ElevenLabs Reaches $400 Million in Revenue on Flat Team Structure

ElevenLabs凭借扁平架构实现4亿美元营收 语音技术将从模仿人类的客户支持，转向具备情商的销售智能体，以及完全绕开人类语言的自动化智能体间协议。 在本次讨论中，Mati Staniszewski详细讲述了ElevenLabs从波兰的高中同窗情谊发展为领先的前沿音频AI企业的历程。通过优先考虑资本效率、即刻变现以及完全远程的团队模式，该公司开发了涵盖文本转语音、翻译和实时对话智能体的完整模型套件。Staniszewski分享了一些反直觉的运营经验，例如保持无职衔的扁平化架构、将工程师直接派驻非技术部门，以及重金投入拥有1000多名专家的专业数据标注团队以捕捉情绪的细微差别。展望未来，他探讨了智能体间协商的前景、情绪智能与通用音频智能等即将推出的功能，以及即将到来的信任瓶颈——即数字安全体系必须发生反转，转而验证由人类授权的真实AI。 ElevenLabs 由高中好友 Mati 与 Piotr 共同创立，源于波兰单人单调译制片的童年体验；公司在 2022 年选择以高资本效率的远程模式切入音频领域，目前仅以约 400 人团队实现了超过 4 亿美元营收。 (0:22) 技术架构已从上下文感知的文本转语音拓展至语音识别、多语言配音、对话编排与音乐生成，接下来的重点是具备情绪感知匹配能力的语音 Agent 以及融合叙述与歌唱的通用音频智能。 (0:51) 语音 Agent 正在从基础客服向高价值的来电与外呼销售场景扩展，同时在乌克兰前线公民服务平台以及 MasterClass 沉浸式互动教学等关键领域实现规模化落地。 (0:57) 为了在大规模下保持敏捷，ElevenLabs 采用无职级的扁平化管理并将团队控制在 10 人以内，同时聘请了由 1,000 多名专家组成的数据标注团队，攻克音频在情绪传递等层面的长线能力短板。 (14:11)

ElevenLabs co-founder Mati Staniszewski argues that voice is emerging as the primary interface for artificial intelligence, built on domain-specific emotional nuance rather than sheer compute scale. To sustain defensibility, the company pairs capital-efficient research and cross-functional automation with long-term investments in audio artistry, platform integrations, and identity authentication. Capital Efficiency and Cross-Functional Engineering (0:22) Founded in 2022, ElevenLabs took a capital-efficient path by focusing on audio—a compute-light domain largely overlooked during the broader technology industry's focus on crypto and the metaverse. The founders recruited remote researchers by scraping GitHub for audio talent in London, Warsaw, and beyond, prioritizing rapid monetization to maintain healthy margins before taking on larger external capital. Operating four years later with over $400 million in revenue and roughly 400 employees, the company retains a flat hierarchy with no formal job titles and limits individual teams to fewer than 10 people. To maintain development speed as non-technical employees increasingly write code using AI tools, ElevenLabs embeds research engineers directly into legal, sales, and people operations. Modeled after tech practices observed in the Ukrainian government, this structure provides internal code review and produces custom automation. For example, the go-to-market and legal teams created an automated scoring system that allocates negotiation points for liability and indemnity clauses scaled by customer size, removing manual legal approval bottlenecks. Overcoming Jagged Intelligence Through Nuanced Data Labeling (0:51) ElevenLabs expanded its model suite from contextual text-to-speech into speech-to-text, localization dubbing, conversational streaming engines, and music generation. Staniszewski highlights key technical breakthroughs, including replicating distinct accents, synthesizing natural laughter with conversational pauses, and localizing Matthew McConaughey's newsletter into Spanish and Portuguese. Upcoming production targets include voice agents capable of emotional state detection and audio general intelligence models that transition seamlessly between speaking and singing in a single stream. However, Staniszewski notes that audio models currently exhibit "jagged intelligence": while dependable for structured customer support and sales, they remain too slow to capture natural emotional cadence and cannot reliably compose hit music. To address these gaps, ElevenLabs invests in long-term data labeling rather than near-term commercial capture. The company employs an annotation team of over 1,000 specialists—including professional voice coaches and musicians—to categorize emotional delivery and musical characteristics on 10- to 24-month development horizons. Commercial Voice Agents and Agent-to-Agent Coordination (0:57) Voice agents are transitioning from customer support phone trees into revenue-generating sales and operational roles. Deliveroo deploys voice agents to call restaurants and verify operational hours, while Deutsche Telekom and ElevenLabs utilize conversational inbound sales agents, finding that prospective buyers communicate pain points far more thoroughly by voice than through static forms. In the public sector, the Ukrainian government deployed ElevenLabs voice channels to provide frontline updates, children's educational lectures, and safety alerts. In education, MasterClass uses the technology to offer dynamic interactive training with instructors like Gordon Ramsay and Chris Voss. Looking toward autonomous agent negotiations, Staniszewski observes that current deployments remain limited to basic logistical coordination and price capturing. Dynamic negotiation will require emotional pacing and strategic interruption rather than simple interruptibility. Furthermore, inter-agent communication will likely abandon human voice altogether; during a hackathon, two voice agents that recognized each other as AI immediately switched from spoken English to an alternative data transmission protocol to transfer information more efficiently. Audio Artistry, Ecosystem Defensibility, and Identity Authentication (23:31) Citing a conversation with Nvidia CEO Jensen Huang, Staniszewski emphasizes that while speech-to-text is pure technology, text-to-speech is artistry requiring deep tuning for emotional resonance across distinct verticals. Beyond core models, ElevenLabs builds defensibility through its platform ecosystem, which integrates telephony systems, enterprise knowledge bases, monitoring tools, and a community library of over 20,000 user-contributed voices. As ambient robotics and smart devices position voice as a primary interface for machine intelligence, Staniszewski identifies identity verification as a critical structural bottleneck. When personal autonomous agents routinely execute transactions and transmit sensitive medical or personal data, the security paradigm will invert: digital communications will be treated as synthetic by default, shifting defensive focus toward authenticating that an AI agent has been explicitly authorized by a specific human principal. While ElevenLabs demonstrates that specialized audio models can scale profitably through lean operations and proprietary data labeling, the long-term viability of ambient voice interfaces depends on overcoming latency deficits in emotional nuance and developing secure infrastructure to authenticate agent-to-agent transactions.

