11-19 更新 实时
首页/ 科技公司/ 正文

Belfort 的 AI 语音初创 Tone 想做华尔街之狼式销售

科技公司 来源:TechCrunch AI 2026-11-17 01:10 0 阅读 机器翻译
Belfort 的 AI 语音初创 Tone 想做华尔街之狼式销售

多年前,随着围绕 WeWork 估值飙升的质疑越来越多,我与联合创始人亚当·诺伊曼坐下来交谈。我们当时正在商谈一场舞台采访,而在采访前的几周里,他要求与我会面两次,一次在旧金山的一家 WeWork,另一次在他纽约的办公室。

我完全清楚,尤其是 SoftBank 给 WeWork 定下的那些数字根本说不通,但离开时我也明白了,为什么投资者愿意与他共进退。

同样容易理解的是,为什么人们会被自称“华尔街之狼”的乔丹·贝尔福特迷住。贝尔福特在 1999 年对与 Stratton Oakmont 相关的证券欺诈和洗钱指控认罪,此事广为人知;这家长岛经纪公司的销售人员以拉高出货的手法推销仙股,而贝尔福特则购置了南安普顿豪宅、167 英尺游艇和至少一架直升机。(他在服用 Quaaludes 和可卡因后神志亢奋时毁掉了后两样,正如马丁·斯科塞斯对贝尔福特自传的著名改编电影《华尔街之狼》所描绘的那样。)

你可能会认为,一个在监狱里待了 22 个月、并被勒令向 1,513 名受害者支付 1.104 亿美元赔偿金的人,不会再有第二次机会。TMZ 在 7 月报道称,受害者最终收到的赔偿“远少于”他们应得的金额,并指出贝尔福特在今年 4 月底完成了法院勒令的赔偿义务。

但贝尔福特身上有一种能激发人的能量,他不断重新出现,包括近年来作为驻迈阿密的销售顾问,以及曾在他位于迈阿密海滩的家中主持加密货币研讨会。(《纽约时报》的一名记者几年前参加过一个周末研讨会,当时的入场费是 1 枚比特币,或按当时价值计算的 40,000 美元等值金额。)

如今,他正切入自己最擅长之事——销售——与全球最热门趋势——AI——的交汇点,打造一个基于语音的 AI 平台。他说,其秘诀在于为语音注入那种能促成交易的语调,比如把陈述句说成问句(这"能提升互动"),有时甚至低语(这"暗示信息稀缺")。

Belfort 表示,这个名为 Tone 的平台已被一些公司使用,这些公司的销售线索多到其人工团队无法处理。如今,Tone 正推出一项自助服务平台,面向希望把自己的书变成"富有表现力的 AI 生成旁白"的内容创作者和作者——此举也可能有助于扩大其训练数据。用户可以从 Tone 的 AI 语音中进行选择,或者随着语音克隆功能的推出,创建自己声音的克隆。

很容易说,市场已经充斥着 AI 语音初创公司和技术,而 Tone 还带有额外风险:Belfort 在这项业务中的联合创始人是资深投资人 Shervin Pishevar,他本人也有一段争议过往。我们也无法独立核实这项技术是否如描述般有效(两位联合创始人都不是工程师)。

但如果说我与 Neumann 打交道的经历教会了我什么,那就是:对数字的怀疑很少会妨碍一场精彩的推介。当 Belfort 为这项新事业寻找投资人支持时,找到信徒不会太难。

以下是我们对话的部分内容,为篇幅和清晰度做了编辑。你打造了什么?

这是一种非常独特的打造声音的方式,它源于我从 20 世纪 80 年代末起一直在教授的一套影响力与说服体系。关键不只是你说什么,而是你怎么说。当你不是面对面时,肢体语言会坍缩为语调。人们日常使用的、极具力量的语调大约有 15 种。我为此发明的说法是「目标导向沟通」。如果我想被看作专家,我就会策略性地运用语调。

这套系统实质上是拿你本人训练出来的吗?

