WEIYANG SOUND · 未央声场
写上“灵婉”还不够：一次“绣花嗓” Sample 实验
从《鹊桥小调·灵婉戏曲》前一分钟出发，四组题目生成八个尚未发布的候选；实验不急着宣布成功，而是追问一位虚拟歌手怎样跨过新歌、风格和专辑仍被认出来。
声音连续性 · Sample 实验 · 19 分钟
从《鹊桥小调·灵婉戏曲》前一分钟出发，四组题目生成八个尚未发布的候选；实验不急着宣布成功，而是追问一位虚拟歌手怎样跨过新歌、风格和专辑仍被认出来。
The first minute of Magpie Bridge Air becomes the seed for four two-take tests. The eight candidates remain unpublished while the experiment asks how a virtual singer can cross songs, styles and an album without losing her identity.
01 · THE NAME IS A PROMISE
“灵婉”先是一个承诺，不是一条自动生效的声纹命令
给虚拟歌手取名很容易。难的是几首歌以后，这个名字仍然指向同一个可辨认的人物。未央声场把灵婉写成清亮、细腻而有骨骼的女声：句子像绣线一样一针一针送出，能唱古典离别，也能走进校园、七夕、鲁迅与摇滚群唱。可是把“灵婉”两个字写进下一次生成，并不会自动兑现这些特征。
名字更像编辑承诺：以后每次使用它，都要拿新声音和已经留下的作品互相核对。音色可以有变化，人物不能毫无理由地消失。虚拟歌手的连续性不靠一张设定卡完成，而靠创作者一次又一次听见偏差、拒绝偏差，或者承认这次变化已经长成另一副声音。
Lingwan Is First a Promise, Not a Voiceprint Command That Executes Automatically
Naming a virtual singer is easy. The difficult part is making the name point to a recognisable character after several songs. Weiyang Sound describes Lingwan as clear, finely articulated and grounded: a voice that carries a sentence like embroidery thread and can move from classical farewell into school memory, Qixi, Lu Xun and a rock chorus. Typing Lingwan into another generation does not automatically deliver those qualities.
The name is closer to an editorial promise. Every new use must be compared with the works already kept. Timbre may change; character should not disappear without reason. Continuity is not completed by a character card. It grows when the creator hears deviation, rejects it, or recognises that a change has become another voice altogether.
02 · WHY SAMPLE
这次不用形容词追声音，而把已经听得见的一分钟交给下一首歌
实验选取《鹊桥小调·灵婉戏曲》的 00:00–01:00 作为 Sample 窗口。选择这一分钟，不是因为它能代表灵婉的一切，而是它保存了一组已经发生的关系：清楚咬字、细线般的高音、带戏曲气质的转折，以及声音没有被配器压平时的近距离。它比“delicate”“clear”“opera-inflected”这些形容词更具体。
Sample 在这里不是复制捷径，而是一块可听的参照物。新的歌词、旋律和场景仍要自己站住；被借来的那一分钟只负责把问题问得更尖锐：当起点真正带着灵婉的声音进入生成，后面的作品会不会更稳定地保留她？
Instead of Chasing a Voice with Adjectives, the Experiment Hands One Audible Minute to the Next Song
The experiment takes 00:00–01:00 of Magpie Bridge Air · Lingwan as its Sample window. That minute cannot represent everything Lingwan can do. It does preserve a relationship that has already happened: clear diction, a fine supported upper line, opera-inflected turns and a close vocal distance that the arrangement does not flatten. It is more concrete than delicate, clear or opera-inflected written in isolation.
Sample is not a shortcut to duplication here. It is an audible reference. New lyric, melody and scene must still stand independently; the borrowed minute only sharpens the question. When generation begins with Lingwan actually sounding, rather than merely being named, does the resulting work retain her more reliably?
03 · FOUR ROOMS, EIGHT TAKES
四个新房间，每个房间两次回答；八个候选先留在实验室
这轮共有四组题目：《旧校门的雨》《水袖过桥》《一封没寄出的信》《雨窗小调》，每组两个候选，共 8 个结果。题目故意把灵婉放进不同距离：校园记忆、戏曲身体、没有寄出的私人文字和靠近耳边的雨窗。若四个房间都只得到同一种表面古风，Sample 就没有真正守住人物，只守住了装饰。
八个结果当前都没有发布，也没有被标记喜欢或不喜欢。这里不提前写“实验成功”，更不把生成完成冒充作品完成。它们是一次有明确问题的候选集：下一步仍要逐首完整听，记录哪一版保住了绣花嗓，哪一版被源歌影子牵得太紧，哪一版在后半段失去人物。
Four New Rooms, Two Answers in Each; Eight Candidates Remain inside the Lab
The four prompts are Rain at the Old School Gate, Water Sleeves Cross the Bridge, A Letter Never Sent and A Little Song at the Rain Window. Each produced two candidates, for 8 results. The titles place Lingwan at deliberately different distances: school memory, theatrical body, private writing and rain beside the ear. If every room produces only the same generic ancient-style surface, the Sample has preserved decoration rather than character.