根据红杉资本对话节目的讨论，ElevenLabs 联合创始人 Mati Staniszewski 阐述了语音如何成为下一代人工智能核心界面的技术路径。他指出，音频模型凭借独特的情感艺术性与更轻量的算力需求，正在从单向语音生成演进为承载销售、政务和教育的交互式代理，而未来的长期壁垒取决于精细化数据标注与身份信任机制的建立。 轻资本路径与音频模型演进 (0:22) ElevenLabs 的创立灵感源自联合创始人 Mati 和 Piotr 在波兰成长的共同经历：传统波兰影视作品长期依赖单一男性声调为全片男女角色进行旁白式配音，剥离了原作的情感张力。这促使他们确立了打破语言与情绪隔阂的创业愿景。与主流前沿大模型动辄数亿美元融资的开发范式不同，ElevenLabs 在 2022 年成立时选择了资本高效路线。团队当时敏锐察觉到音频属于被忽视的细分赛道，且音频模型的参数体量与算力开销远小于文本和视觉模型。他们通过在 GitHub 筛选检索伦敦和华沙等地的顶尖音频研究者，组建了全远程团队，并以快速商业化所产生的现金流支持底层研发。 在技术路线上，公司从具备上下文理解和情感语调的文本转语音（TTS）模型起步，随后拓展至语音转文本与本地化翻译工具，并结合高速推理模型构建起低延迟的实时流式对话引擎。其研发历程中的关键节点包括完美复刻创始人特定口音、还原自然的笑声与犹豫语气词（如“ums”和“ahs”），以及协助演员马修·麦康纳以西班牙语和葡萄牙语向海外受众分发音频简报。团队目前正重点研发两大前沿特性：一是能感知用户情绪并自适应调整语速与抚慰语调的情感智能；二是能在叙述与演唱之间无缝切换的通用音频智能（Audio General Intelligence）。 超越呼叫中心的应用场景拓展 (0:57) 语音代理的商业落地正迅速超越传统的客服问答与电话自动导航，向具备高转化价值的入站及出站销售场景转移。例如，外卖平台 Deliveroo 利用语音代理直接致电餐厅核验实时营业时间，并自动将变更同步至骑手端和用户端；德国电信则引入该技术处理入站销售咨询，取代繁杂的静态文本表单。ElevenLabs 自身运营实践也表明，潜在客户在与语音代理口头交谈时，透露的业务痛点和使用场景深度显著高于传统网页表单。 在公共事务与交互式教育领域，语音交互展现出更具社会价值的落地潜力。乌克兰政府部署了 ElevenLabs 的技术以建立公民支援服务热线，向民众提供前线战况实时更新、儿童教育广播以及主动安全警报。在教育领域，MasterClass 接入了该交互引擎，开发出让名厨戈登·拉姆齐在虚拟厨房动态指导学员、或让前危机谈判专家克里斯·沃斯提供高拟真对抗训练等沉浸式课程，推动教育从单向倾听转变为实时博弈与互动练习。 工程下沉与无头衔敏捷架构 (14:11) 成立仅四年的 ElevenLabs 实现了超 4 亿美元的年化营收，而团队总人数仅略微超过 400 人。公司维持了极度扁平的组织设计，各独立研发、产品、市场与运营单元的规模通常控制在 10 人以下，多数管理者的直属汇报人数在 10 人左右。为确保运转速度，公司采用了一项与乌克兰政府技术团队相似的架构创新：将专职研发工程师直接派驻并嵌入到法务、人力资源和市场拓展等非技术职能部门中。 这一机制有效应对了当前非技术人员利用 AI 编写部署代码时普遍面临的代码审查与系统安全短板。在实际业务中，驻场工程师协助招聘团队搭建了针对候选人全流程的自动化抓取与分析管线，并对法务团队进行 AI 工具培训。针对过往销售谈判中旷日持久的免责声明与责任上限条款审批，团队开发出一套根据客户采购规模自动分配“谈判点数”的自动化评分系统，彻底消除了层层人工流转的瓶颈。同时，公司全面取消正式职务头衔，确保个人的业务晋升与发展完全受实际产出驱动，而非受制于入职司龄。 机器协商、信任瓶颈与数据壁垒 (17:04) 展望代理间交互（Agent-to-Agent）的前景，Mati 指出当前应用仍局限于初级的价格收集与物流协调，但已有早期初创公司尝试让 AI 自主致电多家场馆并在预算区间内进行多轮价格磋商。真正的动态谈判对语调、停顿节奏乃至“策略性打断对手”提出了极高要求。在技术交互层面，早期测试显示，当两个语音代理识别出彼此均为 AI 时，会主动弃用人类语言，切换为传输效率更高的专用数据传输协议，这预示着底层通信协议未来可能分化。 与此同时，代理的大规模普及引发了严重的信任与身份核验瓶颈。当个人 AI 代理常态化代理人类发起医疗或交易通话时，数字环境的防御逻辑将颠倒为“默认一切皆虚假，直至得到验证”，安全重点将转向对人类合法授权的“真实 AI 身份”的签名认证。此外，在应对复杂情绪节拍与高水准流行音乐生成时，音频模型仍存在明显的“参差智能”（jagged intelligence）。为此，ElevenLabs 组建了由专业配音导师、音乐人等千余名专家构成的标注团队，开展跨度达 10 至 24 个月的细颗粒度表现力与声乐特征标注，并结合超过 20,000 个社区声音及电话协议集成，构筑深层技术护城河。 Mati 借用英伟达 CEO 黄仁勋的观点指出，语音转文字是确定性的技术，而文本转语音则是需要深度共鸣的艺术。ElevenLabs 的探索表明，语音作为未来物理机器人与智能环境的核心入口，不仅不会稀释真实人际交往的价值，反而会推高现场实体互动的溢价；但整个行业若要迈向高度自洽的自主代理生态，仍须在解决毫秒级情感韵律断层的同时，在全行业范围内建立起抗伪造的身份鉴权基础设施。

[0:02] Mati Staniszewski: Time is on our side. We are still best friends. It's working out [23:31] Jensen Huang: our speech to text models are technology and text to speech is artistry and we are all artists.

[0:02] So, I love uh line charts and bar graphs [0:05] as much as the next guy, probably more. [0:08] Uh the story of 11 Labs is also [0:10] interesting from a human perspective. [0:12] But, is you started a company with a [0:14] childhood friend. So, maybe take us back [0:16] to 2022 or earlier and just tell the the [0:19] human side of the 11 Labs story to [0:20] start. [0:22] The I have the I have the most luck in [0:25] the story of 11 Labs because well, it [0:26] started in 2022 it felt feels like it [0:28] started 17 years ago when I met my my [0:31] co-founder Piotr. [0:33] Uh all the names in Polish are [0:34] complicated luckily for for for us, but [0:37] we we met in high school, became best [0:39] friends, took all the same classes [0:40] together, and then through the years did [0:42] everything together. So, we traveled [0:43] together, studied together, worked [0:44] together, and time is on our side. We [0:46] are still best friends. It's working [0:48] it's working out. Um and part of what [0:51] started 11 Labs is is is inspiration [0:53] from where we are both from. We are both [0:55] from Poland. [0:57] Uh so, I was of Warsaw. And there's a [0:59] very peculiar thing in Poland. If you if [1:01] you watch any foreign movie in Polish [1:03] language, all the voices, whether that's [1:05] a male voice or a female voice, get [1:08] narrated with one single character. [1:10] So, as you can imagine, pretty terrible [1:12] experience. You have literally one voice [1:15] narrating everything. [1:17] Um it usually also on purpose is kept in [1:19] monotone. So, you are meant to interpret [1:21] your own emotions for that content. And [1:24] uh while we grew up with this, this is [1:25] still happening today for majority of [1:27] content. And that kind of opened our [1:29] eyes into one of the clear things across [1:32] the domain across audio domain um across [1:35] the future will be this ability for [1:37] everybody to speak any language with the [1:39] same emotion, the same intonation. Um [1:41] and we started diving deeper into that [1:43] problem and realized the problem of [1:45] audio exists in so many other domains, [1:46] too. Whether that's narrating the [1:48] content around us, um whether that's the [1:51] books not being available in audio form, [1:54] whether that's the and articles that we [1:56] could read, whether that's that language [1:57] barrier or in the future as we heard in [1:59] the previous conversations, the future [2:01] where human noise the robots are around [2:02] us, the voice will be the primary [2:04] interface to a lot of that technology [2:06] and and something we would love to fix [2:07] and solve. [2:09] Excellent. [2:10] Um and 11 Labs builds frontier models [2:13] for audio. I think there's a paradigm [2:16] now where to build a frontier model, you [2:18] have to start with hundreds of millions [2:20] or billions of dollars and then figure [2:21] out the rest later. 