Audible, AI Creators Deploy Synthetic Voice Tools to Scale Audio

Audible与AI创作者借助合成语音工具扩大音频规模 内容创作者限制全合成AI输出以维持人机协同的真实性，同时利用定制模型与声音克隆来扩大生产规模。 由AI内容创作者与行业领袖组成的小组探讨了创作者如何利用生成式AI工具，在兼顾制作速度与产出规模的同时保持真实的个人风格。发言者分享了关于管理工作流自动化、缓解LLM幻觉，以及利用基于个人档案训练的定制GPT建立安全护栏等实用策略。Audible平台代表重点介绍了以创作者为中心的产品开发（例如ACX上的AI声音克隆），旨在保障创作者权益的同时提高有声书的供给。讨论最终强调，在AI驱动的媒体格局下，人类的真实生活经历、语境理解和独特的创作个性依然是不可替代的核心资产。 创作者通常将重复性行政工作委托给 AI，同时由人工严格主导核心构思与个人风格，以在生成速度与真实性之间取得平衡。 (0:38) 为保护受众信任与个人品牌，Aishwarya Srinivasan 将完全由 AI 克隆的声音与数字人视频输出严格限制在总内容的 10% 以内。 (0:38) 基于过往海量内容训练的定制助手（如 AshGPT）仍需持续调优提示词与人工审核，其初稿准确率往往受限于约 80% 的瓶颈。 (1:22) AI 模型欠缺对现实生活经历与即时互动的感知，而创作者独特的思维跳跃与个性化视角是抵御同质化合成内容的关键壁垒。 (19:56) Audible 依托 ACX 创作者反馈小组等机制推进语音复刻等功能的分阶段公测，在降低有声书制作门槛的同时确保透明标注与创作者权益。 (32:22)

Creators and platform leaders argue that while generative AI rapidly accelerates content pipelines, real-time lived experience and distinct creative identity remain indispensable human safeguards. To maintain audience trust, creators must treat AI as an augmenting assistant capped by accuracy limits rather than an autonomous replacement. The Eighty Percent Ceiling and Contextual Gaps (1:22) AI generation tools consistently struggle to produce finished creative work, capping out at an estimated 80% accuracy threshold that demands ongoing human editing. Aishwarya Srinivasan describes training "AshGPT"—a custom model fed on over 4,500 of her LinkedIn posts, Substack articles, and podcast transcripts—yet finds its drafts require manual refinement to overcome initial drafting friction. Similarly, Ravit Jain notes that prompting an LLM about his five-year-old show yielded entirely fabricated interview subjects and discussions, demonstrating that extensive personal training data cannot eliminate hallucinations. Srinivasan explains that this structural ceiling exists because models lack real-time bandwidth to human cognition and daily life. While static prompts reflect historical writing, they cannot ingest same-day lived context—such as morning conference sessions with researchers like Fei-Fei Li or Jeff Dean. Because generative systems cannot spontaneously process real-world interactions as they occur, creators must actively supply nuanced context and oversight to keep material accurate and relevant. Authenticity and Distinct Biological Hallucinations The panel draws a sharp line between faceless corporate brands, which can deploy fully synthetic media, and individual creators whose value relies on personal connection. Jain emphasizes that audiences readily identify fully synthetic outputs, warning that unedited AI posts erode authenticity and hinder audience growth. Consequently, Srinivasan limits fully cloned synthetic video (using HeyGen and ElevenLabs) to 10% of her total social media output, while Dylan Curious relies on automation like Gling solely to strip filler pauses while keeping ideation human-driven. Curious characterizes human quirks, unpredictable tangents, and eccentric perspectives as "natural biological hallucinations" that serve as an artist's primary differentiator. While generative tools like Google Veo can yield unexpected creative variations from a 30- to 40-second prompt, attempting to eliminate human idiosyncrasies in favor of polished AI uniformity produces homogenized, forgettable content. Preserving imperfect, unrepeatable human voices remains the central defense against blending into automated online noise. Tailoring Creator Tooling Beyond Influencer Stereotypes Effective creator tools must support specialized niches rather than imposing rigid, generic templates. Srinivasan, representing technical communicators within Meta's Creator Group, points out that educational data science creators require conversational, straightforward tooling rather than the hyper-stylized edits, fast cuts, or colorful frames built for traditional lifestyle influencers. While she employs ManyChat for automated direct messaging on Instagram, she preserves direct manual control over her LinkedIn publications and community Slack interactions to protect her core value proposition. Software platforms are structuring product development around iterative creator feedback. Sadaf Kazmi explains that Audible relies on continuous social listening, targeted feature surveys, direct qualitative interviews, and a quarterly ACX Creator Feedback Group composed of independent authors and narrators. This advisory panel reviews roadmaps and tests early features through phased beta rollouts, ensuring new capabilities match practical working requirements before broader deployment. Commercial Voice Scaling and Algorithmic Exploitation (32:22) In commercial audio publishing, synthetic tooling provides an avenue to bypass prohibitive production costs where roughly 97% of books lack audio editions. Audible's ACX Voice Replicas beta allows narrators to license verified AI clones of their own voices, enabling them to scale production and expand into underserved non-English markets while maintaining audition control, output editing rights, and compensation. Kazmi notes that this setup helped one narrator continue working and earning during a week when she lost her voice, with Audible mandating explicit labeling (such as "Sadaf Kazmi's voice replica") to preserve audience transparency. However, expanding synthetic production creates unintended algorithmic consequences. Curious notes the "liar's dividend," in which minor synthetic anomalies—such as unnatural cadences or misaligned eye movements—cause viewers to pause and scrutinize video clips. Social media recommendation engines misinterpret this prolonged watch time as high engagement, algorithmically boosting unverified synthetic trends like a viral, entirely fabricated video of rabbits on a trampoline. Panelists stress that robust attribution, transparent disclosures, and platform detection mechanisms are vital to prevent unauthorized identity exploitation. While AI tooling solves persistent bottlenecks in administrative tasks, post-production editing, and audio catalog scaling, creators must retain manual control over their core ideation and community interactions. A key unresolved question is how major distribution platforms will update recommendation algorithms to distinguish between authentic human engagement and the synthetic anomalies that currently capture outsized algorithmic attention.

行业从业者与内容创作者指出，生成式人工智能虽能显著压缩内容生产周期并降低制作门槛，但受限于实时生活经验的脱节与模型幻觉，初稿质量普遍受困于约80%的准确率上限；创作者必须保持人在回路（human-in-the-loop）的把控，并通过规范的内容标注与版权追踪机制抵御虚假流量和身份盗用风险。 速度与真实性的权衡及八成准确率瓶颈 (0:38) 社交平台算法对更新频率的严苛要求促使创作者寻求自动化，但 Aishwarya Srinivasan 与 Ravit Jain 强调，单纯追求速度极易导致内容陷入同质化与浅薄化。Srinivasan 使用基于自身4500多篇历史文章与播客脚本训练的“AshGPT”起草内容，但指出该输出始终停留在约80%的准确率水平，根本原因在于模型无法即时吸收创作者当天的真实经历与即席思考（例如当日参加的行业前沿演讲）。Jain 的测试同样显示，即使用数千条历史动态训练大模型，AI依然可能彻底编造虚假的访谈经历。 两位创作者均认为，人工审核与微调是不可让渡的底线。为维护受众信任，Srinivasan 将由 HeyGen 和 ElevenLabs 完全生成的克隆视频在 Instagram 与 YouTube 的发布比例严格限制在10%以内，其余内容均保持人工主导。Jain 警告称，受众对纯合成内容具有极高的鉴别力，过度依赖未经编辑的AI输出不仅会透支真实感，还会阻碍频道的长线增长。 创作者分层工作流与定制化工具设计 在生产工具的具体应用中，成熟创作者普遍采用“部分委托、核心留存”的分层架构。Srinivasan 组合使用 HeyGen、ElevenLabs、Captions、Canva 和 CapCut 进行视觉与音频处理，并借助 ManyChat 自动化处理 Instagram 平台的消息互动；但对于技术沉淀核心阵地 LinkedIn 以及 Slack 社群，她坚持全程手动撰写与亲自回复。她呼吁产品研发人员破除对内容创作者的刻板印象，技术与数据科学领域的传播者并不需要生活方式类博主常用的花哨滤镜或剪辑模板，工具更应提供灵活且去模板化的控制力。 Audible 平台的 Sadaf Kazmi 则从平台产品设计角度补充了服务逻辑。Audible 贯彻以创作者为中心的产品理念，通过建立涵盖顶级独立作者与播音员的季度顾问小组（ACX Creator Feedback Group）、定期定向问卷、一对一深度访谈以及小范围渐进式公测，确保 Virtual Voice Studio 等AI工具能够真正减轻创作者的操作负担，而非剥夺其对旁白节奏的精细艺术控制权。 声音克隆对有声出版门槛的重构 (32:22) 针对图书出版行业中约97%的作品因高昂录音成本而缺乏有声版本的现状，Audible 通过旗下 ACX 平台探索声音复制技术以补齐产能缺口。Kazmi 介绍，已试运行约半年的“声音副本（Voice Replicas）”内测项目允许播音员录制并生成本人的专属AI语音模型，从而将业务拓展至预算有限的非虚构作品及小语种市场。在实际案例中，曾有专业播音员在嗓音完全失声的一周内，借助已验证的声音副本无缝承接项目并保障了个人收入。 该模式的前提是保障创作者的完整权益与知情同意权。播音员享有选择接单、试听调校最终音频以及获取公平版税的全部控制权。同时，平台对所有采用合成声音的书籍实施强制性显式标注（例如标明“某某的声音副本”），以确保消费端的知情权，并建立起防范声音资产被未授权滥用的制度屏障。 算法红利诱因与数字身份伦理防线 (38:10) 生成式内容在传播分发环节引发了新型伦理与算法失真问题。主持人 Dylan Curious 指出，当前视频平台存在典型的“说谎者红利（Liar's Dividend）”现象：当AI生成的视听内容出现轻微瑕疵（如视线偏移或语调卡顿）时，观众往往会延长驻留时间仔细辨别破绽，算法系统却将这种额外的观看时长误判为高互动意向，反向助推了虚假合成视频（例如全网扩散但完全虚构的蹦床兔子视频）的病毒式传播。 此外，在视频直播等实时场景中，直接朗读AI即时生成的复杂学术回答屡次因事实性幻觉遭到观众反驳，迫使创作者必须将AI论述降级为参考假说。Jain 与 Kazmi 呼吁，除创作者自身的审慎态度外，主流平台亟需建立强有力的AI合成内容识别过滤机制与版权归属溯源体系，遏制未授权商业仿冒对创作者个人品牌的伤害。 尽管AI技术有效降低了音频与视觉内容的制作门槛，但创作者独特的真实生活阅历与非标准化的个人风格仍是抵御内容均质化的核心壁垒；行业如何在借助算法扩充产能的同时，健全跨平台的版权收益分配与合成内容监管机制，仍是决定该生态可持续性的关键课题。