None of the eight results is published, liked or disliked at the time of writing. This essay does not declare the experiment successful simply because generation completed. The candidates still need complete listening: which take holds the embroidered voice, which remains trapped beneath the source song's shadow, and which loses the character after its opening?
04 · PERSONA, SAMPLE, COVER, MASHUP
四种方法不是四个时髦按钮，它们回答的是不同问题
在未央声场的工作语言里，Persona 先命名一条持续声线路线；Sample 把一段具体声音当作新生成的参照；Cover 让既有作品在新的演唱、结构或人物选择中重新回答自己；Mashup 则把不同来源或音乐关系推到同一场碰撞里。它们可以互相连接，却不能被混成“都只是换风格”。
这不是替所有平台和音乐学制定统一词典，而是为了让本站的创作记录不含糊。每次实验先说明用了什么、想检验什么，再由完整成品决定它最后属于版本、姐妹歌、新作还是只应留在工作档案。方法不能替作品命名，工具标签也不能替人的判断。
Four Methods Are Not Four Fashionable Buttons; They Ask Different Questions
In Weiyang Sound's working vocabulary, Persona names a continuing vocal route; Sample supplies a concrete piece of sound as a reference for new generation; Cover asks an existing work to answer itself through another performance, structure or character choice; Mashup deliberately brings different sources or musical relationships into one collision. They may connect, but they do not all mean changing style.
This is not a universal dictionary for every platform or for musicology. It is a way to keep this archive's production record precise. State the method and the question first; let the complete result determine whether it becomes a version, sister song, new work or an item that remains in the lab. A tool label cannot name the final work on behalf of human judgment.
05 · THE THREE-EAR TEST
第一耳听绣花嗓，第二耳听源歌阴影，第三耳一直听到结尾
评估第一问很直接：绣花嗓还在不在？这里说的不是复制某个瞬间的频谱，而是字头是否清楚、气息是否贴近、细高处是否有支撑、情绪是否仍然带着灵婉那种克制的骨骼。声音可以换房间，人物不能只剩一个名字。
第二问是源歌阴影是否过强。如果新歌一开口就让人只能想到《鹊桥小调》，新的文字和旋律便可能没有获得自己的空间。第三问必须听到结尾：后段会不会越唱越虚、音色开始发飘，或者高潮为了证明力度突然过火。一个好开头不能替后半首歌通过验收。
First Hear the Embroidered Voice, Then the Source Shadow, Then Stay until the Ending
The first question is direct: does the embroidered voice remain? This does not mean reproducing one moment's spectrum. It asks whether consonants stay clear, breath stays close, the fine upper line keeps support, and emotion retains Lingwan's restrained spine. A voice may enter another room; the character cannot survive as a name alone.
The second question asks whether the source song casts too heavy a shadow. If every entrance only recalls Magpie Bridge Air, the new lyric and melody may not have earned space of their own. The third question demands the ending: does the later voice turn airy and indistinct, or does a climax suddenly overfire to prove its power? A strong opening cannot pass the second half by proxy.
06 · THE ALBUM GATE
单曲“可以发”，不等于专辑“应该收”
创作中已经出现过这种难题：一首歌单独听有可取之处，却因为后段声线发飘、清晰度下降，或者高音和高潮突然过火，没有进入原定专辑。拒绝它并不是否认整首歌，而是在保护一张专辑内部更严格的声音关系。
专辑需要的不只是若干首各自不错的歌。它还需要人物在相邻曲目之间保持可信的距离，响度和高潮不互相争抢，结尾能够把下一首歌交出来。生成时代真正稀缺的编辑行为，往往不是再加一首，而是敢于说：这一首可以独立存在，但现在还不属于这张唱片。
A Track Can Be Publishable without Belonging on the Album
The creative process has already produced this difficult result: a song can have clear standalone value and still be withheld from its intended album because the later voice floats away, clarity declines, or the upper register and climax suddenly become excessive. Rejecting that placement does not reject the entire song. It protects a stricter relationship inside the album.