11 Labs did not take [2:24] that path. Can you talk a little bit [2:26] about your approach towards building [2:29] this company, why you have this hasn't [2:31] been replicated, is that even possible [2:33] in 2026, [2:34] etc. [2:36] Yeah, that goes I think that continues [2:38] that great luck and timing because we [2:40] started in 2022. For those of you [2:43] working in the domain at the time, that [2:44] that was the year of crypto [snorts] [2:47] and metaverse. Nobody was still working [2:49] on the AI side. [2:50] Um [2:51] even further, [2:52] people were starting to work, of course, [2:54] on the text models and the visual [2:56] models, but audio as a domain was still [2:59] considered a big niche with so few [3:01] researchers in the space working on on [3:03] that work. So, for us that was [3:05] a good part of picking that domain where [3:08] A, we were excited about where that [3:09] future is called, we felt that people [3:11] around just didn't realize the value of [3:13] that domain, but three, the requirements [3:15] of what you needed to solve were very [3:16] different. The audio models were [3:18] smaller, [3:19] so you don't need as much computer as [3:20] you need for some of the other sister [3:22] domains. [3:23] The data needs are big, um but while [3:26] there's a lot of audio data, we knew [3:28] that the thing to actually get that [3:29] audio working, you will need to figure [3:31] out how to transcribe a lot of that data [3:33] and annotate a lot of that data, which [3:34] we knew we can do. And then ultimately [3:36] it all boiled down to architectural side [3:39] of can we can we solve that part [3:42] in a good way. And here my co-founder is [3:45] one of the smartest people I know [3:47] and and a great researcher and has been [3:49] able to assemble some of the best people [3:51] in audio to to help us. and we took a [3:53] slightly untraditional approach at the [3:54] time. We started we started in London. [3:57] We had a lot of people between London [3:58] and Warsaw and started a company in [4:01] remote completely remote way. So we [4:03] wanted to hire the best researchers [4:04] wherever they were. [4:06] We were going through the classic GitHub [4:08] scraping and and and trying to reach [4:10] people based on their work instead of [4:12] based on their presence and based on [4:14] that work we would reach out to those [4:16] people. [4:17] We would always share our examples and [4:18] try to get them to join the team and [4:20] that's how we assemble the first the [4:21] first set of of of people who we think [4:23] are some of the best [4:25] researchers in that audio domain and [4:27] through the years they still help us [4:29] crank a lot of those models in into [4:31] production. [4:32] Um [4:33] then we launched the product. I think [4:34] the slightly different approach we took [4:36] was monetizing very quickly. So trying [4:38] to get some of the revenue stream back [4:41] so we can fund a lot of the work and the [4:43] models. We try to stay uh [4:47] stay [4:48] healthy on the margins so we can [4:50] continue investing with the assumption [4:52] that it's better for us to figure out [4:54] that stream and be be able to be [4:55] independent in that development. Um but [4:58] then as the ambitions grew we knew that [4:59] we needed to train models so we of [5:01] course brought a lot of [5:02] money externally as well. And I think [5:04] like projecting to today one thing [5:06] that's clear for us is there's still so [5:09] many of those niches that people don't [5:11] tackle that that you can start with and [5:13] then step-by-step start opening them [5:16] them up. [5:17] I think a lot of customers see 11 Labs [5:20] through their narrow needs, right? Um [5:23] maybe take a zoomed out view. Like what [5:25] is the suite of models that 11 Labs [5:28] works on? How do you prioritize them? [5:29] How do you organize R&D? Etc. [5:32] Yeah, so we started um we started with [5:35] the first text-to-speech model. So the [5:36] model that could finally understand the [5:38] context of what's being written and [5:39] based on that context context [5:41] understanding get the right [5:43] emotion the right intonation from text. [5:45] So if it's a happy sentence you get that [5:47] happiness out. If it's a uh it's a [5:49] dialogue, it can pronounce the dialogue [5:50] out. And then continuously started [5:53] adding that. So, it started with the [5:54] problem of of breaking down language [5:55] barriers. [5:57] Um [5:58] the things you need to solve dubbing is [6:02] transcription, so understanding, then [6:04] the the translation, and then [6:05] text-to-speech. So, we first solved [6:07] text-to-speech. Then we knew we needed [6:09] to add the other component, which is [6:10] speech-to-text and being able to [6:11] transcribe content in a in a great way. [6:14] Then how we combine those models [6:15] together. [6:16] So, that's