一开始是这样的。作为非工程师,我从第一性原理出发:我会怎么教一个人做这件事?一次教一种语调。我有 30 年超高质量的数据,于是我开始教这个模型。那是一个突破性的时刻,因为它不仅听起来和我一模一样,还精准掌握了那种语调。接着我教它第二种语调,然后教它如何把两者衔接起来。这一切的关键是数据。哪怕只有一点点脏数据,也会大幅抵消你已经取得的进展。

你和 Shervin Pishevar 是怎么走到一起的?

我和 Shervin 是多年的私交好友,有天晚上我们一起吃晚饭。此前有家 AI 公司向我推销,说他们有能做销售的 AI。我听了一下,效果糟透了。但这让我脑子里有了这个想法。我原以为我只是要教模型怎么做销售,但我找不到一个足够好的声音。需求是发明之母,所以我们自己造了一个。

你们的客户是谁?他们如何使用它?

我们有律师事务所和健康保险公司客户。我们在民调和汽车领域也做了很多。它会带你完成整个健康保险销售流程。它会问一些资格筛选问题,比如你想要哪种保险、你的预算、你的年龄。然后它会选择一个计划,并解释其特点和优势。我们的延迟大约是600毫秒。如果付款金额太高,它会说可以降低付款,并解释如何提高你的免赔额。如果你不喜欢这个计划,它会给你换一个。它确实完成了整个销售流程。

AI 如何处理偏离脚本的销售通话?

它从不说“这是个好问题”,除非这问题确实很好。我讨厌那种说法。太讨好了。任何销售都有一个理想路径,即我们希望完美对话如何进行,每一个字都可以预先规划好。但人们会有问题和顾虑,会把你带离直线。所以诀窍在于,模型既能进行那段完美对话,又能应对抛给它的任何问题,优雅地回答并回到直线上,同时不破坏融洽关系,也不产生幻觉——或者,用销售员的话说,不撒谎。我们已经训练模型做到这一点。

你们从 Tone 上获得任何收入了吗?你们为它筹集过资金吗?

收入大概一百万美元——也许70万美元。我们完成了亲友轮和种子轮融资,约1000万美元。

我想谈谈房间里的大象。你有欺诈定罪记录。Pishevar 曾面临不当行为指控,但他否认这些指控。为什么银行、医院或消费品牌应该信任你们两人推出的这款产品?

我能为自己说的就是:我确实犯过错误。那是很久很久以前的事了,而过去 20 多年我一直在给世界各地的人授课。世界各地的公司聘请我去教销售和说服技巧。如果有些公司不想与我们合作,那是他们的决定。但我认为,会有足够多的公司看到我们创造的东西的价值。

与典型的销售团队相比,Tone 的成交率如何?

我仍然觉得它还不如顶尖的销售高手。该领域的顶尖高手在逐通电话的比拼中都会胜过它。我认为这会在未来六个月到一年内改变。要让模型响应得这么快,你就不能用前沿模型。我们用的是自己的非常小、非常轻量的模型,所以它们的智能没那么高。我认为它会明显胜过平庸的成交人员。

它在哪里使用最有效?低金额的销售机会?我倒是觉得,一个人花的钱越多,就越会要求更多的人与人接触。

我不认为这关乎取代人类。我在客户那里看到最多的情况是,他们根本没有可用的人力。他们能获得的线索远超有足够人手去打电话的数量,或者他们需要在一天中的不同时段、用不同语言进行覆盖。我认为永远都会需要优秀的真人销售。我真的这么相信。

你在与 ElevenLabs 和 Cartesia 以及许多其他语音 AI 公司竞争。Tone 是你的优势所在吗?

我不认为听起来像人真的是这件事的关键。六个月或一年后,它们都会听起来像人。让我们与众不同的是语调的应用:在何时选择某种语调背后的智能。我们为此拥有专利。ElevenLabs 是一家很棒的公司,Cartesia 也是,但那些声音对我们不适用。

FCC 在 2024 年裁定,AI 语音受《电话消费者保护法》下的机器人电话规则约束。外呼销售电话是如何运作的?

