中文精读
AI音乐的机会：把“能生成一首歌”变成一次有目的的表达
42章经单期精读 · AI音乐与内容消费
嘉宾：李岩，当期为快音创始人，节目介绍其曾于2015—2018年负责字节整体增长策略｜主持：曲凯
发布：2024年4月6日｜时长：30分04秒｜整理：2026年10月18日
官方单集与节目介绍
当音乐制作变容易，谁会多做一首歌
主持人曲凯用Suno为节目生成片头，近百首候选里挑中一首。这个开场已经同时展示了AI音乐的吸引力和成本：普通人能制作，得到满意结果仍需要选择。
本期不是简单预测歌手会不会被替代。李岩把生产、修改、分发和表达放到同一条链上：模型能生成音乐之后，人为什么需要它，怎样得到适合场景的版本，又如何被听见？节目介绍称快音曾有300万以上日活，这是历史介绍口径，未独立核验。以下技术与市场判断也均属于2024年访谈。
1. 内容单位随场景变化，“一首完整歌曲”不是唯一产品
线索：01:50–07:26。 访谈先谈传统平台的版权支出与议价，再谈短视频如何改变分发。李岩认为，注意力与宣传资源更分散后，更多创作者可以找到受众；主持人则观察到，用户常记住十几秒背景音乐，却不知道歌手是谁。
音乐因此可以是专辑中的歌、独立单曲，也可以是视频里的一小段。节目用唱片容量解释历史长度习惯，这是双方的概括，本篇不据此作完整音乐史考证。李岩称团队测试中，一分至一分半比较合理，并补充独立听歌与视频背景的长度要求不同。
编辑推论：不要先决定做“无限曲库”，再寻找使用者。先问用户是在认真听、给视频配乐、传递祝福还是学习记忆。同样的十几秒，在一个场景足够，在另一个场景可能根本没展开。团队经验也不足以成为所有歌曲的时长规范。
版权讨论还提醒，制作更便宜与权利清楚是两件事。节目转写中的平台亏损、收入比例等有数量级或术语疑点，本篇不引用来计算平台利润，也不将相关讨论当法律判断。
2. 一次生成很方便，局部可改才可能进入制作流程
线索：08:13–10:11、15:05–15:49。 访谈把端到端生成与分层编辑作比较。端到端指从输入直接得到整体音乐，适合快速得到成品；如果第二句不好，重新生成可能连满意部分也改变。音乐人却可能只想改某一小节。
另一种路线保留更多结构和可编辑部分，适合接入制作工作流。节目借图层解释这类差异，但转写对MIDI的定义不够可靠，不能据此把它理解成普通音频波形或完整分轨工程文件。
李岩称团队两条路线都做，因为既想产生整曲，也希望成为可编辑的助手。编辑推论：应按用户修改需求选路线。临时生日祝福可能允许整首重来，已有旋律与明确交付要求的制作任务更在意保留满意部分。专业与普通用户的差别，不只是按钮多少，而是他们知道自己具体要改哪里。
对Suno训练方案的讨论带有嘉宾推测，本篇不写成已核实的内部架构，也不把端到端与可编辑视为永久互斥。
3. 生成便宜之后，筛选与修改成为可用作品的成本
线索：13:07–16:18。 李岩设想，同一素材可以尝试很多音色，再选较好版本，做吉他、钢琴等不同风格，并配合不同媒介形成传播组合。规模化生产扩大了实验空间，但大量版本不会自己变成需求。
主持人追问制作成本，李岩把人力放在重要位置：有人要听、筛、改。团队当时有模型判断一致性或基础质量，但他认为“好听”的标准很多，自动评分主要保底，仍需人的审美选择。片头近百首里挑一首，正是类似机制。
编辑推论：核算单位应是一首可交付作品，而非一次生成。若生成几十版才选出一版，试听、修改与确认可能比算力更影响交付速度；多版本的权利和适用范围也需要逐项确认。节目没有提供完整成本账或经验证的毛利。
李岩还认为，有大量生产能力以后，可以通过分发和内容运营获利，不必只卖版权许可。这是商业模式方向，不能据此认定平台已建立稳定收益。
4. 新需求可能从营销和祝福开始，听歌习惯则需要时间
线索：16:40–23:31、28:25–29:24。 节目举出音乐电影解说、品牌宣传、生日祝福、诗歌演绎和儿童记忆等场景。它们共同的特征是有一个表达任务：同样的话，用旋律、节奏和音色传出去，可能更抓注意或更容易参与。
李岩认为短期营销目的明确，较容易理解为什么使用；个人日常表达的改变则需要适应。主持人将音乐想象为新的表达形式，李岩用手机摄像头作比喻：门槛降低之后，即使不以摄影为业，也可能拍更多照片。
比喻提出的是需求假设。音乐能否帮助记忆、提升传播、转化为习惯，都需要在具体任务中测量，不能把演示好听当成效果证明。节目中关于名人音色和定制歌曲的设想，也未证明相应授权已经存在。