kind of was the first three [6:17] models in the first first couple of [6:19] years. And then of course the other [6:21] things started happening across the [6:22] space, which is that a lot of the [6:24] reasoning models started becoming quick [6:26] enough and smart enough at the same [6:27] time, where you could imagine those [6:29] interactive experiences being possible. [6:31] And that's where we started launching [6:32] our [6:33] more of the real-time streaming models [6:35] across audio, and then combining those [6:38] into conversational experiences. So, we [6:40] added effectively all the stack, all the [6:42] turn-taking and orchestration to create [6:44] a voice engine for a voice for a voice [6:46] agent. Um and then on the other side as [6:49] we realized that the emotionality is [6:51] something we can solve, uh we added some [6:53] of the hardest modality in in in audio, [6:56] which is music and being able to produce [6:58] music. So, today we spun entirety of the [6:59] research of audio, where it's [7:01] text-to-speech, speech-to-text, um [7:03] combining those models together in both [7:06] localization with dubbing, with uh [7:08] orchestration with voice agent voice [7:09] engine, and then uh and then being able [7:11] to do that across music as well. [7:14] And what's the um [7:15] all those things and all that [7:17] interesting development work [7:19] uh was there any oh wow moment in terms [7:21] of what these products are capable of [7:23] that you can you can remember? [7:25] Yeah, there's so many and it's kind of [7:26] the bar changes for all of us. The first [7:28] moment for us uh was [7:31] well, first moment for us they always [7:32] use my voice as a testing voice because [7:34] it has this weird accent. And uh and the [7:37] first time was like when when we could [7:38] replicate my voice based on a good [7:40] sample, that was like a first wow moment [7:42] to to myself and you always go for this [7:44] moment like this is not how my voice [7:45] sounds like and then you listen to [7:47] yourself side by side and it's like [7:49] definitely how it sounds like. [7:51] Uh uh unfortunately, [7:54] uh then the the second moment was where [7:56] we first got it to laugh [7:58] and people were like, "Okay, this is [8:00] actually the thing that that makes the [8:01] the whole experience more human. The the [8:04] laughter, the pauses, the ums, the ahs, [8:06] the imperfections." So, we started [8:08] getting those uh out and uh those were [8:10] the moment for us because we made it to [8:11] the top of Hacker News with the first AI [8:14] that can laugh uh uh a model, which was [8:16] a very proud moment for us. One of the I [8:18] feel like pinnacles of the voice [8:20] performance, Matthew McConaughey giving [8:23] his newsletter and his iconic lines in [8:26] in in Spanish and Portuguese, where for [8:27] the first time his family who speaks [8:29] that language could hear him speak those [8:31] languages, too. Um but for more recent [8:34] pieces, the two two ones that we are [8:36] excited about bringing to production, [8:39] I think the first one is uh finally [8:42] figuring out the emotional intelligent [8:44] in that interactive experience. So, in [8:46] the voice agent experience where it [8:49] doesn't only get the right uh intonation [8:52] emotion, but can understand the other [8:53] side. If somebody is stressed, it gets [8:55] and delivers that uh that that soothing, [8:58] reassuring emotion. If if someone is [9:00] excited, maybe it matches that. If [9:02] someone speaks slowly, it makes sure to [9:04] slow down. And that emotional [9:05] intelligence is something that we are [9:07] finally seeing internally uh a path to [9:09] solving, which will be just a a a [9:12] a continuous step change to to what's [9:14] possible. And then the second one, [9:17] uh which will apply there, but also [9:19] apply into general audio spaces, audio [9:21] general intelligence where you can [9:23] combine audio models together in one [9:25] stream. So, you could theoretically have [9:27] a model that narrates, then pauses and [9:30] let's say starts singing with that same [9:32] continuous voice. And that's something [9:34] that's extremely hard to combine today [9:36] and and something that would be would be [9:38] possible, I think, very very soon. [9:41] And voice imagine, you know, voice [9:42] agents. And it seems like everybody is [9:44] at least on the customer side, [9:45] everyone's buying a voice agent. Um [9:48] and I think intuitively you think [9:49] customer support, um [9:51] you know, the old phone tree [9:52] replacement. Uh what's actually going on [9:55] in the world of voice agents? And what [9:57] do you think are are the most [9:58] interesting overlooked opportunities, [10:00] spots where startup founders should [10:01] focus? [10:03] Yeah, the of course the customer support [10:05] is probably the one that everybody heard [10:06] and and knows about very very well. Um I [10:09] think the second thing and the second [10:10] thread we are seeing is increasing shift [10:12] to revenue-generating opportunities [10:14] where voice agents can act in sales, [10:16] whether