## Background 本内容由 Ai4 于 2025 年 8 月 27 日发布在 YouTube 平台上。页面简介显示，本场圆桌研讨会汇聚了多位行业从业者与全网创作者，共同探讨 AI 驱动的工具如何重塑创作者的生产流程、提升效率并拓展创意边界，同时重点探讨如何在人机协同中坚守真实感与艺术控制权。 ## TL;DR 圆桌围绕创作者在 AI 时代的“生产速度与个人辨识度”博弈展开深度剖析。嘉宾们一致指出，AI 是增强而非取代创作者的工具，创作者必须牢牢掌握选题、立意等漏斗顶端核心，坚守“人类在环（Human-in-the-loop）”的底线，以突破 AI 初稿“永远只完成 80%”的瓶颈。同时，讨论还涵盖了利用个人知识库定制模型、Audible 填补 97% 有声书空白的声音克隆实践，以及 AI 标注透明度与“说谎者的红利”等版权与治理挑战。 ## Atmosphere & style 整场圆桌对谈氛围务实、坦诚且富有行业前瞻性，没有空洞的营销说辞，而是充斥着一线创作者与产品负责人的实操反思与真实痛点。主持人 Dylan 与嘉宾之间的接抛梗自然流畅，从下班后的“思维混乱”到对抗 AI 均质化的“生物性自然幻觉”，对话在严谨的技术探讨与幽默风趣的经验分享之间自如切换。听众能真切感受到创作者们在拥抱效率红利的同时，对坚守个人独特声线与受众信任所抱有的清醒与审慎。 ## Key takeaways - [03:56](https://www.youtube.com/watch?v=2kzMx4egyDY&t=236s) AI 时代创作者面临速度与辨识度的博弈，过度依赖自动化易导致内容风格扁平化并失去受众信任，守住独特的“数字指纹”是核心资产。 - [04:29](https://www.youtube.com/watch?v=2kzMx4egyDY&t=269s) 创作者应牢牢把控“创作漏斗顶端”，选题框架与核心观点完全由本人主导，仅将 AI 工具局限在执行层（如仅占 10% 的数字人视频），并坚持原创脚本。 - [07:58](https://www.youtube.com/watch?v=2kzMx4egyDY&t=478s) AI 能够极大释放机械重复劳动的生产力，但生成文案前需通过大量个人历史语料定制训练模型，以确保输出内容保持本人的独有语气与真实感。 - [09:49](https://www.youtube.com/watch?v=2kzMx4egyDY&t=589s) 平台方将 AI 定位为“增强”而非“替代”创作者，例如 Audible 通过虚拟声音工作室（Virtual Voice Studio）赋予创作者对 AI 旁白的全流程精细化编辑权。 - [11:51](https://www.youtube.com/watch?v=2kzMx4egyDY&t=711s) 创作管线存在人机分工边界：前期的选题灵感与主观观点无法被机器替代，而后期剪辑与降噪（如 Gling）能实现显著的自动化提效。 - [13:22](https://www.youtube.com/watch?v=2kzMx4egyDY&t=802s) 无论模型如何定制迭代，AI 交付的初稿在创作者心理预期中“永远只完成了 80%”，仍需人工介入（Human-in-the-loop）进行深度校对与润色重写。 - [15:51](https://www.youtube.com/watch?v=2kzMx4egyDY&t=951s) 模型幻觉风险依然存在，特别是在即兴发散与背景事实生成中极易编造内容，创作者发布前必须保持审慎并逐一核实。 - [19:40](https://www.youtube.com/watch?v=2kzMx4egyDY&t=1180s) 人类瞬息万变的现实经历与鲜活交流无法被实时灌输给 AI，创作者自带的“生物性自然幻觉”与独特思维切角构成了不可替代的个人核心壁垒。 - [23:58](https://www.youtube.com/watch?v=2kzMx4egyDY&t=1438s) AI 工具开发者应摒弃刻板的泛娱乐预设，针对不同垂直领域（如严肃技术科普）提供灵活模块化能力，并建立“创作者至上”的敏捷反馈闭环。 - [29:17](https://www.youtube.com/watch?v=2kzMx4egyDY&t=1757s) Audible 推出声音克隆（Voice Replicas）Beta 测试，让朗读者在保留完全使用控制权的前提下承接更多有声书录制，保障了工期与版权报酬的连续性。 - [31:05](https://www.youtube.com/watch?v=2kzMx4egyDY&t=1865s) ChatGPT 可作为多模态数据分析助手，通过输入频道后台数据、爆款封面图及社群反馈，横向对比竞品并辅助优化运营策略。 - [32:58](https://www.youtube.com/watch?v=2kzMx4egyDY&t=1978s) 市面上约 97% 的在售图书缺乏有声版本，AI 旁白与声音克隆技术能够大幅降低制作门槛，有效填补独立出版与非英语国际市场的音频缺口。 - [36:10](https://www.youtube.com/watch?v=2kzMx4egyDY&t=2170s) “说谎者的红利”（Liar's Dividend）使得逼真但微带破绽的 AI 内容因用户驻足而更易获得算法流量倾斜，因此严格的透明标注与版权溯源机制至关重要。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=2kzMx4egyDY&t=4s) 圆桌开场与嘉宾背景介绍 主持人 Dylan Curious 轻松开场，定下了本场圆桌随性且富有干货的基调。他鼓励台下观众在尾声环节积极提问，并邀请台上的嘉宾依次介绍自己的职业经历与背景，以便观众清楚该向谁寻求哪方面的行业洞察。 率先发言的是 Fireworks AI 的开发者关系负责人 Aishwarya Srinivasan（大家也叫她 Ash）。她笑称自己其实有两份工作：朝九晚五的本职工作是在 Fireworks AI——一家专注于生成式 AI 推理引擎的初创公司——带领开发者生态；而“晚五朝九”的业余时间，她则是一位全网内容创作者。她自嘲自己最初是 LinkedIn 上的“意外创作者”，在深耕该平台 10 年后意外积累了庞大粉丝群，如今阵地已拓展至 Substack、YouTube 和 Instagram。前一天她刚在 Ai4 大会上主讲了可扩展 AI Agent 的技术议题，今天则换上创作者的身份参与交流。 随后发言的 Ravit Jain 是《The Ravit Show》的创始人兼主持人，专注于数据与 AI 领域领袖专访已长达五年半。他创办节目的初衷是抵制市面上铺天盖地的销售套路（Salesy Stuff），为企业高管和从业者提供真正有价值、能够助力规模化发展的深度对谈。目前该节目在 LinkedIn、YouTube 和 Instagram 等社交渠道上已建立起覆盖近 100 万名专业人士的社群。与身兼两职的 Aishwarya 不同，Ravit 将其作为唯一的全职事业倾注心力。 紧接着，Audible 创作者与国际产品负责人 Sadaf Kazmi 介绍了自己在有声书领域的专注方向。她的团队致力于为创作者与平台的交互打造无缝体验，旗下核心业务之一是 ACX（Audiobook Creation Exchange，有声书创作交易平台），支持创作者将独立内容发布至 Audible 和 Amazon。Sadaf 补充道，作为高品质音频叙事的领军者，Audible 目前拥有超过 100 万部有声书作品，服务遍及全球 180 多个国家、支持 50 多种语言。 - **Aishwarya Srinivasan (Ash)**：Fireworks AI 开发者关系负责人；Substack、YouTube、LinkedIn 技术内容创作者。 - **Ravit Jain**：《The Ravit Show》创始人兼主持人，深耕数据/AI 访谈 5.5 年，社群规模近 100 万人。 - **Sadaf Kazmi**：Audible 创作者与国际产品负责人，主管 ACX 平台，链接全球 180+ 国家创作者。 - **Dylan Curious**：AI 题材 YouTuber，聚焦 AI 发展对人类社会及人机交互形态的冲击。 --- ### [03:56](https://www.youtube.com/watch?v=2kzMx4egyDY&t=236s) 速度与辨识度的博弈：如何在 AI 辅助下保持独创风格？ 主持人 Dylan Curious 切入圆桌的核心议题：速度与个人辨识度（Speed vs. Voice）之间的权衡。AI 工具确实能让制作速度倍增，但负面效应同样突出：有时生成的劣质内容需要反复返工，整体效率反受拖累；更有甚者，AI 生成的内容风格过于扁平、平庸套路，导致创作者原本的艺术构想与独特声线被抹杀。Dylan 由此向全场抛出疑问：当工具泛滥时，创作者究竟该如何守住属于自己的“数字指纹”（Creative Fingerprint）？ ```mermaid flowchart TD accTitle: 创作者在 AI 时代面临的创作权衡 subgraph Dilemma["创作的核心博弈"] Speed["生成速度与高频更新<br/>（平台算法要求）"] Voice["个人独创性与独特风格<br/>（品牌核心资产）"] end Speed <-->|"存在冲突与妥协"| Voice Dilemma --> Danger["主要风险：风格扁平化、返工成本高、丧失受众信任"] ``` --- ### [04:29](https://www.youtube.com/watch?v=2kzMx4egyDY&t=269s) 漏斗顶端控制与平台差异化策略 针对这一痛点，Aishwarya Srinivasan 坦言创作者的处境极其艰难——不仅要保持真实真诚，还被迫迎合社交平台严苛的高频更新算法，这极易让人陷入对自动化的过度渴望与焦虑之中。她的破局策略是坚决把控“创作漏斗的顶端”：选题框架、核心思想、论述观点以及社交账号的主题调性必须百分之百由本人主导构思；只有进入特定制作环节时，才定向引入针对性工具。 在实际管线中，Aishwarya 会结合 HeyGen 的数字人分身与 ElevenLabs 的声音克隆工具来制作视频。但她设定了严格的红线：这类全 AI 生成的视频在 Instagram 或 YouTube 的产出占比仅保持在 10% 左右，且脚本必须完全出自她手，仅仅用于在日程拥挤时节省实拍时间。她强调，创作者必须分清“观众眼中的真实感”究竟源于你亲自露脸出镜，还是源自你观点本身的深度与共鸣。此外，她还必须在“个人表达欲”与“受众实际需求”之间寻找平衡，并通过平台分流来化解这种冲突——在 LinkedIn 上发布前沿硬核的技术剖析，而在 Instagram 上则将复杂的 AI 概念拆解为微学习（Microlearning）短篇。 当 Dylan 追问未来是否会长期维持“人机协作回环”（Human in the Loop）时，Aishwarya 斩钉截铁地表示：“绝对会，我绝不可能把自己的社交账号完全托管给自动化。”她随后补充说明，自动化并非绝对的好与坏，如果是没有个人具象化形象的企业级品牌或产品品牌，利用 AI 虚拟化身进行全流程自动化完全合情合理且能节省巨额开支；但对于依赖个人声誉与信任纽带的个人 IP，主观意志与真实的个人印记永远不可让渡。 - **工作流切分**：创作者牢牢掌控想法与选题（Top of Funnel），仅在执行层局部使用工具。 - **AI 数字人克隆占比**：由 HeyGen 视频克隆 + ElevenLabs 声音克隆生成的视频仅占其全网内容的 10%，底线是手写原创脚本。 - **平台分发差异**：LinkedIn 承载进阶高深的技术观点，Instagram 主打轻量化的 AI 知识点微学习。 - **个人品牌 vs 企业品牌**：企业级品牌可以通过无面孔或虚拟形象做高度自动化，而个人 IP 必须坚定保留 Human in the Loop。 --- ### [07:58](https://www.youtube.com/watch?v=2kzMx4egyDY&t=478s) 效率工具与个人声线的模型训练 Ravit Jain 接过话题并高度认同 Aishwarya 的判断。他直言：“速度固然重要，但绝不能以牺牲身份特质（Identity）为代价。”在内容生产体系中，人类的主导把关必须时刻在线。他将 AI 在内容中的应用拆解为两层逻辑：一方面是重复机械劳动的彻底剥离，另一方面是基于个人语料对底层模型的定制训练。 针对单调乏味的工作（Monotonous Jobs），AI 能够极大释放生产力。Ravit 描述了一个典型的现场场景：当他在舞台上进行访谈时，AI Agent 可以实时将音视频转录为文本，并根据预设规则触发多渠道分发机制，当场向各大社交平台推送他正在 Ai4 现场演讲的动态。这种速度是纯人工难以企及的。然而，要确保自动分发的内容读起来依然像是他本人在说话，核心在于如何“训练模型”。在利用 ChatGPT 等大语言模型生成文案前，必须反复向模型注入自己的过往表达风格、行业阅历以及惯用语气，唯有当模型透彻理解了创作者的表达习惯与独有声线（Voice），协同生成的文字才具备可信的真实度与情感温度。 ### [09:49](https://www.youtube.com/watch?v=2kzMx4egyDY&t=589s) 真实性与平台视角：AI 是增强而非替代创作者 创作者与粉丝社群建立深厚连接的核心在于其独特性与真实感。许多粉丝早在五年前就因为创作者真实的说话风格和未经修饰的原始视频而选择关注，那时甚至还没有完善的生成式 AI 或像 HeyGen 这样的数字人工具。在当下的创作生态中，约有半数创作者在重度依赖 AI 工具，但观众和社群往往能一眼识别出机械化生产的痕迹，这种缺乏真实性的内容很容易导致粉丝增长停滞，而社群拓展与深度交流正是创作者业务的生命线。 从平台方的角度来看，AI 的应用绝不仅仅是使用类似 ElevenLabs 的语音合成工具来加快有声书制作速度。Audible 将创作者视作核心客户，其产品设计的核心指导原则始终是“AI 的存在是为了增强创作者的能力，而非取代创作者”。例如，Audible 推出的虚拟声音工作室（Virtual Voice Studio）是一款具备 AI 旁白编辑能力的产品，它允许创作者深入到音频制作的每个细节进行精细化调整与二次编辑，确保最终交付的作品能够精准还原其艺术构思，在大幅提升制作效率的同时牢牢把握创意的自主权。 - **社群认同与真实感**：早期粉丝多被创作者原生的人格特质与未修饰的视频风格吸引；过度使用未经润色的 AI 容易破坏受众信任并拖累粉丝增长。 - **Audible 的产品哲学**：将 AI 定位为创作增强工具，通过 Virtual Voice Studio 赋予创作者对 AI 旁白的全流程编辑控制权。 --- ### [11:51](https://www.youtube.com/watch?v=2kzMx4egyDY&t=711s) 创作管线的边界：从选题灵感到剪辑自动化的取舍 将 AI 引入视频制作管线并非能够实现全流程无缝接管，不同环节对 AI 的适应度存在巨大差异。在视频前期的选题与素材搜集阶段，直接依赖 ChatGPT 抓取文章往往收效甚微，因为创作者必须亲自在 Twitter（X）或订阅的邮件通讯（Newsletter）中挖掘那些自己抱有强烈观点、能够产生共鸣的选题切入点，AI 无法凭空替代人类的情感种子与主观判断。 然而在后期制作环节，AI 工具的提效表现极为惊人。诸如 Gling 这类 AI 剪辑软件能够自动识别并剔除口癖、语气停顿（如 "um" 和 "ah"），甚至能智能识别出创作者为了追求完美而反复录制两三遍的句子，并自动保留最后一遍的最佳发挥。此外，自动补充空镜（B-roll）和降噪清理等 AI 功能也在飞速进化。但整个内容创作的基底必须始终植根于创作者自身的思考与火花，无法从零开始全部交由机器代劳。 ```mermaid flowchart TD accTitle: 创作者内容管线中的人机协作分工 subgraph HumanCore ["人类主导环节（真实性与情感核心）"] A["选题灵感与素材筛选<br/>(Twitter / Newsletter)"] B["主观观点与核心立意<br/>(不可替代的个人表达)"] end subgraph AIEnabled ["AI 赋能与辅助环节（效率与工具化）"] C["初稿撰写与结构搭建<br/>(Custom GPT / 提示词辅助)"] D["智能后期剪辑<br/>(Gling 去除口癖/重录识别/B-roll)"] end A --> B B --> C C --> D ``` --- ### [13:22](https://www.youtube.com/watch?v=2kzMx4egyDY&t=802s) 定制“Ash GPT”与提示词迭代：永远停留在 80% 的初稿体验 针对模型幻觉与创作偏差，构建深度的个人定制化知识库是当前行之有效的方案之一。通过在 ChatGPT 上构建专属的“Ash GPT”，可以为其投喂个人历年积累的约 4,500 篇 LinkedIn 帖子、Substack 深度博客文章，以及个人 AI 播客《Weekend Musings》的完整逐字稿。这使得模型能够深度学习创作者的技术深度、类比手法、叙事逻辑以及语言性格特征，从而胜任高水准的初稿撰写工作。 ```mermaid flowchart LR accTitle: Ash GPT 知识库构建与人机协作工作流 subgraph DataSources ["训练数据源"] A["4,500+ LinkedIn 帖子"] B["Substack 博客文章"] C["Weekend Musings 播客逐字稿"] end subgraph CustomModel ["定制化模型配置"] D["Ash GPT 核心引擎"] E["多轮长系统提示词<br/>(持续增删结构与格式)"] end subgraph ReviewStage ["严格的人工审核（Human-in-the-Loop）"] F["第一版初稿输出"] G["人工校对与风格重写"] end A & B & C --> D E --> D D --> F F --> G ``` 维持定制化模型的表现依赖于持续迭代的长系统提示词（System Prompt），提示词的结构和格式需要经历数周甚至数月的反复修改与回炉打磨。即便如此，在创作者晚间下班、遭遇思维混乱（Brain Fog）时，AI 输出的初稿依然无法做到直接发布。创作者绝不能建立一条完全无人值守、让 AI 自动写脚本、配合 HeyGen 生成数字人视频再由 Captions 自动排版发布的“纯自动化流水线”。由于没有任何一次 AI 的初稿能够达到完美无瑕，"Human in the loop"（人工介入）是不可忽视的底线。 - **Ash GPT 语料储备**：涵盖约 4,500 篇 LinkedIn 动态、长篇博客及播客逐字稿，用于锁定个人语调与思维模型。 - **提示词迭代工程**：系统提示词并非一蹴而就，而是历经高频次的内容与格式追加调整。 - **“80% 瓶颈”心理学**：当 AI 能力提升时，人类对优质内容的期望值也在同步拉高，因此创作者主观上总会觉得 AI 的交付成果“永远只完成了 80%”，依然需要耗费半小时到一小时进行人工审校与重写。 --- ### [15:51](https://www.youtube.com/watch?v=2kzMx4egyDY&t=951s) 幻觉陷阱与 Audible 的高标准质量基准测试 模型幻觉依然是内容生产中的重大隐患。在 ChatGPT 刚问世的早期阶段，创作者尝试输入提示词让其介绍自身背景或《The Ravit Show》播客节目时，模型曾凭空编造出创作者从未做过的访谈嘉宾和完全不相干的话题内容。尽管当前模型智能已大幅演进，但在发布前人工核对每一处内容是否符合真实发声仍是刚性需求；即便给模型投喂海量历史数据或耗费上百小时进行针对性训练，AI 依然无法做到绝对精准，创作者在借用其风格生成内容时必须保持高度审慎。 对于像 Audible 这样处理预先撰写完毕的书籍文本的平台而言，质量控制的侧重点则在于语音交付层面的基准测试（Benchmarking）与严格评估。Audible 团队与亚马逊 AGI 部门旗下的内部语音专家紧密合作，建立了一套严苛的内部测试指标，重点监测听众对 AI 朗读效果的实际听觉反馈，以此判断声音表现究竟是仅仅达标还是达到了卓越水准，从而持续优化高保真的听觉体验。 - **虚构访谈幻觉案例**：早期模型在回答公开个人资料时容易出现“无中生有”，编造不存在的播客期数与对话主题。 - **Audible 质检体系**：通过高频质量基准测试与听众体验指标联动，依托亚马逊内部 AGI 语音团队攻坚高质量文本转语音技术。 - **即兴发挥中的幻觉风险**：在视频录制中进行无脚本即兴发散（如探讨量子力学等专业复杂概念）时，依赖未经校验的 AI 生成信息极易引入事实性错误。 ### [19:40](https://www.youtube.com/watch?v=2kzMx4egyDY&t=1180s) 创作中不可替代的“人类内核”与独特表达 主持人 Dylan Curious 承接上一话题提到，创作者在使用 AI 辅助时已经学会对其给出的答案“保持审慎（take it with a grain of salt）”，避免直接照本宣科而被评论区纠错。随后他抛出一个核心议题：在 AI 席卷创作全流程的背景下，**到底什么是人类创作者根本无法割舍、必须亲自保留的内核？** Aishwarya Srinivasan 认为，之所以目前的 AI 工具充其量只能完成 80% 的工作，是因为模型并未与人类大脑建立物理连接。创作者每天接收的信息是高度动态且实时的——包括当天读到的文章、遇到的人、与同行专家（如 Fei-Fei Li 或 Jeff）的最新交谈。这些未被数字化的鲜活体验和瞬息万变的思考，根本无法实时灌输给 AI 模型或检索增强生成（RAG）管道。正是这种受最新生活经历影响的情感与思考，赋予了内容最本真的“人类可信度与真实性（authenticity）”。 ```mermaid flowchart TD accTitle: Human Authenticity vs AI Generation subgraph S1["人类即时输入与体验"] A["日常阅读与思考"] B["真实社交与行业交流"] C["即时灵感与情绪反应"] end subgraph S2["独特性过滤层"] D["生物性自然幻觉与偏见"] E["独家叙事语调与审美控制"] end subgraph S3["输出成果"] F["高辨识度原创内容"] G["与受众建立真实信任纽带"] end A & B & C --> D D --> E E --> F --> G ``` Ravit Jain 简短补充道，人类大脑依然远比当前的 AI 模型聪明与高级。AI 输出结果的优劣，最终取决于人类如何构思并提出 Prompt；想要得到真正真实、有深度、贴近自身思想的内容，创作者就必须把自己脑中的心智共享给 AI。 