An album needs more than several individually good songs. Character must remain credible between adjacent tracks; loudness and climaxes cannot all compete; one ending must know how to hand the listener to the next. In an age of abundant generation, the scarce editorial act is often not adding another track but saying: this may exist independently, yet it does not belong on this record now.
07 · BEGINNINGS ARE NOT ENOUGH
Sample 锁住开头以后，最危险的地方反而在一分钟以后
声音种子天然会让人注意开头是否相似，可一位歌手的身份要经过整首结构才成立。第二段主歌的咬字、桥段的气息、最终副歌的力度和尾声的收束，才会暴露声音有没有在生成惯性里逐渐漂走。Sample 若只让前六十秒漂亮，它仍然没有解决专辑级问题。
因此下一轮提示方向会特别保护后半程：结尾清楚、稳定，最终副歌克制，不要为了制造戏剧性而突然拔高，也不要把细腻误解成没有支撑的气声。要求不是“永远别唱高”，而是让每一次抬升都属于人物和歌曲，而不是模型自动添加的壮观。
Once Sample Holds the Opening, the Most Dangerous Place Begins after the First Minute
A voice seed naturally draws attention to opening resemblance, but a singer's identity must survive a whole structure. Diction in verse two, breath through the bridge, force in the final chorus and control in the outro reveal whether the voice gradually drifts under generative momentum. If Sample makes only the first sixty seconds beautiful, it has not solved the album-level problem.
The next prompt direction therefore protects the latter half: a clear stable ending, a restrained final refrain, no sudden ascent merely to manufacture drama, and no confusion of delicacy with unsupported airiness. The rule is not never sing high. Every lift must belong to the character and the song rather than arrive as automatic spectacle.
08 · THE SOURCE SONG MUST REMAIN WHOLE
种子只取一分钟，来源作品仍然必须让听者完整听完
实验窗口是 00:00–01:00，但《鹊桥小调·灵婉戏曲》在本站仍以完整 3:36 MP3 与影片出现。Sample 的剪取是制作方法，不是面向听者的缩水发行。想理解为什么选这段声音，最诚实的办法仍是把原作从第一声听到尾声。
这也划开了实验证据和公开作品的边界。页面下方四首歌都是已经核实的完整灵婉母带；上面四组八个结果则仍是待评候选。可播放的不假装是 Sample 实验的新歌，待评的也不偷借既有播放器制造已经发布的印象。
The Seed Uses One Minute; the Listener Still Deserves the Entire Source Work
The lab window is 00:00–01:00, but Magpie Bridge Air remains a complete 3:36 MP3 and film on this site. Cutting a Sample is a production method, not a reduced public release. The most honest way to understand why the minute was chosen is still to hear the source from first sound to outro.
This separates experimental evidence from released work. The four recordings below are verified complete Lingwan masters. The four groups and eight new results above remain candidates awaiting judgment. The playable works do not impersonate the new tests, and the pending tests do not borrow an existing player to create the appearance of publication.
09 · LISTEN ACROSS STYLES
同一人物要跨过元曲叙事、昆曲气质、晚唐夜雨和七夕小调
只听一首来源歌，很容易把人物误认成一种风格。所以下方导听把《鹊桥小调》与《长亭送别》01A、昆曲风味 02A、《玉炉香》并列。两首《长亭送别》甚至保留同一组文字与同一 Persona，却以不同速度、句尾和戏曲密度证明：灵婉不是“昆曲风”三个字。
四首合计 12:22。这不是为 Sample 候选背书，而是一条公开的基线：听者可以先认识已经成立的灵婉，再理解为什么新实验要同时检验清亮、骨骼、距离和结尾。人物连续性不是追求每首一模一样，而是在变化以后仍听得见是谁作出了这次选择。
One Character Must Cross Yuan-Drama Narrative, Kunqu Colour, Late-Tang Rain and a Qixi Air
One source song can make character look like genre. The listening route therefore places Magpie Bridge Air beside Long Pavilion 01A, Kunqu-inflected 02A and Incense from the Jade Burner. The two Long Pavilion recordings even keep the same text and Persona while different tempos, endings and theatrical density prove that Lingwan is not another name for Kunqu style.
The four works run 12:22. They do not endorse the pending candidates. They establish a public baseline: hear the Lingwan who already exists, then understand why the new lab must test clarity, spine, distance and ending together. Continuity does not require every song to sound identical. It means that after change, we can still hear who made this choice.