it's inbound or outbound sales [10:18] of of sales. It doesn't replace the [10:20] entire experience, but takes and [10:21] amplifies part of that experience. Um [10:24] maybe a good example is Deliveroo, where [10:26] Deliveroo will have voice agents that [10:28] contact the restaurants to capture their [10:31] opening times. And based on their [10:33] opening times, they can update the [10:34] riders and drivers. And of course, the [10:36] people ordering on when to get to to [10:38] that work. All the way through to the [10:40] inbound sales, where increasingly [10:43] people, that's a good example of [10:45] Deutsche Telekom, will be contacting to [10:47] inquire about the service, inquire to [10:49] buy a [10:50] a product. And [10:52] instead of going through the drop down, [10:53] instead of going through the form, you [10:54] can speak with the voice agent to leave [10:56] that information. Uh we do it ourselves, [10:58] too, so we have a good metrics of an [11:00] understanding of what's happening there. [11:02] One, of course, so much simpler and [11:04] quicker to go through instead of going [11:07] through that form. But the second thing [11:08] that started happening in that inbound [11:09] sales flow is we uh we had a lot more [11:13] information that people started leaving [11:15] because they would speak about the use [11:16] case they are coming with, but then [11:18] where it's not working, where it's [11:19] working, some of the other use cases [11:20] that they are evaluating. Which we can [11:22] combine and then just deliver such a [11:24] much better experience afterwards. On [11:26] the overlooked side, [11:28] uh [11:29] I think my favorite example there's the [11:32] citizen support education and healthcare [11:34] will completely change. On the citizen [11:36] support, like all of us would [11:39] would benefit from just generally better [11:41] government access, whether that's [11:43] understanding how to fill in the taxes [11:45] that I think many of you went through [11:47] earlier this this month, [11:49] all the way through to just learning how [11:51] what is the policy for travel abroad and [11:53] and and and and how that might affect [11:55] the the the space. We've recently seen [11:57] that work deployed in government of [11:59] Ukraine. We think is like one of the [12:00] most advanced governments on that front. [12:03] We traveled to Ukraine [12:05] working with their team and what they [12:06] are trying to solve is they they have a [12:09] a government app which every citizen can [12:11] access and get information about what's [12:13] happening, but given the war, given the [12:16] the the frontline and lack of that [12:17] access, they wanted to figure out a new [12:19] channel for people to be able to call in [12:21] and get that information. So, they [12:23] created voice agent effectively [12:25] where you can where you can call in and [12:27] get the information about what's [12:29] happening on the frontline. You can get [12:31] education help and some of the lectures [12:33] delivered to your to your kids all the [12:35] way through to proactive engagement [12:38] about about staying safe and staying [12:40] staying out there. And maybe last [12:42] example on education front, and that's [12:43] probably my favorite one as I think [12:45] about that changing, it's it's just how [12:47] incredible would it be to have a uh [12:50] someone that is an incredible teacher [12:52] available 24/7 we can ask him questions [12:54] whether it's Karpathy all the way [12:56] through to Richard Feynman and and you [12:58] can learn physics with them on the [13:00] headphones while you are teaching that [13:01] subject or learning that subject. And [13:04] and that's something that we are seeing [13:06] pockets of. Like a great example is [13:08] MasterClass where MasterClass of course [13:10] collaborates with incredible teachers to [13:12] deliver static lectures, but recently [13:14] they launched an interactive version of [13:16] that. Uh so, for I don't know if that [13:19] will be a good reference for for this [13:21] this audience, but we [13:23] we recently worked with them on bringing [13:24] Gordon Ramsay that can teach you [13:27] cooking. So, while you're in the kitchen [13:29] he can shout at you effectively [13:31] to get to get better or maybe a better [13:33] one there is a Chris Voss where you can [13:35] of course learn negotiation but you can [13:36] learn by negotiating with Chris live on [13:40] the phone to to to to get better which I [13:43] thought was a phenomenal subject. Having [13:45] negotiated against Mati a number of [13:47] times around financing rounds, I [13:48] understand now. I think it helps you to [13:50] say this but I think the [laughter] [13:51] opposite opposite is true. [13:53] Um [13:55] Uh ask more questions I want to save [13:56] time for the audience as well. Um maybe [13:58] one [13:59] as Constantine mentioned more than 100 [14:01] million of net new ARR in Q1. Obviously [14:03] the business is going very well. [14:05] Um and you're sort of pioneering the [14:07] startup founder building a foundation [14:09] model applications. [14:11] Um any