Dylan Curious 则提出了一个趣味视角——**“人类自带的生物性自然幻觉（natural biological hallucinations）”**。他认为正是每个人奇特的看世界方式、天马行空的思维跳跃和个性化的“怪癖”，才让内容显得有趣而迷人： - 每个人看世界的独特切角不应被 AI 的均质化所抹除。 - 在使用 Google Veo 等视频生成工具尝试生成 30 至 40 秒视频时，AI 的脑洞常常与人类预想大相径庭。这种“人机想象力之间的落差（delta）”往往能反哺创作，拓展创作者的喜剧风格或写作语调。 - 认清并坚守自己独特的“幻觉模式”，是在 AI 时代避免跟风盲从、保持个人声音吸引力的核心关键。 --- ### [23:58](https://www.youtube.com/watch?v=2kzMx4egyDY&t=1438s) 赋能创作者：以创作者为中心的产品开发哲学 讨论转向“以创作者为中心的 AI 产品开发（Creator-centric AI Product Development）”，探讨工具开发者应如何为不同领域的创作者量身定制产品生态。 Aishwarya Srinivasan 分享了她日常使用的工具流水线与创作边界： - **日常工具链**：日常高频使用 HeyGen、ElevenLabs、Captions、Canva 以及 CapCut 进行视频制作与排版。这些工具充当了“趁手的工具箱”，既大幅提升效率，又把真正的创意决策权留给创作者。 - **互动界限划分**：在 Instagram 上使用 ManyChat 自动化回复消息以提效；但在 LinkedIn（核心根据地）和专属 Slack 社群中，坚持全部由本人亲自互动，以保持真实的情感联结和观点输出。 - **拒绝刻板化产品设计**：作为一名专注于数据科学与 AI 的技术创作者，她加入 Meta 创作者计划时发现大部分产品仍偏向传统泛娱乐创作者。早期许多 AI 剪辑软件预设了过于花哨的动效和彩虹滤镜，完全不契合严肃、对话式的技术科普风格。AI 工具开发者必须理解不同垂直细分领域（Niche）创作者的独特语调与剪辑诉求，提供高度灵活的模块化能力，而非用刻板印象限制创作者。 ```mermaid flowchart LR accTitle: Audible Creator-Centric Iteration Loop subgraph Loop["Audible 创作者中心化迭代循环"] direction TB A["创作者倾听<br/>(社交监听 / 问卷 / 1v1访谈)"] B["提炼真实痛点<br/>(音频制作 / AI工具瓶颈)"] C["产品路线图优先级排序<br/>(Creator Obsession)"] D["快速上线与功能迭代<br/>(功能反馈闭环)"] A --> B --> C --> D --> A end ``` Audible 的创作者与国际产品负责人 Sadaf Kazmi 阐述了平台方的设计哲学： - **“创作者至上（Creator Obsession）”**：Audible 将“以客户为中心”的原则延伸至创作者端，坚信只要解决创作者的真实痛点，商业回报便会水到渠成。 - **敏捷反馈闭环机制**： - *宏观社交监听（Social Listening）*：全面捕捉社区对音频与 AI 新技术的宏观讨论趋势。 - *定期与专项调研（Creator Surveys）*：针对成熟功能保持常规满意度跟踪，针对新上线特性进行针对性体验诊断。 - *深度 1 对 1 访谈*：深入挖掘头部与核心创作者的真实情绪与细分场景诉求，将反馈直接反哺至产品路线图的敏捷迭代中。 ### [29:17](https://www.youtube.com/watch?v=2kzMx4egyDY&t=1757s) Audible 的创作者反馈机制与声音克隆（Voice Replicas）测试 Audible 创作者与国际产品负责人 Sadaf Kazmi 介绍了平台近期推出的两项核心举措。首先是在独立出版平台 ACX（Audiobook Creation Exchange）上成立了“ACX 创作者反馈小组”（ACX Creator Feedback Group）。该小组汇集了顶尖的独立创作者，Audible 团队每季度与他们进行深入会谈，探讨即将推出的新功能、正在评估的产品方向、早期测试（Early Access）意向以及市场营销方案，借此获得极具价值的一线洞察。 其次，对于整个创作者产品线（尤其是 AI 相关功能），团队坚持采取小步慢跑的 Beta 测试策略。通过分批引入创作者进行入驻培训，观察实际运行效果并收集改进建议。Sadaf Kazmi 强调，近期在 ACX 推出的“声音克隆”（Voice Replicas）功能正是这一机制的典型代表——旁白朗读者（Narrator）可以为自己的声音创建 AI 副本，以此在 ACX 市场上竞标并承接更多有声书录制工作，同时对该声音副本的使用方式拥有完全的控制权，真正践行了“人类在环”（Human-in-the-loop）的赋能理念。 ```mermaid flowchart TD accTitle: ACX 创作者赋能与声音克隆工作流 subgraph Feedback["创作者反馈闭环"] A["顶尖独立创作者"] -->|"季度深入访谈"| B["ACX 反馈小组"] B -->|"评估新功能 / 营销方向"| C["产品团队迭代"] end subgraph BetaRollout["AI 声音克隆 (Voice Replicas)"] D["朗读者录入声音"] --> E["生成 AI 声音副本"] E --> F["保留完全控制权 / 竞标工作"] F --> G["明确标注并在 Audible 上架"] end ``` --- ### [31:05](https://www.youtube.com/watch?v=2kzMx4egyDY&t=1865s) Dylan 的数据分析工作流与有声书制作的现实痛点 Dylan Curious 对此表示赞同，认为未来社交媒体的虚拟化身（Avatar）必然会与人类真实声音深度贴合，而保障创作者所有权是其持续获利的关键。随后，Dylan 分享了自己将 ChatGPT 用作“反馈回路”的日常实操流程。他通常会从后台导出 YouTube 视频的各项数据并上传给 ChatGPT，输入一段非常直接的 Prompt：“这个视频获得了这么多播放量，我想让我的频道规模进一步扩大，你怎么看？”接着以此展开深度对话，要求模型对竞品进行横向对比。 Dylan 还会将各种多模态数据源喂给 ChatGPT，包括视频逐字稿、YouTube 每月后台数据、未来视频的脚本文档，甚至是那些逻辑看似莫名其妙却在 YouTube 上暴火的视频封面图（Thumbnails）。不论是日常分析还是处理来自 Discord 社区的反馈，ChatGPT 已经成为他不可或缺的高频工具。接着，Dylan 将话题引回有声书与 AI 旁白领域，他指出高质量的非虚构类（Non-fiction）有声书凝聚了作者深思熟虑的思想和编辑的反复打磨，制作极为耗时且昂贵，并询问 AI 旁白与自愿参与的声音克隆技术将如何帮助更多作者出版有声书。 - **Dylan 的多模态数据输入源**： - 单个视频与频道的每月 YouTube 后台分析数据（Analytics） - 视频生成脚本（Scripts）与已有视频的逐字稿（Transcripts） - 各类在 YouTube 上爆火但模式反常的封面缩略图（Thumbnails） - 来自 Discord 等社群的互动与反馈信息 --- ### [32:58](https://www.youtube.com/watch?v=2kzMx4egyDY&t=1978s) 填补 97% 的有声书空白：AI 旁白与创作者赋能 Sadaf Kazmi 指出，根据行业估算，目前市面上高达 97% 的在售图书根本没有制作有声书版本；而与此同时，有声书却是整个出版业中增长最快的细分领域，供需之间存在着巨大的断层。Audible 的核心目标是通过 AI 降低出版商、作者以及旁白员的准入门槛，让创作者自主选择如何呈现作品，使原本无力承担高昂录制成本的独立创作者和预算有限的作品得以有声化。此外，这一技术对于非英语的国际市场尤为重要，因为全球绝大多数语种的有声书资源严重匮乏。 Sadaf Kazmi 进一步介绍了在 ACX 上进行了约 6 个月的“声音克隆”Beta 测试情况。创作者对此享有全方位的控制权——他们自主决定为哪些作品试音、亲自编辑修整音频，成品在上架 Audible 销售时会进行极具透明度的明确标注（例如清晰注明为“Sadaf Kazmi 声音克隆版”）。测试反馈显示，创作者利用该功能有效拓展了商业规模并节省了大量时间。其中一个令人印象深刻的案例是：一位专业朗读者在完全失声的一整周里，借助自己的 AI 声音副本继续履行录制任务，不仅没有延误工期，也保住了该阶段的收入。更重要的是，创作者在此过程中获得了公正的版权报酬。 - **图书市场供需鸿沟**：约 97% 的现有图书缺乏音频版本，而音频又是出版业增速最快的板块。 - **国际市场语言失衡**：英语有声书数量远超全球其他所有语种，AI 旁白有助于填补国际小语种缺口。 - **声音克隆测试成效**：测试已开展约 6 个月；保障公平报酬；遭遇失声等意外状况时可保障创作者收入连续性。 --- ### [36:10](https://www.youtube.com/watch?v=2kzMx4egyDY&t=2170s) AI 标注透明度、版权保护与“说谎者的红利”（Liar's Dividend） 在探讨 AI 内容的标注与公众信任时，Ravit Jain 表示，在视频主要由 AI 生成或由数字虚拟化身出镜时，主动进行透明标注至关重要，这能维系创作者与观众之间的信任基石。但如果真人创作者全程参与，且 AI 使用比例仅占 10% 到 20%，那么是否单独标注的必要性相对较低。Ravit 强调，强制标注与合规监管对于防范侵权尤为关键——创作者最担忧的莫过于自己的声音或肖像在不知情的情况下被他人盗用并在其他社交平台（如 Instagram）上分发变现。因此，在推广声音克隆的同时，必须严格落实创作者署名与版税结算机制。 Dylan Curious 则引入了传播学中的“说谎者的红利”（Liar's Dividend）现象进行补充。在 YouTube 和 Instagram 等社交平台上，当某些 AI 生成内容的逼真度极高、仅在眼神或动作上有微米级的细微破绽时，观众往往会出于好奇反复驻足凝视，而平台算法会将这种异常的停留时长识别为高互动度并给予流量倾斜。这就导致未经合规标注的虚假内容反而更容易病毒式传播。Dylan 举例提到，此前网络上有一段“小兔子在蹦床上跳跃”的视频风靡全网，其实完全由 AI 生成，但由于创作者之间层层转载搬运，最终溯源链条极其漫长且找不到源头责任人，这正是当前平台治理亟待解决的挑战。 ## Notable quotes > [04:29](https://www.youtube.com/watch?v=2kzMx4egyDY&t=269s) 绝对会，我绝不可能把自己的社交账号完全托管给自动化。 — Aishwarya Srinivasan > [07:58](https://www.youtube.com/watch?v=2kzMx4egyDY&t=478s) 速度固然重要，但绝不能以牺牲身份特质（Identity）为代价。 — Ravit Jain > [09:49](https://www.youtube.com/watch?v=2kzMx4egyDY&t=589s) AI 的存在是为了增强创作者的能力，而非取代创作者 — Sadaf Kazmi > [13:22](https://www.youtube.com/watch?v=2kzMx4egyDY&t=802s) 永远只完成了 80% — Aishwarya Srinivasan > [19:40](https://www.youtube.com/watch?v=2kzMx4egyDY&t=1180s) 人类自带的生物性自然幻觉（natural biological hallucinations） — Dylan Curious > [23:58](https://www.youtube.com/watch?v=2kzMx4egyDY&t=1438s) 创作者至上（Creator Obsession） — Sadaf Kazmi > [31:05](https://www.youtube.com/watch?v=2kzMx4egyDY&t=1865s) 这个视频获得了这么多播放量，我想让我的频道规模进一步扩大，你怎么看？ — Dylan Curious > [36:10](https://www.youtube.com/watch?v=2kzMx4egyDY&t=2170s) 说谎者的红利 — Dylan Curious ## Mentioned Dylan Curious, Aishwarya Srinivasan, Fireworks AI, Ravit Jain, The Ravit Show, Sadaf Kazmi, Audible, Amazon, Amazon AGI, ACX (Audiobook Creation Exchange), HeyGen, ElevenLabs, Virtual Voice Studio, Gling, ChatGPT, Ash GPT, Weekend Musings, Fei-Fei Li, Google Veo, Captions, Canva, CapCut, ManyChat, Meta ## Video description Dylan Curious, AI YouTuber, Dylan Curious (YouTube) Alyssa Christensen, Newsletter Writer, GreenPulse Talent 2.0 Aishwarya Srinivasan, Head of Developer Relations - Growth, Fireworks AI Sadaf Kazmi, Head of Creator and International Products, Audible Ravit Jain, Founder & Host, The Ravit Show Abstracts This session will explore how AI-powered tools are transforming the creative process for content creators, enabling faster production, enhanced creativity, and personalized content experiences. Panelists will discuss how AI is being used to automate repetitive tasks, generate visual, audio, and written content, and optimize creative workflows. Discover how creators are leveraging AI to push the boundaries of creativity while maintaining authenticity and artistic control. The panel will also address the challenges and opportunities of using AI in creative industries and how to strike the right balance between human ingenuity and machine assistance. [Source](https://www.youtube.com/watch?v=2kzMx4egyDY)

[4:29] Aishwarya Srinivasan: 绝对会，我绝不可能把自己的社交账号完全托管给自动化。 [7:58] Ravit Jain: 速度固然重要，但绝不能以牺牲身份特质（Identity）为代价。 [9:49] Sadaf Kazmi: AI 的存在是为了增强创作者的能力，而非取代创作者 [13:22] Aishwarya Srinivasan: 永远只完成了 80% [19:40] Dylan Curious: 人类自带的生物性自然幻觉（natural biological hallucinations） [23:58] Sadaf Kazmi: 创作者至上（Creator Obsession） [31:05] Dylan Curious: 这个视频获得了这么多播放量，我想让我的频道规模进一步扩大，你怎么看？ [36:10] Dylan Curious: 说谎者的红利

[1:22] 80% accuracy [0:04] Ravit Jain: The Ravit Show [1:22] Aishwarya Srinivasan: Weekend Musings [23:58] Sadaf Kazmi: Customer Obsession [23:58] Sadaf Kazmi: Voice Replicas [19:56] Dylan Curious: What Stays Human [32:22] Sadaf Kazmi: Sadaf Kazmi's voice replica [38:10] Dylan Curious: liar's dividend

[0:04] Thanks to everybody for coming out. This [0:06] um should be a really kind of casual fun [0:07] conversation. We're going to just try to [0:09] share some insights into what we've [0:11] learned over the years. Um also we'll be [0:14] open to like some questions here at the [0:16] end here too. So I'd love to kind of [0:18] hear from everybody. But um yeah, we [0:20] should probably just start with a little [0:21] bit of a better intro for everybody cuz [0:24] when you're giving your intro, maybe you [0:25] could tell us like your name, who you [0:27] work for, but also a quick little uh [0:30] kind of linear progression of like [0:31] things you've done in the past and like [0:33] you know so we know where to ask [0:34] insights from. Yeah. Do you want to [0:36] start on the end? [0:38] >> Hi everyone, I'm Ashwar Shinavas and I [0:41] also go by Ash. Um I have two jobs I [0:44] would say like 9 to5 is I lead the [0:47] developer relations uh at an AI startup [0:49] called Fireworks AI which is a [0:51] generative AI inference engine and my 5 [0:54] to9 is being a creator. So I started off [0:58] as an accidental creator on LinkedIn. Um [1:00] something that I've been on that [1:02] platform for 10 years and accidentally [1:05] ended up having a lot of followers ended [1:07] up becoming a creator. Now I started [1:10] like on other platforms like Substack, [1:12] YouTube, Instagram etc. So that's what [1:14] I'm doing um outside of my 9 to5. So [1:17] very excited to be here at AI4. [1:20] Yesterday I had a more technical [1:22] conversation