10 · FAILURE IS DATA
发飘、过火和太像源歌，都不是丢脸的废片
当目标是稳定一位虚拟歌手，失败版本会告诉我们名字、Persona、Sample 窗口和后段约束各自做到了什么。声线发飘说明人物支撑没有贯穿结构；高潮过火说明作品的戏剧弧线压过了歌手分寸；太像源歌则说明新歌没有挣到足够的自主空间。
这些判断值得进入档案，但不必全部进入公开曲库。保存失败的理由，比保存一串没有上下文的文件名更有价值。它让下一次生成不只是重新摇一次骰子，而是带着上一次听见的问题继续工作。
Floating, Overfiring and Resembling the Source Too Closely Are Not Embarrassing Waste
When the goal is to stabilise a virtual singer, failed versions explain what the name, Persona, Sample window and later-section constraints each achieved. A floating voice shows that character support did not survive the structure. An overfired climax shows the dramatic arc overwhelming the singer's proportion. Excessive source resemblance shows the new song failing to earn autonomy.
Those judgments deserve a place in the archive without every file entering the public catalogue. Preserving why a candidate failed is more valuable than preserving a row of contextless file names. The next generation can then continue with a heard problem rather than simply rolling the dice again.
11 · A SCALABLE ALBUM ENGINE
如果这条路成立，它会成为专辑引擎；如果不成立，也要知道卡在哪里
Sample 路线最吸引人的地方，是它可能把一副已经成立的声音带进一批新歌，让专辑不必每首都从零寻找主唱。但可扩展不等于批量复制。四组题目正是在测试：同一声音种子面对不同叙事时，能否保留人物，同时让作品各有场景、旋律和情绪命运。
真正可用的专辑引擎需要三层闸门：每首歌自己的完整结构、主唱跨曲目的连续性，以及整张唱片的高潮与结尾秩序。任何一层失败，都可以退回单曲、实验或档案状态。这样扩容才不会把“产量”误写成“阵容”，也不会让灵婉被自己的名字淹没。
If the Route Works, It Can Become an Album Engine; If It Fails, the Failure Must Be Located
The attraction of Sample is its potential to carry an established voice through a group of new songs, so an album does not have to rediscover its lead on every track. Scalable does not mean duplicated. The four prompt rooms test whether one seed can retain character while each work earns its own scene, melody and emotional fate.
A usable album engine needs three gates: the complete structure of each song, vocal continuity between songs, and the order of climaxes and endings across the record. Failure at any layer may return a result to single, experiment or archive status. Expansion then does not mistake output for a cast, and Lingwan is not drowned by her own name.
12 · KEEP THE VERDICT OPEN
先公开问题与方法，等耳朵决定哪一版真正留下
这篇文章公开的是一次正在进行的创作实验，而不是八首新发行的宣传稿。四组题目、八个候选、一个一分钟来源窗口和三条评估问题已经明确；最终保留哪一版、是否进入专辑、是否需要重做，仍然开放。
我愿意让这个未完成状态被看见，因为它恰好说明 AI 音乐里人的工作在哪里。机器可以快速给出八种可能，Sample 可以让声音记忆更具体，档案可以保存关系；只有完整聆听、比较、舍弃和承担发布后果，才会把其中某一个可能交到听者手里。名字是一项承诺，作品才是兑现。
Publish the Question and Method First; Let Complete Listening Decide What Remains
This essay makes an experiment in progress public; it is not a press release for eight new songs. Four prompts, eight candidates, one one-minute source window and three evaluation questions are now explicit. Which take remains, whether any joins an album, and whether the work must be remade are still open.
That unfinished state is worth showing because it locates human work inside AI music. A system can produce eight possibilities quickly, Sample can make vocal memory concrete, and an archive can preserve the relationships. Only complete listening, comparison, rejection and responsibility for release place one possibility in a listener's hands. The name is the promise. The work is its fulfilment.
本文依据“灵婉 Sample 锁定绣花嗓”实验记录整理：来源为已公开完整作品《鹊桥小调·灵婉戏曲》，选取窗口 00:00–01:00；四组题目各生成两个候选，共 8 个结果。成文时八个候选均未发布，也未完成喜欢／不喜欢标记或最终听觉判定，因此本文不宣称 Sample 已经成功锁定声线，不公开候选链接，也不把它们计入本站公开作品。关于绣花嗓、发飘、过火、源歌阴影与专辑一致性的表述属于 William Lee 的创作判断和待验证标准，不冒充声学测量、真人演唱或平台通用定义。下方 12:22 导听只包含四首已经核实的灵婉完整母带。