我们只呼叫已选择加入(opt-in)的对象。你不会用 AI 给人打陌生推销电话。我们不会允许那样做。否则你会被全世界每一个 TCPA 律师起诉。要么是已选择加入的潜在客户,要么是现有客户。例外是民意调查,其中有一项豁免允许你在未经选择加入的情况下致电私人住宅。

你们如何验证同意,并在声音克隆中防止身份冒用?

我们本可以明天就推出克隆功能。我们没有。如果你在我们的系统中创建一个克隆,你要对着电脑说话,所以我们知道是谁在创建它,并且那个人要签署同意书。如果有人想上传音频——我们不推荐这样做——你必须实时匹配一个现场声纹。

原文较长,译文为节选;译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(Jordan Belfort’s new AI voice startup, Tone, wants to sell like the ‘Wolf of Wall Street’)

Years ago, as questions mounted about WeWork’s soaring valuation, I sat down with co-founder Adam Neumann. We were negotiating a stage interview, and he asked to meet me twice in the weeks leading up to it, once at a WeWork in San Francisco and again at his office in New York.

I knew full well that the numbers that SoftBank in particular had assigned to WeWork made little sense, but I came away understanding why investors wanted to throw in their lot with him.

It’s similarly easy to understand why people are enchanted by Jordan Belfort, the self-proclaimed “Wolf of Wall Street.” Belfort famously pleaded guilty in 1999 to securities fraud and money laundering tied to Stratton Oakmont, the Long Island brokerage where salespeople pushed penny stocks in pump-and-dump schemes while Belfort acquired a Southampton mansion, a 167-foot yacht, and at least one helicopter. (He destroyed the latter two while high on Quaaludes and cocaine, as depicted in Martin Scorsese’s famous adaptation of Belfort’s autobiography, “The Wolf of Wall Street.”)

You might think someone who spent 22 months in prison and was ordered to pay $110.4 million in restitution to 1,513 victims wouldn’t get another shake. The victims ultimately received “far less” than they were owed, TMZ reported in July, noting that Belfort completed his court-ordered restitution obligations in late April of this year.

But Belfort, whose energy has a galvanizing quality, keeps resurfacing, including in recent years as a Miami-based sales consultant and one-time host of crypto workshops at his Miami Beach home. (A New York Times reporter attended a weekend workshop several years ago where the price of admission was one bitcoin or its equivalent value at the time of $40,000.)

Now, he is moving into the intersection of what he knows best — sales — and the hottest trend on the planet, AI, building out a voice-based AI platform. The secret sauce is that it imbues its voices with the kind of tonality that wins deals, he says, for example, by phrasing declarative statements as questions (which “builds engagement”) and even whispering at times (which “implies scarcity of information”).

Belfort says the platform, called Tone, is already being used by companies that have more sales leads than their human teams can handle. Now, in a move that could also help expand its training data, Tone is launching a self-service platform for content creators and authors who want to turn their books into “expressive AI-generated narration.” Users can select from Tone’s AI voices or, as a voice-cloning feature rolls out, create a clone of their own voice.

It’s easy to argue the market is already saturated with AI voice startups and technologies, and Tone comes with added risk: Belfort’s co-founder in the business is longtime investor Shervin Pishevar, who has a controversial past of his own. We also can’t independently verify that the technology works as described (neither co-founder is an engineer).

But if my time with Neumann taught me anything, it’s that doubts about the numbers rarely get in the way of a good pitch. When Belfort looks for investors to back him in this new venture, finding believers won’t be too hard.

Here is part of our conversation, edited for length and clarity.

What have you built?

It’s a very unique way of going about building a voice, and it stems from a system of influence and persuasion I’ve been teaching since the late 1980s. It’s not just what you say, but how you say it. When you’re not in person, body language collapses into tonality. There are about 15 very powerful tonalities that people use all the time. The phrase I invented for this was “goal-oriented communication.” If I’m trying to be perceived as an expert, I apply tonality strategically.

Is this effectively trained on you?

It started that way. As a non-engineer, I went to first principles: How would I teach a human to do this? One tonality at a time. I have 30 years of super high-quality data , and I started teaching the model. It was a breakthrough moment, because it not only sounded exactly like me, it had nailed the tonality. Then I taught it a second one, and then how to link them. The key to all of this is data. Even the littlest bit of corrupt data dramatically negates the progress you’ve made.