about AI agents and [1:24] building them scalably which was as part [1:26] of what I do at Fireworks and today I'm [1:28] here as a creator. So yeah, very excited [1:31] to have this conversation. [1:34] >> Fantastic. Good morning everyone. I'm [1:36] the founder and host of the Rav show. uh [1:38] I've been interviewing leaders in the [1:40] data and AI space since last 5 and a [1:42] half years now. The idea when we started [1:44] was obviously to amplify various voices. [1:46] There's a lot of salesy stuff that [1:48] happen and we see it all the time. I [1:50] wanted to make sure there are you know [1:53] conversations that really help those [1:55] enterprise leaders to scale fast [1:58] understand what's in the space and at [2:00] the same time uh they feel confident [2:03] about uh you know investing their time [2:06] in content. I understand there's a lot [2:08] of content around AI data uh but I just [2:12] wanted to you know make a mark where we [2:14] can have uh authentic conversations on [2:17] our show today uh it's almost a 1 [2:20] million big data AI community that we [2:22] have across all social channels YouTube [2:24] LinkedIn Instagram uh like Aisha uh it's [2:30] always a full-time gig does two jobs I [2:32] just do one and that is for myself uh [2:35] but Um yeah that's uh the idea is always [2:39] to you know make sure where we have the [2:42] right content for the right people out [2:44] there and yeah I'm excited to be here on [2:47] the panel and super excited to chat [2:49] about various things around data and AI [2:51] and how space is moving. [2:52] >> All right [2:53] >> thanks. [2:54] >> Um hi everybody I'm Saddaf Kazmi. I'm [2:57] head of creator and international [2:59] products at Audible. uh and my team is [3:01] really focused on ensuring that creators [3:03] have a seamless experience when they [3:05] interact with our company. Uh one of the [3:07] things that falls under my team's [3:09] purview is ACX or the audiobook creation [3:11] exchange and that is our platform for [3:14] independent publishing onto Audible and [3:17] Amazon for your audio titles. Uh for [3:19] people who may not be familiar with [3:21] Audible, Audible is the leading producer [3:24] of audio storytelling [3:27] of premium audio storytelling. Uh, and [3:29] we have over a million titles in our [3:31] catalog and we have customers in over [3:34] 180 countries and in over 50 languages. [3:37] >> Yeah, this will be great, too. I'm [3:38] excited to talk about long form versus [3:40] short form, too, in terms of audiobooks [3:41] because I'm a huge fan. So, um, and I'm [3:43] Dylan Curious. I have a YouTube channel. [3:45] I cover like all things artificial [3:47] intelligence, especially around like [3:48] society and, uh, the way people are like [3:50] relating to them, um, and just what it's [3:53] like to have an agentic world out there [3:54] along with humans. But let's go ahead [3:56] and start the conversation by talking [3:58] about sort of a speed versus voice kind [4:01] of conundrum that everybody's in. Like [4:02] these tools in some ways make things [4:04] super fast and then sometimes they [4:06] create things that you have to undo. [4:08] Sometimes they create things that don't [4:09] actually make the whole process faster [4:11] and sometimes you lose your own creative [4:13] vision in it. So um AI does make [4:16] creation fast, but style can be very [4:18] flat. It can be very generic sometimes. [4:20] How do you preserve your creative [4:22] fingerprint when you're using AI tools? [4:25] um like what's your method for keeping [4:26] it in your voice? [4:27] >> Yeah. [4:29] >> So, I'll start with saying that being a [4:31] creator is definitely very hard because [4:33] it's not just about being authentic. [4:35] It's not just about sharing your [4:36] thoughts, but the platform requests [4:39] consistency, which means you have to [4:41] post every day or like multiple times a [4:43] week. So, that does get to you at a [4:46] point where you're like, how much should [4:47] I automate and how well do I automate? [4:50] So from my perspective, I use a bunch of [4:52] different tools to help me through that [4:54] creator journey. But one of the things [4:56] that's helped me is planning out my [4:58] content. That helps me to break down the [5:01] automation pipeline in a way that I'm [5:03] controlling the top of the funnel, which [5:05] is like what are the ideas that I want [5:06] to put, what are the thoughts that I [5:07] want in there, what are the topics that [5:09] I want to cover, what are the theme that [5:11] I want on my on my socials. And then I [5:14] use each of these tools to do certain [5:17] parts of that workflow. So for example, [5:19] I use a combination of Hijen and um 11 [5:23] Labs for a video clone plus a voice [5:24] clone. [5:26] But those videos are only a part of the [5:28] videos that I put out. So 10% of my [5:31] videos that go out on Instagram or [5:32] YouTube are those AI generated videos [5:34] which again has my script that I [5:37] provided but it's generated because it [5:40] saves me time for producing those [5:42] videos. So it's a balance that you have [5:44] to make on how do you want to connect [5:46] with your audience? How do you maintain [5:48] that authenticity? Is it through just [5:51] recording a video of yourself? Does that [5:53] resonate as authenticity or does your [5:55] ideas resonate as authenticity? And uh [5:58] that's basically like how you maintain [6:00] that balance. And part of what I've [6:03] learned as a creator is [6:06] there are things that you would want to [6:08] talk to your audience about, but there [6:10] are things that the audience wants to [6:12] hear from you. And that's also a balance [6:15] which is very hard to crack because [6:17] there are days when I'm just craving to [6:19] put out the things that I want to say [6:21] but then I know that it may not be super [6:23] relevant for the people or like what [6:25] they're looking for or it might be too [6:27] advanced for them. So that's why I [6:29] choose different platforms for it. Like [6:31] LinkedIn is where I put more advanced [6:33] technical stuff. Instagram is more [6:35] breaking down AI topics in a more [6:36] microlearning manner. So, it's a balance [6:39] of like how do you balance what your [6:40] audience requires and what satisfies [6:43] your ego to like, you know, just put put [6:45] your thoughts out there. [6:46] >> Sometimes I feel like such a slave to [6:47] the algorithm. It's like the world is [6:49] having a conversation and they want to [6:50] hear our opinions and you just have to [6:52] throw it in there. But it's not [6:53] creative, you know. Um, so and just [6:55] really quick followup. Are you going to [6:56] keep a human in the loop, do you think, [6:57] for a long time in the future? [6:59] >> Absolutely. Absolutely. [7:00] >> There's no there's no way that I'm going [7:01] to automate my socials. [7:03] >> Okay. So it depends again like the the [7:06] answer is not that it's good to automate [7:08] or bad to automate. I do consult with [7:11] companies for whom it makes sense to [7:13] automate their socials. So if your brand [7:16] doesn't have a face, it makes sense to [7:18] automate because that saves you time [7:20] that still builds brand trust. There are [7:22] things that you can do outside of just [7:24] putting somebody's face out there that [7:26] builds trust. There are AI avatars. [7:28] There's a lot of social media where it's [7:30] entirely like nonhuman people who look [7:33] human, [7:34] >> right? [7:34] >> That's the face of the brand. But it's [7:37] not a right or wrong answer for me as a [7:39] creator. Like I I sort of have a [7:41] personal brand and that's where it makes [7:42] sense for me to like still maintain that [7:45] authentic videos for me. But if it's a [7:47] brand brand like an enterprise brand or [7:49] like a product brand, then it it is okay [7:51] to automate certain sections of that [7:53] more than a individual person. [7:55] >> Okay. What are your thoughts? Yeah, I [7:58] think Aisha covered uh interesting [8:00] points here. For me, I think speed is [8:03] important for sure, but not at the cost [8:05] of identity. I feel like uh human in the [8:08] loop has to be there when you're kind of [8:10] talking about AI or uh even working in [8:14] you know working with different tools [8:16] obviously helps a lot. There are two [8:17] different things how we can look at it. [8:19] uh maybe for the monotonous jobs where [8:22] you feel that okay uh you know you want [8:25] to schedule your content that's where AI [8:28] can play a very important role if I'm [8:30] doing an interview in my case uh and if [8:32] I want an agent to actually go out and [8:35] transcript everything while I'm [8:36] recording things and also create a [8:39] trigger where it can actually post [8:40] things for me socially uh that I'm at [8:43] AI4 doing this panel discussion and it [8:46] you know while maybe my co-founder there [8:49] is recording is actually triggering [8:51] everything