counterintuitive lessons about [14:13] building a company in this era that for [14:15] the founders in the audience they may [14:16] want to take home with them. [14:19] So we are just for reference we are [14:22] uh uh just over 400 people over 400 [14:24] million in revenue uh but still keep the [14:27] teams extremely small so it's it's like [14:30] rough arbitrary a little bit cop is is [14:32] less than 10 people is for each of the [14:34] the research product even the the [14:36] go-to-market ops talent teams are all [14:39] smaller than that size. Um [14:41] most of people will have 10 direct [14:43] reports or so it keeps it relatively [14:44] flat and allows us to move move a little [14:46] bit quicker. One thing that we've done [14:48] which is in this model and very [14:50] surprisingly this is very similar model [14:51] that we've seen actually with the [14:52] government of Ukraine. Each of the teams [14:55] even the teams that aren't [14:57] uh technical teams will have engineers [14:59] within them. [15:00] So our people team our go-to-market team [15:02] our uh legal team will have an engineer [15:05] in that team that helps to build of [15:07] course automation upskill uplevel the [15:09] the rest of the people and recently that [15:11] really helps because as I'm sure many of [15:13] you are going through everybody will be [15:15] wipe coding and coding a lot of the the [15:18] help even if they are not technical so [15:20] now that kind of shifted the [15:22] responsibility not the not the [15:23] responsibility but shifted the [15:25] requirement of how good the review needs [15:27] to be for all of that work. Whereas [15:28] security infrastructure implications, [15:31] you will want to make sure that the [15:32] output is right. [15:34] And everything on the engineering side [15:35] you can put that expectation on the [15:36] non-engineering side the [15:38] the ability to do is that is is [15:40] relatively hard. So that technical [15:41] research in those teams helped us a lot [15:43] to to figure this this out. And and in [15:46] general there's just so many incredible [15:48] work you can do by having that whether [15:50] that's the scraping on the hiring and [15:52] recruiting front or analyzing what [15:53] worked in the past to improve in the [15:55] future, whether that's [15:57] upscaling the legal team on how to use [15:59] those tools and then figuring out ways [16:02] of recently introduced a scoring system [16:05] for those on the go to market on the [16:07] sales side. You frequently will end up [16:09] in this negotiation with your sales team [16:10] of can I give [16:12] indemnity provisions? What's the [16:14] liability cap? Can I give this set of [16:16] clauses? And then you kind of need to [16:18] draw the line of how many things you [16:19] give. And I ended up being in so many of [16:21] those conversations that we gave already [16:23] a lot or we didn't. So now we introduced [16:25] a scoring system that you can give per [16:27] per size of the customer you can just [16:30] give a few of those points out and in. [16:32] Um [16:33] we just made it so much easier and of [16:35] course that's fully automated now with [16:36] with with how we work across that team. [16:38] So that was one of the unintuitive small [16:41] teams bringing technical talent in the [16:43] non-technical teams, [16:45] keeping relatively flat. We also have no [16:47] titles which allows us [16:49] to to to bring people and and really [16:51] optimize for impact that they are having [16:53] and then you can grow as quickly as as [16:55] as as you want. The tenure will not [16:57] define this [16:58] and and many more. So we'll see. It's [17:00] four years old company. So we'll see if [17:02] that's helps. [17:04] Any questions? [17:07] Oh no. Okay, Sonia. [17:10] Are you seeing people deploy voice [17:12] agents to actually negotiate on their [17:14] behalf? And then when you are you [17:17] starting to see agents actually [17:18] negotiate with agents? Um and sorry, I I [17:21] do three-part questions. Uh [17:24] when that when that world happens, do [17:26] you think the agents are actually [17:27] talking to each other the way that [17:28] humans talk to communicate and [17:30] negotiate, or do you think it's beep [17:32] boop beep boop? Do you think it's, you [17:33] know, it's all done instantaneously? [17:34] Like how how's how's that world going to [17:36] look like? [17:37] So, one early inklings of that, we [17:40] haven't seen any truly successful on the [17:41] negotiation front. It was like more, you [17:43] know, kind of order taking. What's the [17:44] price? Can we capture that? And then [17:46] kind of goes back to the team. So, not [17:47] real negotiation. But, there's uh [17:49] there's few startups that we see [17:51] especially on any any like [17:53] organizational shifts of can I organize [17:56] this event, calling calling a lot of [17:58] places, getting the price, and then [17:59] calling again with like our budget. So, [18:01] that is happening, and I think this will [18:03] shift. I think it emotional intelligence [18:05] will like This is the big part that will [18:07] start being important