知道什么时候停下来
作品越来越多，整理、匹配、生成摘要和适配渠道当然需要自动化。但自动化一旦把同名版本、音频起点或画面身份连错，效率就会立刻变成噪声。一次地址能够打开，也不等于里面就是应该出现的那首歌。
所以我的原则越来越清楚：可以小批量前进，必须逐项核实；遇到歧义宁可暂缓，不能用猜测补齐一个看似漂亮的页面。工具应该扩大人的判断，而不是绕过人的判断。能停下来，是一套创作系统成熟的标志。
Automation Must Know When to Stop
As the catalogue grows, automation is necessary for organizing records, matching versions, generating summaries and adapting material for different channels. But when it connects the wrong same-named version, starts an audio file at the wrong point or assigns the wrong visual identity, efficiency turns into noise. A link that opens is not necessarily a link to the right song.
My operating principle has become clearer: move in controlled batches and verify each item; when identity is ambiguous, wait rather than filling a polished page with a guess. The purpose of a tool is to extend human judgment, not bypass it. Knowing when to stop is not a failure of an automated creative system. It is evidence that the system has learned what it is responsible for protecting.
14 · THE HUMAN PART
AI 是乐器，感动还得靠人
我希望未央声场最终不是一个炫耀生成数量的仓库，而是一座可以慢慢走进去的音乐房子。有人从一首中文慢歌进来，有人因为叶芝、世界语、法语香颂或宝莱坞旋律停下，也有人先认识某位虚拟歌手，再沿着他的声音听完整个系列。
AI 给了我一支前所未有的乐队，也给了我几乎无穷的岔路。真正属于人的部分，是知道为什么出发，替每一次选择负责，并在热闹过去以后仍愿意问：这首歌能不能陪一个人多走一段路？如果答案是能，那么这场中文慢歌实验就不是技术演示，而是在为今天的人，继续寻找可以唱、可以记、也值得完整听完的歌。
AI Is the Instrument; Feeling Still Belongs to People
I do not want Weiyang Sound to become a warehouse that boasts about the number of things generated. I want it to feel like a music house one can enter slowly. One visitor may arrive through a Chinese ballad; another may stop for Yeats, Esperanto radio, French chanson or a Bollywood-inflected melody. Someone else may first recognize one virtual singer, then follow that voice through an entire series.
AI has given me an orchestra unlike any I could have had before, along with almost infinite branching paths. The human part is knowing why we set out, taking responsibility for each choice, and continuing to ask after the excitement fades: can this song accompany one person a little farther? If it can, the slow-song experiment is no longer a demonstration of technology. It is a search for songs that people today can sing, remember and want to hear all the way to the end.
本文依据未央声场现有 202 条歌曲登记记录、创作者反馈、制作背景、声线路线与完整媒体验收原则整理。数字仅采用已核对的本地档案；未公开的版本判断、临时制作稿和无法确认的媒体链接不写入作品宣传。