and sending it on social at [8:53] the same time. That's speed. Uh it [8:56] definitely requires my voice to be on [8:59] the social at the same time. But how do [9:01] you train the model? You need to [9:03] obviously make sure where the model when [9:06] you chatting with may maybe it's Chad [9:09] GPD uh who's kind of you know triggering [9:12] all my social posts. uh I want to make [9:14] sure that chat GPD knows how I talk, my [9:17] experiences, the voice and that will [9:20] help create good content is what I feel. [9:23] So speed is important. Speed can be [9:25] maybe you know automating some [9:27] monotonous tasks and then uh you know [9:31] the voice can be how I train it well. [9:34] But again very uh good point where you [9:37] know still keeping your voice or making [9:40] sure what you're wanting to share uh in [9:43] that authenticity has to be there when [9:45] you're working with brands or even if [9:47] you're sharing it with with your [9:49] community because they look at you or [9:51] they've started even following you uh [9:53] for a particular reason of how you talk. [9:56] Maybe 5 years back there was no AI or no [9:59] hey gen out there uh and they were [10:01] always just maybe looking at you your [10:05] raw videos your authenticity and that's [10:07] why they started you know even following [10:08] and in the world where we are I I feel [10:11] like 50% of the creators still use a lot [10:13] of AI and it for me I can easily see [10:17] it's an AI so it uh and I'm pretty sure [10:20] the community also knows it so it [10:22] doesn't feel that authentic at times uh [10:26] so you can get slow on getting followers [10:29] and I'm pretty sure in the creator [10:30] business you would want more followers [10:32] and more community to follow you and [10:34] share their insights too. So yeah. [10:37] >> So this kind of a more interesting [10:38] question for you because you're thinking [10:39] of it from a platform point of view. So [10:41] is speed just mean like something like [10:44] 11 Labs for their audio books or do you [10:45] also think about the future where [10:47] they're interacting in different ways? [10:48] So yeah, same kind of question for [10:50] >> Yeah, it's a it's a great question. And [10:52] so as a product person, you know, we [10:53] think a lot about creators. That's our [10:55] primary customer. Um and uh you know, [10:58] for us, um one of our guiding principles [11:01] is that AI is there to actually enhance [11:04] what creators do, right? It's not there [11:05] to replace creators. So I love what you [11:07] guys had to say on this topic and we [11:09] think about it very very similarly which [11:12] is that these AI tools help they might [11:14] help speed up the process but we also [11:17] have our for example our virtual voice [11:19] studio product which is an AI narration [11:22] editing tool and that actually enables [11:24] creators to go in um you know make [11:27] updates to sort of edits al all types of [11:30] edits uh to the final product to ensure [11:33] that their creative vision is realized [11:35] um through uh the the end state or the [11:38] output of this which is assisted [11:39] obviously by AI tools um but at the end [11:43] of the day this is really about [11:44] enhancing what the creator does helping [11:46] them to be more efficient um but [11:48] ultimately help them achieve their [11:50] creative vision. [11:51] >> Yeah. Yeah. And I I mean I find it very [11:53] tricky. There's I'm I've thought about [11:55] my entire pipeline from coming up with [11:56] an idea for a YouTube video to actually [11:58] filming it. And what I thought I could [12:00] automate at first, which was something [12:01] like chatgbt, finding the articles [12:03] doesn't work for me because I don't [12:04] there's I have to go find the articles [12:06] on Twitter, on my newsletters that I [12:09] have some opinion on and it just can't [12:11] do that part. But then surprisingly, it [12:13] does amazing at editing. There's these [12:15] tools like Gle that I'm using for [12:16] editing where it cuts out all my ums and [12:18] o's and I've learned to start sentences [12:20] over and over again and it will [12:22] recognize if I've started it two or [12:24] three times and just give me the last [12:25] version. So, it it is a kind of amazing [12:28] way to make videos go faster. Um, and I [12:30] see all sorts of AI tools for just [12:32] cleaning things up and adding B-roll and [12:33] stuff sort of on the horizon. So, I like [12:36] those, but then um I can't I can't have [12:39] the conversation from the beginning done [12:41] by AI. It has to somehow have a seed of [12:43] of me in it that I I build from. So, um [12:46] but yeah, so for the next question, I [12:48] thought we could maybe move on to uh [12:49] like when AI gets things wrong. There [12:52] is, it leads into it well because [12:54] there's guard rails on certain models. [12:56] There's um halluc hall hallucinations in [12:59] some cases. There's all sorts of strange [13:02] administration things that pop out of [13:04] these and they're always changing. So, [13:05] how are you dealing with the um kind of [13:07] guard rails? And maybe you could share a [13:08] moment where AI generated some sort of [13:10] research or some kind of thing that you, [13:12] you know, used incorrectly and and it [13:14] was a hallucination or some other way [13:17] that went wrong. Can I give us a warning [13:19] where those might be? [13:22] So I'm going to give you a very naive [13:24] answer but that's some some something [13:26] that still works very well for me is I [13:28] have created something called as ash GPT [13:30] which is a custom GPT that sits on my [13:32] chat GPT and it's trained on about 4,500 [13:35] of my LinkedIn posts that I've made so [13:37] far and some of my Substack blogs as [13:38] well. Uh I've also like inserted a [13:40] transcripts of this um AI video podcast [13:43] that I used to do called weekend [13:44] musings. So it basically has a lot of my [13:48] content from the last few years that [13:51] helps it understand my style, the [13:53] technical depth that I want to get into [13:56] the anology style, the storytelling, all [13:59] of my personality traits. That's [14:01] something that's easier for me to like [14:03] use as a draft script writer. [14:07] Second thing, I have a really really [14:08] really long system prompt that goes into [14:11] it really. [14:12] >> So and it's not something that I came [14:14] overnight. It was something that I [14:15] started with a week later I was like [14:17] this is let's go back reiterate it [14:20] write it again 2 days later again add [14:22] something more structure more format [14:24] more so it's it's a constant iteration [14:27] process it's still not not the most [14:28] perfect thing I still have to do a lot [14:30] of review and edits after it gives me [14:33] something but it's still good to give me [14:35] a first draft because there are times [14:37] when I'm coming back from work at like 6 [14:39] p.m. and complete brain fog, you know, [14:41] so I'm like get gets me started with [14:44] something and then I can go and edit [14:46] edit it out. I think that's where like [14:48] you know Raith was mentioning the human [14:50] in the loop part, right? It cannot be [14:51] neglected. I can never have I can never [14:53] imagine having a pipeline where I use [14:56] even ash GPT to like be the script [14:58] writer and hun to be my video uh [15:00] companion and captions to be my AI edit [15:02] and just post it on socials. each and [15:05] every time I have to manually go and [15:07] check every single thing because it [15:10] hasn't been one situation where any of [15:13] the output was perfect in the first [15:15] draft. So I don't trust it. [15:17] >> And and how close is it getting? I mean [15:19] is it 80% there, 50% there? I mean how [15:21] many hours after you get your prompt? [15:23] >> It's I I would say it's very hard to say [15:24] because every time that I feel it's 80% [15:27] there and it improves and then I go back [15:29] and I'm like it's still 80% there. [15:31] >> It's almost like a human problem. your [15:32] your expectation from it also keeps [15:34] increasing, right? So you always feel [15:36] that it's constantly consistently at [15:38] 80%. [15:39] >> So that's that's what I feel about it. [15:41] And yeah. [15:42] >> Yeah. Part of me just thinks it's always [15:44] going to be like a half hour or an hour. [15:45] Like you just have to get tired of [15:46] iterating and then you feel like you're [15:48] finally done. [15:48] >> Yeah. [15:49] >> But anyways, sorry. Go ahead. [15:51] >> Yeah. I just remember one of the you [15:53] know um facts that actually happened [15:56] with me when you know when Chad GBD kind [15:58] of came in and it was almost just like [16:00] actually two years later when we kind of [16:02] started uh trying to create content out [16:05] of it and somehow I saw you know wanted [16:08] to put out a prompt that say you know [16:10] wanted to just ask about