in a lot of that [18:09] work, where it's not only the content [18:10] that matters, but how you deliver, when [18:12] you pause, uh uh that work. And then [18:14] maybe the extreme version of that, which [18:16] agents are uh [18:18] are not like most of the people wouldn't [18:20] do it, and and they are not good at [18:22] that, is today you will see a lot of [18:24] interruptibility built in, where human [18:26] can interrupt the agent. But, with [18:28] negotiation, you also want the opposite, [18:29] where agent will interrupt the human. [18:31] It's kind of the extreme version of [18:32] that. On the second part, on the [18:34] agent-to-agent part, [18:36] uh uh [18:37] Some of you might have seen this uh the [18:39] the we did a hackathon over a year and a [18:41] half ago, and that was exactly the case [18:43] where agent was speaking with another [18:44] agent. They detected that they are [18:47] both agents, and they swapped over to [18:49] the to the different language, and [18:50] that's like a more of a uh [18:52] more efficient transmitter of [18:54] information than just the the classic [18:57] um [18:58] um spoken word. And uh I think this will [19:00] happen. I 100% like the the the big [19:03] question will be really voice, will it [19:05] be other transmission of information? Um [19:08] it depends truly on what the [19:10] infrastructure is built for, and I think [19:12] this will define that that experience. [19:15] Yes. [19:18] Okay. [19:19] So you're going to catch box. [19:21] Hey. [19:21] >> [snorts] [19:22] >> Curious how you're thinking about the [19:24] need for voice in a future where agents [19:26] do more and more of the work. [19:29] So basically what are the kind of use [19:30] cases maybe where human conversation I [19:33] think it's more of a follow up to the [19:34] last question. [19:36] Like first [19:38] you all of us will have so many [19:40] different devices around us and [19:43] step from that you will have robots [19:44] around us. So of course voice will be [19:46] such an important interface to to [19:48] instruct and and be able to interact [19:50] with those those those those devices. In [19:52] many ways I feel like the you know we [19:54] see a lot of developments of of [19:55] intelligence but then the the real [19:57] bottleneck of the future will be how we [19:59] communicate with that intelligence and I [20:00] have voice and visual part will will be [20:02] a big unlock to be able to actually get [20:04] the most of that intelligence value in [20:06] those settings which [20:08] which which which isn't [20:10] yet possible. [20:11] But on the flip side it's a it's yet the [20:14] value of the human to human interaction [20:16] will only increase. So like the the [20:19] whether that's the events like this one [20:20] whether that's events with your favorite [20:22] artist [20:23] will will increase in value [20:26] with that ability of having voice all [20:29] all around you. [20:31] But the trust will be such a big part [20:33] and something we optimize for like in [20:34] between the agent and human of you know [20:36] in the in the future where all of you [20:38] will all of us will have a voice agent [20:39] for example to call and book a [20:41] restaurant or [20:43] give information to a health care [20:45] appointment. [20:46] All of that will require such a high [20:48] degree of trust that this is you and and [20:50] and authenticated user. There'll be like [20:51] a level of [20:53] encoding and decoding for real then [20:54] encoding decoding for water might opted [20:58] in human and then by default everything [20:59] else will be [21:01] fake which is kind of the opposite of [21:02] how it is today. You detect for AI but [21:05] you will detect for real authenticated [21:07] AI in the future and assume it's fake. [21:12] Uh if you could pass it that way. [21:17] Uh Andre spoke earlier about jagged [21:19] intelligence. Do you see similar odd [21:21] places in audio where models are good [21:23] and the bad that you might not expect [21:25] and yeah, what are they? [21:29] The [21:30] um there's still so much on the on the [21:32] on the bad side. I think the [21:34] you know, like we spoke a little bit [21:35] about where we see the voice agents [21:37] working. So like this combination of the [21:38] models together um and support settings [21:42] works really well, works reliably. And [21:44] early sales starts working, but like the [21:46] moment you start swapping to a true [21:48] emotional interaction, not yet working. [21:51] It's um it doesn't get the emotion that [21:53] that well. It's slightly too slow. [21:56] Uh so that is still like um I think a [21:58] big step change that should work. Um [22:01] same will apply on a on a very different [22:03] domain on on the music side. I think in [22:05] the music side you [22:07] uh [22:07] you can get you can get good production [22:10] music. You cannot get top charts music [22:12] even with artist input. Um I think this [22:15] will change over the next over the next [22:17] year or two. Can I ask a follow-up on Of [22:18] course. Andre's take was that the the [22:20] the reason for that was that the labs [22:21] were basically training for the stuff [22:23] that had economic value. Where you're [22:25] training your models, is that true of [22:27] you? Are you basically training for the [22:28] things that make the most money or is it [22:30] that there are some challenges that are [22:32] genuinely harder than others? [22:35] The [22:36] I don't know. We we try to train the [22:37] models, build a product and an ecosystem [22:39] that will drive of course the biggest [22:40] impact for for for for all our [22:43] customers, all users, which should [22:44] correlate of course with the revenue in [22:46] the long term. So like that long term [22:48] perspective, that's it's going to be [22:50] like minimal over the next few years, so [22:52] not next year. [22:53] Um [22:54] so frequently we will train the models [22:56] that might not provide that value in the [22:57] short term. Or even step before, we'll [23:00] like I so much time labeling the data, [23:02] not only the what of audio, but also how [23:04] of audio. Like what emotions did I use? [23:07] Um what is my voice described as? What [23:09] is this music described as? So we [23:10] assembled a team of now thousand plus [23:13] people that have been voice coaches, [23:14] musicians, artists before that can help [23:16] us annotate that behind the scenes. And [23:18] that will not provide value in the next [23:20] 6 to 12 months, but we think it will in [23:22] the next 10 to 12 to 24. Um and then you [23:25] of course need to collect that data, [23:26] which frequently just isn't that [23:28] accessible as well. [23:31] Hey, last one and then we'll go to [23:32] lunch. [23:37] Hey. [23:38] Can you hear me? Thanks. Big fan of [23:40] yours and 11 Labs. Thank you. [23:42] What do you think from from the model [23:44] air perspective, what do you think are [23:47] the moats here with with audio models? [23:50] The labs are going there, not going [23:52] there. What are the kind of, you know, [23:55] in this sausage making of making a real [23:57] good frontier audio [snorts] model? What [23:59] what are the the main defensible parts [24:02] there? [24:03] At the um [24:04] So of course we do a [24:06] a variety of models and [24:08] recently had a pleasure of meeting [24:09] Jensen and he was commenting on a few of [24:11] those models and he said that our speech [24:13] to text our speech to text models are [24:15] technology and text to speech is [24:17] artistry and we are all artists. [24:19] So he gained a client for life. [24:22] But of course we do believe there's a [24:24] little bit of that to to really fix text [24:26] to speech and fix that emotionality. You [24:28] you you you will need to be really [24:30] focused on that space. You really need [24:32] to get in front of users, collect the [24:33] data, collect the preferences, use that [24:36] to fine tune the models and then there [24:37] is a domain specificity in how you [24:39] actually bring those models to [24:40] production. In health care very [24:42] different than in financial services, [24:44] very different than in education or [24:46] experiences. So that's on the model [24:47] layer. I think there will be continuous [24:49] advantage that if you actually care [24:50] about the quality, [24:52] that like actually spending the time on [24:55] the model work will will will help you [24:56] keep that advantage. But to your point, [24:59] the models and like a lot of use cases [25:01] will use a model as just a small part of [25:03] their stack. And that's where we spend a [25:04] lot of time and like beyond going beyond [25:07] the research on on the product side of [25:08] how you understand a user's problem, the [25:10] workflow that they need. [25:12] Um invoice agents is combining the audio [25:14] models with knowledge [25:15] uh and bringing that inside of the the [25:17] the system, how you bring it outside [25:19] with telephony systems so you can [25:20] interact across channels, how you [25:22] evaluate, test, and monitor. Um and then [25:25] as you create, whether that's in the [25:26] agent space, whether that's in the [25:27] creative space, that's some [25:28] understanding, you build the ecosystem. [25:30] And that's what we hope to build across [25:31] 11 labs, a place where whether that's [25:34] distribution and brand that people can [25:36] trust, the platform where you have [25:38] pre-existing set of work that you can [25:40] start off, whether it's a template for [25:42] creating an agent, template for creating [25:44] uh a workflow in a creative space, or [25:46] whether that's a voice. And we had a [25:48] pleasure now of of having over [25:51] 20,000 voices that people created, [25:52] contributed, that you can you can use [25:54] across language styles and voices. And I [25:56] think that will be an increasingly [25:57] important layer of how you are able to [26:00] cater to that diversity, make it easy [26:01] for people to start, and really [26:03] understand that that workflow. [26:06] All right, I'm going to hand it back to [26:07] Konstantin Mali. Thank you. Andrew, [26:09] thanks for being partner. Yes. [26:10] >> [applause] [26:11] >> Amazing. [cheering] Thank you, guys.