企业场景还有不同交付。李岩举夜店定制音乐的例子：老板未必自己操作模型，可能请音乐人或服务商完成。编辑推论：个人产品重视能否轻松完成表达，企业服务还要明确交付用途、修改范围与验收。个人试玩热度不能代替商户持续订单。
5. 无限供给不会自动消除创作者、社区和审美差异
线索：10:28–13:36、23:31–28:11。 即使未来设备能在本地生成音乐，李岩仍认为用户需要创意、音色、灵感和精选。他追问：懂音乐的人能否做得明显更好？若可以，消费者可能更愿意听别人做好的作品，而不是人人从零创作。节目用80分与90分说明选择差异，这是思想实验，并非测量结果。
主持人觉得Suno V3已经很好听；李岩提醒，听众可能因AI超出预期而给予优待，最终仍要与平常消费的作品比较。这一分歧比宣布“已经够好”更有用：新奇感与持续消费采用的质量标准可能不同。
李岩还发现，团队清洗训练资料时把“喊麦”排除了，导致模型难以满足这种风格需求。他要求重新处理。这暴露的未必只是模型能力不足，还可能是供给在进入训练前就被团队审美筛掉。
编辑推论：评估集应包括目标用户真的喜欢的风格，而非只包含团队认为高级的音乐。社区价值也要落到消费者收益：更好的成品、更容易获得灵感或更省筛选时间。供给数量大、生成人人可用，都不能单独证明社区成立。
李岩关于头部依靠IP、中部承压、尾部获得更多任务，以及传统IP份额被压缩的说法，保留为2024年的产业预测。廉价新供给与旧作品可能并存，节目没有验证最终就业与市场份额变化。
用一条表达任务测试，而不是只比“像不像真人唱”
以下是编辑建议。选定一种用途，例如为活动制作三十秒宣传歌，邀请实际需要它的人写下受众、必传信息和发布位置。先生成，再让同一个人筛选和修改，记录：
| 要回答的问题 | 保留什么证据 |
|---|---|
| 是否把信息讲清楚 | 不看提示的听众能否说出核心信息 |
| 能否改掉一处问题 | 修改前后，满意片段是否保持可用 |
| 成品到底花多久 | 生成、试听、筛选、修改和确认分别用时 |
| 是否真会使用 | 用户把哪一版放进实际传播流程，为什么 |
这个测试能帮助定位表达与交付障碍，不能直接证明营销收益。若目标是生日祝福，则改看收件人理解、分享意愿与实际发送；若目标是持续听歌，则需要更长观察，不能沿用宣传任务的标准。
回听导航与关联阅读
- 04:43：短视频怎样改变音乐的分发与片段消费。
- 08:38：整体生成与局部编辑的取舍。
- 13:39：算力之外，人怎样筛出好听的作品。
- 16:42：营销、祝福与叙事的新用途。
- 24:43：创作者质量差与社区价值。
- 26:12：音乐表达，以及喊麦被数据筛选排除的问题。
相关主题：创作工作流、内容分发、用户审美。可对照FlowGPT的创作者与消费者，再读Hidecloud谈生成过程的消费价值，比较“生成能力”和“愿意消费”的距离。
资料边界：依据完整提供的自动转写、旧精读与官方介绍重新组织，未逐句复听。音乐与歌词不作为访谈观点引用；未核财务数字、版权状态、模型训练细节及性能。预测与审美判断保留归因，不作当前已验证事实。
English edition
The AI Music Opportunity: Turn “Generate a Song” Into Purposeful Expression
42 Zhangjing episode analysis · AI music and content consumption
Guest: 李岩, introduced as founder of 快音, with responsibility for ByteDance’s overall growth strategy during 2015–2018 in the program introduction | Host: Qu Kai
Published: April 6, 2024 | Duration: 30 minutes 4 seconds | Prepared: October 18, 2026
Official episode and introduction
When making music becomes easier, who needs another song?