my show about [16:12] uh who Ravit is or the Ravit show is uh [16:16] just wanted to test the grounds there as [16:17] well and what I see is uh it actually [16:21] you know told me about something which I [16:23] never did an interview which I never [16:25] hosted and also spoke about something [16:28] some topics which were totally [16:30] unrelated. So that's when you know it [16:32] kind of uh felt like okay there's some [16:36] mixup here and that's the real [16:38] hallucination that can happen. But over [16:40] the time obviously we've seen uh the [16:43] models becoming more intelligent. [16:45] There's obviously GPD5 which has already [16:47] come out recently. So things have [16:49] changed but like I should have said I [16:52] think uh at every stage when now we want [16:55] to publish something we want to go out [16:58] and check if it really has my voice or [17:02] not at every uh like I can train model [17:05] as much as I can um you know having [17:08] 4,500 posts maybe on uh which can be [17:12] trained on LinkedIn on chat GPD uh I [17:16] feel is huge but still If Aisha feels [17:19] it's she's 80% there. Imagine uh for [17:22] someone like us who maybe I've just [17:24] trained it uh I don't have my model over [17:27] there but uh um yeah if if I feel if uh [17:32] you know I'm going out and uh you know [17:35] even training it for like say 100 hours [17:38] it will still not get there. So you need [17:41] to make sure when you are going out and [17:43] even prompting something or even putting [17:47] out your content, you can make it [17:48] intelligent but it won't be perfect. It [17:52] will take time. It's getting there. Uh [17:54] but we still make sure everything's in [17:57] place. I make sure where if there are [18:00] certain interviews that we've done and [18:02] we want to you know try and copy that [18:05] style of content uh we kind of still go [18:08] out and you know uh put it into insert [18:11] the content and then get the results out [18:13] of it. So there's still a long way to go [18:16] but yeah I think uh you need to be very [18:19] careful when you're using it. [18:21] >> Okay. So is this is this a problem at [18:23] all for you since the books are [18:25] pre-written by authors? Well, uh, you [18:27] know, we have obviously AI hallucinates [18:30] and, um, you always need to take a look [18:32] at quality. So, we spend a lot of time [18:34] actually benchmarking quality and doing [18:37] a lot of testing related to quality. So, [18:39] that's where it sort of manifests, uh, [18:41] for us. And we also look at um, you [18:44] know, how do listeners receive this? Um, [18:47] and so when we do our own internal [18:50] testing to ensure something meets our [18:51] own high quality standards, um, we think [18:54] about what are the metrics that we use [18:56] and how do we ensure that it's a great [18:58] listening experience. So listeners will [19:00] tell us if they think it's, you know, if [19:03] it's good, if it's good enough, if it's [19:05] excellent, right? And that's all [19:07] information that we take in and [19:08] insights. And do you have your own model [19:10] that you're developing at Audible for [19:12] the voice uh like text to voice or [19:16] >> so we actually we partner with our [19:18] internal voice experts uh within uh the [19:21] AGI organization at Amazon. [19:23] >> Okay. Yeah, that's I'm sure they're [19:25] working on some really amazing stuff [19:26] there too. I've had Yeah, for me [19:28] hallucinations have kind of been a [19:29] problem in certain times especially when [19:31] I'm going on tangents. I sometimes do [19:34] these uh train of thoughts where I'm on [19:37] camera and I'm like let's just go there [19:38] audience like how does quantum mechanics [19:40] work and then I ask the question and [19:41] just read the answer and then I see the [19:43] comments come in and they're like this [19:44] that's not right and then I'm trying to [19:46] fix it. So you know I definitely have [19:48] learned to take it with a grain of salt [19:49] and always sort of couch my comments [19:51] when I read it directly from there but [19:53] it's getting better. Feels like it's [19:54] more trustworthy now than ever. Um let's [19:56] talk about for the next topic what stays [19:59] human um in the whole idea of creation [20:03] like what is the most important thing [20:05] that that you you fundamentally just [20:07] can't let go of when you're um running [20:10] Audible or when you're creating your own [20:12] personal content that um it just has to [20:14] be you if you get rid of that you don't [20:16] feel the same way about your product. [20:20] So, [20:21] >> so I'll probably piggyback on what we [20:22] were talking about previously, right? [20:24] Like the reason why I always feel that [20:26] any of the AI tools that I'm using is [20:28] 80% there is it's not physically [20:30] connected to my brain. So, every time [20:32] I'm feeding it information of the posts [20:35] that I've written or like my thoughts, [20:37] all of that are very partial [20:38] information. Every single day, the [20:41] information that I am grasping, things [20:43] that I read, people that I meet, [20:44] conversations that I have, that still [20:46] stays only in my brain. So that [20:48] information is not something that I can [20:50] share with the model or it's like or [20:52] I'll have to like physically connect my [20:53] brain to it. And that's where the human [20:56] part comes in, right? Like the [20:57] experiences that I have, the thoughts [20:59] that I that changes every single day for [21:01] me. Um maybe the kind of ideas that I [21:05] have every single day or like every [21:06] single minute. That's not something that [21:08] I can communicate with an AI model all [21:10] the time. And those are something that [21:13] makes that content more authentic. If [21:15] you were to like ask me to write about a [21:18] specific topic right now, it is going to [21:20] get influenced by the recent activities [21:22] that has happened, it's going to get [21:24] influenced by what I heard from Fay [21:25] Philly this morning, what I heard from [21:27] Jeff last morning. That that information [21:30] doesn't exist in those AI models. It [21:31] doesn't exist in ag. So all of that is [21:34] what makes the authentic content more [21:37] human. And that's why I hope I'm still [21:39] needed for my social media more than the [21:41] AI models out there. [21:44] >> Okay. [21:45] I think that totally makes sense. I'll [21:47] keep it short and I'll you know [21:48] obviously just focus on something which [21:50] is very important uh when you even [21:53] chatting with AI which is human. So [21:55] human brains are still way too smart uh [21:59] more intelligent and [22:02] we in the end it's us who's going to [22:04] prompt and you know get the results out [22:06] of it. So I feel if you need the right [22:10] authentic content, you need to prompt it [22:12] that way and you'll get those answers. [22:15] And I think nothing better than human [22:18] brain. So uh if you want that authentic [22:21] content, you'll have to go out and share [22:25] mind share you know your uh thought [22:27] processes with the AI and you'll get [22:31] results maybe close to what you're kind [22:33] of thinking. That's it. [22:35] >> Yeah. [22:36] Yeah. So, do you have any thoughts? If [22:38] not, I'll [22:39] >> I'm interested to hear you. [22:41] >> I think I just think humans have each [22:42] human has their own brand of natural [22:45] biological hallucinations that make them [22:48] interesting and it's not worth cutting [22:50] those out. I think when people go on [22:52] tangents and they see the world in a way [22:54] that's very different from you, it's [22:55] it's fascinating. It's interesting. It [22:57] It's not necessarily something you want [23:00] is some like perfect everybody acts the [23:02] same kind of thing. But um you know I I [23:05] was I'll throw in a couple tools like [23:06] there were tools like um Google VO and [23:09] I've been prompting it and in some ways [23:11] when I ask for a video a 30 or 40 second [23:14] video um I give it a few prompts and I [23:17] imagine something and what it imagines [23:19] sometimes is very entertaining but it's [23:21] very different from my brain and the the [23:23] delta between there like the difference [23:25] between there is something that I don't [23:28] want to close a lot of times. So, I want [23:30] to keep prompting it until I get [23:31] something close to what I want, but then [23:33] sometimes I'm also impressed by it. And [23:35] I feel like it changes my style of [23:37] comedy um or my style of writing. So, [23:40] yeah, I don't know. I think having our [23:42] own hallucination is kind of the main re [23:43] like knowing how you hallucinate, how [23:45] the weird way you see the world is and [23:47] just being very core to that is going to [23:49] be the key in an AI future. So, you [23:50] don't just kind of wander around and [23:52] copy people like and you keep a voice [23:54] that