Host Qu Kai used Suno to generate an opening theme, selecting one from almost a hundred candidates. This already demonstrates both the attraction and a cost of AI music: an ordinary person can produce it, while a satisfying result still requires selection.
The episode goes beyond predicting whether singers will be replaced. Li connects production, revision, distribution and expression. Once a model can generate music, why does someone need it, how do they obtain a suitable version and how does it get heard? The introduction reports that 快音 once exceeded three million daily active users. This historical program claim was not independently verified. The technical assessments and market judgments below also belong to the 2024 interview.
1. Content units change with context; a complete song is one possible product
Episode reference: 01:50–07:26. The discussion begins with licensing expenses and bargaining power in conventional platforms, then moves to short video’s effect on distribution. Li argues that fragmented attention and promotional resources let more creators find audiences. The host observes that people often remember a few seconds of background music without knowing its singer.
Music can consequently be a song on an album, a standalone single or a short passage within a video. The participants explain historical duration conventions through recording capacity; this article does not treat their account as a complete verified history. Li reports that one to one and a half minutes worked reasonably well in team tests, while distinguishing standalone listening from background music in video.
Editorial interpretation: Identify users before committing to an “infinite music library.” Are they listening attentively, scoring a video, sending a greeting or learning something? The same few seconds can be sufficient in one context and too brief in another. A team’s experience cannot establish a duration standard for all songs.
The licensing discussion also separates cheaper production from clear rights. Transcribed platform losses and revenue proportions contain possible errors in scale or terminology. They are not used here to calculate platform profitability, and the discussion is not treated as a legal determination.
2. One-shot generation is convenient; local revision can make it a production tool
Episode reference: 08:13–10:11 and 15:05–15:49. The interview compares end-to-end generation with more structured editing. End-to-end means receiving a whole piece of music directly from an input, useful for obtaining a quick result. If the second line is unsatisfactory, regenerating may also change sections the user already likes. A musician may instead want to change one specific bar.
Another route retains more structure and editable components, helping it fit a production workflow. The participants use image layers as an analogy, but the transcript’s MIDI definition is unreliable. It should not establish that MIDI is an ordinary audio waveform or a complete multitrack project file.
Li says his team pursues both routes, wanting to generate whole songs and function as an editable assistant. Editorial interpretation: Choose according to revision needs. A temporary birthday greeting may tolerate replacing the whole piece. A production task with an established melody and explicit delivery requirements may need satisfactory parts preserved. Professionals and casual users differ in how specifically they know what they want to change, beyond how many buttons they need.
The discussion of Suno’s training approach includes the guest’s conjecture. It is not presented as verified internal architecture, nor are end-to-end generation and editability treated as permanently incompatible.
3. After generation gets cheap, selection and revision still cost time
Episode reference: 13:07–16:18. Li imagines trying many voices for the same material, selecting stronger versions, making guitar or piano variations and combining them with different media for distribution. Large-scale production expands experimentation, but more versions do not create demand automatically.
Asked about production costs, Li emphasizes human labor: someone listens, selects and changes things. His team then used models to assess consistency or basic quality, but he argues that pleasant music has many standards. Automated scoring can maintain a floor while human taste chooses. Selecting one opening from nearly a hundred candidates illustrates the same mechanism.
Editorial interpretation: The accounting unit should be a deliverable work rather than one generation. If dozens of candidates produce one usable result, listening, revision and approval may matter more to delivery speed than compute. Rights and permitted uses also need checking across versions. The interview provides no complete cost ledger or validated margin.
Li also suggests that substantial production capacity could earn through distribution and content operations rather than licensing alone. This describes a business direction, not proof of stable platform earnings.
4. New demand may begin with marketing and greetings; habits take longer
Episode reference: 16:40–23:31 and 28:25–29:24. Examples include musical film commentary, brand promotion, birthday greetings, poetry performance and children’s memorization. Each has an expressive task: delivering words through melody, rhythm and voice might attract attention or encourage participation.
Li considers marketing relatively understandable in the short term because its purpose is explicit. Changes to personal daily expression require adaptation. The host imagines music as a new expressive form; Li compares it with phone cameras. Lower barriers can encourage more pictures even among people who are not photographers.
This analogy proposes a demand hypothesis. Whether music improves memory, distribution or habit requires measurement in a concrete task. An appealing demonstration is not evidence of the intended outcome. Speculation about celebrity voices and customized songs also does not establish that relevant permissions exist.
Business delivery differs as well. Li gives the example of a nightclub commissioning its own music. The owner may hire musicians or a service provider rather than operate a model. Editorial interpretation: A consumer product needs easy completion of an expressive task. Business services must also establish intended use, revision scope and acceptance. Consumer experimentation cannot substitute for repeat commercial orders.
5. Infinite supply does not eliminate creators, communities or differences in taste
Episode reference: 10:28–13:36 and 23:31–28:11. Even if devices eventually generate music locally, Li expects people to need ideas, voices, inspiration and curation. He asks whether someone with musical knowledge can make substantially better work. If so, consumers may prefer that work to creating from scratch. An eighty-versus-ninety-point comparison illustrates the idea; it is not a measured quality difference.
The host finds Suno V3 already enjoyable. Li asks whether AI is receiving favorable treatment for exceeding low expectations and argues that it ultimately must be compared with normally consumed music. This disagreement is useful: novelty and continued consumption may apply different standards.
Li also discovered that his team had removed 喊麦, a rhythmic shouted-vocal style, while cleaning training material. The model consequently struggled to serve that preference, and he asked the team to address it. Missing capability may therefore originate in taste-based filtering before training, as well as in model limitations.
Editorial interpretation: Evaluation material should include styles target users actually enjoy, beyond the team’s preferred notion of sophisticated music. Community value must produce a consumer benefit: better finished works, easier inspiration or less selection time. Large supply and universal access to a generate button cannot establish that value by themselves.
Li’s predictions about leading musicians retaining IP advantages, pressure on mid-tier workers, more tasks for others and a falling share for conventional IP remain forecasts from 2024. Cheap new supply and existing works could coexist. The episode does not verify eventual employment or market-share changes.
Test an expressive task instead of only asking whether it sounds human
The following is an editorial exercise. Choose a use, such as a thirty-second promotional song for an event. Ask someone who needs it to specify the audience, essential message and publication setting. Generate music, then have that person select and revise it, recording:
| Question | Evidence to retain |
|---|---|
| Is the message clear? | Can listeners identify the central information without reading the prompt? |
| Can one problem be fixed? | After revision, do satisfactory passages remain usable? |
| How long does a deliverable take? | Separate generation, listening, selection, revision and approval time |
| Will it actually be used? | Which version enters a real distribution process, and why? |
This test can locate expression and delivery obstacles; it does not directly establish marketing returns. For a birthday greeting, examine recipient understanding, willingness to share and actual sending instead. For sustained music listening, use a longer observation period. A promotional task’s standard cannot simply be reused.
Listening guide and related reading
- 04:43: Short video, music distribution and fragment consumption.
- 08:38: Whole-piece generation versus local editing.
- 13:39: Human selection beyond compute cost.
- 16:42: Marketing, greetings and new narrative uses.
- 24:43: Creator quality differences and community value.
- 26:12: Musical expression and the exclusion of 喊麦 during data selection.
Related themes: creative workflows, distribution and user taste. Compare FlowGPT’s creators and consumers, then Hidecloud on consuming the generation process, to explore the distance between generation capability and willingness to consume.
Source boundary: Reorganized from the complete supplied automatic transcript, earlier reading note and official introduction; audio was not replayed sentence by sentence. Music and lyrics are not quoted as interview views. Financial figures, rights status, training details and performance were not verified. Forecasts and aesthetic judgments retain attribution rather than becoming verified current facts.