attracts people. So, um but yeah, [23:57] we can jump to the next question now. [23:58] Let's talk about uh creator centric AI [24:01] product development. Um this is [24:04] definitely something where I think [24:05] audible will have some insights for us [24:07] but um we'll still go down the line I [24:10] guess but creator centric AI product [24:12] development. Um it sounds like you not [24:14] only have this huge prompt but you've [24:15] done some finetuning um how how do you [24:17] think we build products? [24:20] >> How do you think who builds products? uh [24:22] for for yourself like creator central uh [24:25] create creator centric AI product [24:27] development um so like the products [24:29] you're using the products you might want [24:31] to create around your own brand [24:34] >> I do have some pipelines that I have [24:35] built for myself and I really appreciate [24:39] platforms like um I'm just talking about [24:41] the ones that I use every day which is [24:43] hey gen 11 labs captions canva uh cap [24:46] cut to some extent so all of these tools [24:50] give me the ability to do things that I [24:54] want to do without really taking away [24:56] the [24:58] creativity part out of me. So, it still [25:00] gives you like the toolkit of what you [25:02] want to create, but then at the end of [25:04] the day, you're the one who's creating [25:05] it. [25:06] >> Yeah. And also throwing in like the [25:07] community like are using Discord, are [25:09] you using AI to like read all of your [25:11] comments? [25:12] >> Uh stuff like that. [25:13] >> Many chat is what I use for automation [25:16] on Instagram, but LinkedIn is something [25:17] that I always do myself. like it's it's [25:19] my OG social media platform. So, [25:22] LinkedIn is something that I completely [25:23] do myself. Um I do have like Slack [25:26] communities where I engage with people [25:28] but it's also something which I do [25:29] myself. So, there's parts of it which I [25:32] don't want AI to be involved because I [25:35] feel that's the value prop that I have [25:36] to provide the audience that I'm able to [25:38] like share my latest thoughts or like [25:41] you know honest feedback about certain [25:44] things. But I I feel like the whenever [25:46] you're building any product which is [25:48] keeping creators in mind, it is to [25:50] understand that everybody has a unique [25:52] voice. Everybody has a unique style of [25:54] storytelling of creating of topics that [25:56] they want to create about. Um, a lot of [25:59] times people think it's funny that I'm a [26:01] data science and AI creator because [26:03] whenever even with the meta team that I [26:05] met recently, I'm part of their meta [26:06] creators group and I'm probably one of [26:09] the very few tech creators because [26:11] creator in the in its natural sense [26:13] meant something very different until [26:15] recently that we started having tech [26:16] creators. So having that in mind is [26:19] important that there's going to be [26:21] different kinds of creators creating in [26:23] different niche. So you don't want the [26:26] product to bottleneck them in certain [26:29] categories or you don't you want to like [26:31] basically understand their mindset, the [26:34] kind of content they'll create, the kind [26:35] of formats they'll use, the kind of [26:37] editing that they would require um and [26:40] the kind of tools that would be [26:42] democratized to all of the creators [26:44] without and giving them the flexibility [26:46] to use it the way they need it. So [26:49] rather than like trying to stereotype of [26:51] what kind of creators come in like I've [26:53] used some earlier apps um for video [26:56] editing and it felt too crafty for me. I [26:59] was like [27:00] >> it's it's too much you know like I don't [27:02] want it to be so crazy edited like my [27:05] videos are very conversational. They are [27:06] about technical topics. So if it's like [27:08] edited with rainbows around them like [27:10] that's not that's not my kind of videos. [27:12] So it it needs to like let people have [27:15] their own creativity and like flourish [27:17] in their own niche rather than trying to [27:19] um trying to like do something which is [27:22] not not their authentic voice. [27:24] >> So yeah. [27:24] >> Yeah. [27:26] >> Go ahead. [27:27] >> Okay. Sure. I can I can chime in here. [27:29] Um so uh customer obsession is a sort of [27:31] a a core guiding principle uh within [27:33] Audible. It's our people principle uh or [27:36] our first people principle. Uh for my [27:38] team that customer is the creator, [27:41] right? So, we obsess over what creators [27:42] think. Um, so we're always thinking [27:44] about whether we're meeting our creators [27:46] painpoints. Um, are we solving the [27:49] challenges that they're facing? And [27:50] within both audio as well as the AI [27:54] space, things are changing so quickly [27:56] that it is so critical for us to um just [27:59] iterate and sort of um uh solicit input [28:03] rapidly uh and make sure that we're [28:04] responding to that. Um, so as we think [28:06] about those pain points, we use that as, [28:10] you know, into our we pull that [28:12] information into how we prioritize our [28:15] road maps. Um, and what we've actually [28:18] seen, and I think the results, you know, [28:20] bear this out over and over. If you're [28:22] meeting the needs of your creators, then [28:24] the business results follow. So, we [28:26] really think about creator obsession as [28:28] sort of a core um a core tenant and core [28:31] value of the company. [28:32] um we look at a bunch of different [28:35] pieces of information to ensure that [28:37] we're incorporating uh input into the [28:39] feedback the feedback and insights into [28:42] our overall development process. So part [28:44] of that is you know we'll do social [28:46] listening we'll understand what the [28:48] conversations are that are happening. We [28:50] want to get that sort of broad-based [28:51] view. We'll do creator customer surveys. [28:54] Sometimes those will be at a regular [28:56] cadence for a more mature product, but [28:58] as we roll out new capabilities, we'll [29:00] ask specifically about, you know, did [29:02] you like this feature? What can we do to [29:04] to to make it better? Um, we also have [29:07] one-on-one conversations when we're [29:09] trying to go deeper to really understand [29:12] um uh you know, where creator sentiment [29:14] is on a particular topic. Um, and there [29:17] are a couple of things that I think are [29:18] are worth mentioning as well that we're [29:20] doing now, uh, that I think are really [29:23] relevant to this. The first is within [29:26] ACX, our indie publishing, um, [29:28] independent publishing product that I [29:30] mentioned earlier, we started recently [29:33] the ACX creator feedback group. And [29:36] through that, which is the group is made [29:38] up of some of our top independent [29:40] creators. We meet with them quarterly [29:42] and we have more detailed conversations [29:45] about uh new features we're rolling out, [29:48] things we're considering, um, you know, [29:50] capabilities that are already out there, [29:52] how do they, you know, how are they re [29:53] reacting to it, are they interested in [29:55] early access, you know, marketing [29:57] initiatives, etc. And that's been an [29:59] incredibly valuable way to really go [30:01] deeper in in that conversation. Um, and [30:04] the second piece is that uh, and we do [30:07] this more broadly within my creator uh, [30:10] product organization, but especially for [30:12] our AI capabilities, we roll everything [30:15] out slowly as a beta and we ramp up and [30:17] that gives us an opportunity to [30:19] incorporate creator feedback regularly [30:22] into the development process. You start [30:24] with a small batch of creators. We on [30:26] board them. We see if it's working. We [30:28] see what they like. We see where they [30:30] think that there's an opportunity for [30:32] improvement. And we've done this most [30:34] recently with a capability that we [30:36] launched called voice replicas again on [30:38] ACX. And this enables narrators to [30:41] actually create a an AI replica of their [30:44] voice and then use that to secure jobs [30:47] on the ACX marketplace um and also have [30:50] full control over how that voice is [30:52] used. Um so this goes back to the sort [30:54] of you know human and in AI. um you know [30:57] this really enables full control uh from [31:00] the creator on something that is based [31:02] on their voice um and that sounds like [31:04] them. [31:05] >> That's great. Yeah, that's kind of the [31:07] way I always imagine social media going [31:08] like we've given it all this data and [31:10] eventually the avatars become closer to [31:11] closer to our voice and that's cool that [31:13] you're keeping keeping it in line with [31:15] what the creators created in the first [31:16] place so they can keep getting uh [31:19] supported by that. But yeah, I'm uh for [31:21] me I I just have sort of a simple prompt [31:23] that I'm using with chat GPT and lots of [31:25] little data sources. So I will export my [31:28] analytics from a video. I will upload it [31:29] to Chad GPT and say uh this video got [31:32] this many views. I'd like my channel to [31:34] be bigger. Thoughts question mark and [31:36] then dump it in. And then that sometimes [31:38] like starts a conversation depending on [31:40] what it says back to me that I can dig [31:41] into or like ask like ask it to look at [31:44] my competitors or something like that. [31:46] And I do it with all sorts of data [31:47] sources. Sometimes it's like the [31:49] transcript from a video. Sometimes it's [31:50] like a monthly analytics download from [31:52] the YouTube back end. Um, sometimes it's [31:54] like the scripts that I'm preparing for [31:56] a future video. Sometimes I'll throw in [31:58] images of thumbnails that seem like they [32:00] make no sense, but they're blowing up on [32:01] YouTube. So, um, yeah, I I find the CHBD [32:05] is kind of the the tool that I'm using [32:07] almost all the time for any kind of [32:08] feedback loop, including like Discord or [32:11] something like that. So, um, yeah, I [32:13] find that a really interesting way just [32:15] to pull that out. But um let's move to [32:17] the next question. So let's talk about [32:19] uh AI narration and voice replicas. I [32:22] guess we started talking there, but [32:24] there's questions about access um cost [32:27] um security. So um especially with [32:30] audiobook creation, it's it's expensive. [32:32] Like you know, one thing I love about [32:34] audiobooks is that it's not somebody sat [32:37] there and thought about the most [32:38] important things in their brain. They [32:40] refined their ideas. They bounced it off [32:42] an editor for the most part. And [32:44] especially because I'm into really like [32:46] non-fiction, it's just so time inensive [32:47] and I don't want to lose that kind of [32:50] sense of content in there. Um, how does [32:53] AI narration and opt-in voice replicas [32:56] let more authors publish audio books? [32:58] >> Yeah, this is a great question. Um, so [33:00] by some estimates, 97% of books that [33:04] exist today actually don't have an audio [33:07] version. uh and at the same time audio [33:09] within publishing is actually the [33:11] fastest growing category within that [33:13] particular industry. So it's a pretty [33:16] dramatic disparity. Um and so a lot of [33:20] what we think about is how to lower the [33:23] barriers to entry for all of our [33:25] creators. So that's publishers, that's [33:27] authors, that's narrators, and enable [33:30] them to have choices for how they bring [33:33] their work to life. Um, and so we think [33:36] AI narration actually really really [33:38] helps give uh creators choice and lowers [33:42] those barriers to entry and increases [33:45] the the creators ability to produce more [33:48] if they choose to do that. Of course, [33:50] that's always up to the creator. Um as a [33:53] result we think that that gives real [33:55] access to um titles that would not have [33:59] been produced in audio otherwise um [34:01] because it can you know uh have higher [34:04] barriers to entry for audio titles. Um [34:07] this also means that it really benefits [34:09] independent creators who are often on a [34:11] budget uh as well as international [34:13] marketplaces that are often underserved [34:15] by audio content. So there are way more [34:18] titles for example that exist today in [34:20] English than they do in a lot of other [34:22] languages and most other languages [34:24] across the world. Um so this is a really [34:26] important um sort of foundational uh [34:29] piece of of why we're doing this um as [34:32] we think through our voice replica beta [34:35] on ACX. One of the things that we've [34:38] seen and people are using this today um [34:41] but this has really enabled our [34:43] narrators to create a replica that [34:46] sounds like them and then to use this to [34:48] scale scale their production [34:50] capabilities and scale their uh their [34:53] business um again if they they choose to [34:55] do so. Um one of the things that we [34:57] think is great about this is that [34:59] narrators do retain full control, right? [35:02] They choose which works they want to [35:03] audition for. they're able to edit their [35:06] title, make it their own, and feel [35:07] really good about whatever is ultimately [35:10] published uh and sold on Audible. It's [35:13] also clearly labeled, so it'll be [35:15] labeled as, you know, if I were [35:16] narrating, and nobody wants to hear me [35:19] narrate an audio book for the record, [35:20] but um it would be labeled as Saddaf [35:22] Kazmi's voice replica. Um so we also [35:25] retain a lot of transparency. Um this [35:27] has actually been in beta for about 6 [35:30] months. Um, and so we have a lot of [35:32] really great feedback already. And what [35:33] we've heard is that people are using [35:35] this to scale their businesses. Um, [35:38] they're saving a lot of time in the [35:40] process. Uh, one narrator actually said [35:43] that she lost her voice and she was able [35:45] to continue to work during the week [35:47] where she had totally lost her voice. [35:49] So, she didn't have to forgo the work [35:50] and she also didn't have to forgo income [35:53] that that brought in. Uh, which was [35:55] really great. Um, and narrators have [35:58] also really appreciated that they are [36:00] compensated fairly for the use of their [36:02] voice through this program. So, lots of [36:04] really great initial feedback as we [36:06] continue to refine and incorporate the [36:08] ideas and feedback into our development [36:10] process. [36:10] >> Okay. Yeah. And kind of jumping to trust [36:12] too. Um, are you guys label like how [36:14] much do you feel like you need to label [36:15] or disclose about how much is AI or like [36:19] does it erode the audience at all in any [36:21] way or like where do you kind of think [36:23] about all that stuff? Yeah, that's a [36:25] good question. And I feel like sometimes [36:28] uh you know it's very much necessary to [36:31] go out and put out that it is uh an AI [36:35] you know generated video and you know [36:37] it's maybe an aftar that's kind of [36:39] talking just to make sure that there's [36:41] transparency and the you know the [36:44] audience kind of is connected to you [36:46] very well that you're being transparent [36:48] about things but then for me at least [36:52] I'll make sure that it's if it's just [36:55] 10% or 20% [36:57] um then I want to make sure that uh and [37:00] the whole video is not an AI uh [37:02] generated avatar then okay it's fine if [37:05] I want to or I don't want to because I'm [37:07] still there in uh but I think labeling [37:09] is very much important because sometimes [37:12] uh you can be replicated by others as [37:15] well and your content can be running [37:18] somewhere else where you don't know and [37:20] that is a big red flag for creators and [37:24] they feel uh threatened because uh maybe [37:28] your content is being you you are [37:30] promoting someone's content on some [37:32] Instagram which you don't know about and [37:35] uh to you know so I kind of feel AI [37:38] replicas and the voice replicas are [37:40] fantastic uh but it just needs to be [37:43] made sure where the creators get the [37:46] best uh attribution for uh what they're [37:49] doing and the royalties are in check. uh [37:52] but yeah if uh it all works out then [37:56] yeah I think uh you know labeling is a [37:59] very important part not only just from [38:01] your end but anyone who's you know even [38:04] who you're working with and are creating [38:06] maybe replica voice replicas of yourself [38:09] >> yeah I know on uh social media more so [38:11] on the YouTube space but also kind of on [38:13] Instagram there is something called the [38:14] liars dividend which is this idea that [38:16] you can make something so close to human [38:18] that there's extra attention given to it [38:21] as you're looking at it cuz the eyes are [38:23] just micrometers off or VO has done [38:26] something. And then of course the [38:27] algorithm loves that extra attention. So [38:30] the stuff that's improperly labeled [38:32] sometimes, at least recently, has been [38:34] doing really well. And I I obviously I [38:36] know YouTube thinks that's a problem and [38:38] they're trying to figure out how to kind [38:39] of get around this. But I don't know if [38:40] any anybody see those little rabbits [38:42] that were jumping on the trampoline like [38:43] went super viral but they were all fake [38:46] you know and it took a long time for all [38:48] the creators to sort they kept sourcing [38:50] other humans on humans on humans it went [38:52] down this long chain and finally it was [38:53] like there is nobody at the end of this [38:55] chain everybody just copied it from [38:56] someone else