How did you and Shervin Pishevar come together?

Shervin and I have been personal friends for a long time, and we were out for dinner one night. I had been pitched by an AI company that said it had an AI that could do selling. I listened to it. It was awful. But it put the idea in my head. I thought I was just going to teach models how to sell, but I couldn’t find a voice that was good enough. Necessity is the mother of invention, so we built our own.

Who are your customers, and how do they use it?

We have law firms and health insurance companies. We do a lot in polling and automobiles. It will take you through an entire health insurance sale. It asks qualifying questions, like what type of insurance you want, your budget, your ages. Then it chooses a plan and explains the features and benefits. Our latency is like 600 milliseconds. If the payment is too high, it’ll say it can bring the payment down and explain how to raise your deductible. If you don’t like the plan, it swaps you to a different one. It does literally the whole entire sale.

How does the AI handle a sales call that goes off script?

It never says “That’s a great question,” unless it really is. I hate that. It’s so ingratiating. In any sale there’s a happy path, how we’d want the perfect conversation to go, and every last word can be mapped out. But people have questions and concerns, and they take you off the straight line. So the trick is a model that can carry that perfect conversation, handle anything thrown at it, answer elegantly and return to the straight line, without breaking rapport and without hallucinating — or, as a salesperson, lying. We’ve trained the model to do that.

Have you generated any revenue from Tone yet? Have you raised funding for it?

Revenue, probably about a million bucks — $700,000, maybe. We’ve raised friends and family and seed, about $10 million.

I want to address the elephant in the room. You have a fraud conviction. Pishevar has faced misconduct allegations that he denies. Why should a bank or a hospital or a consumer brand trust this product coming from the two of you?

All I can say for myself is that I did make mistakes. It was a very, very long time ago, and I’ve spent the last 20-plus years teaching people all over the world. I’ve been hired by companies all over the world to teach sales and persuasion. If some companies don’t want to work with us, that’s their decision. But I think there would be more than enough that would see the value of what we created.

What close rate does Tone have versus a typical sales force?

I still don’t think it’s quite as good as a top producer. A top producer in the field would outdo it call for call. I think that will change in the next six months to a year. To make a model respond this quickly, you’re not using frontier models. We’re using our own very small, very light models, so they don’t have quite as much intelligence. I think it will significantly outperform a mediocre closer.

Where is it used most effectively? Low-dollar sales opportunities?I’d think the more money someone is spending, the more human contact they’d demand.

I don’t think it’s about replacing humans. What I see most with customers is that they simply don’t have the human capital available. They could get a lot more leads than they have people to call, or they need coverage at different times of day or in different languages. I think there’s always going to be a need for great human salespeople. I really believe that.

You’re competing with ElevenLabs and Cartesia and a lot of other voice AI companies. Tone is your edge?

I don’t think sounding human is really what this is about. They’ll all sound human in six months or a year. It’s the application of tonality that makes us different: the intelligence behind when it chooses a certain tonality. We have a patent on it. ElevenLabs is a great company, and Cartesia as well, but those voices didn’t work for us.

The FCC ruled in 2024 that AI voices fall under robocall rules under the Telephone Consumer Protection Act. How does outbound sales calling work?

We only call opt-in. You’re not going to use AI to cold call people. We would not allow that. You’d get sued by every TCPA lawyer in the world. It’s either opted-in leads or existing customers. The exception is polling, where there’s an exemption that lets you call private homes without opt-in.

How do you verify consent and prevent impersonation in voice cloning?

We could launch the cloning tomorrow. We didn’t. If you create a clone in our system, you speak into the computer, so we know who is creating it, and that person signs a consent. If someone wants to upload audio, which we don’t recommend, you’d have to match a live voiceprint in real time.

Who created the technology? You’re not an engineer, and Pishevar isn’t, either. I saw a name on the patent, Jordan Anthony Metzner.

No. Our VP of engineering is Ben Sussman. He’s a brilliant engineer. I’m the judge and the jury and the executioner of what sounds perfect. I’ll prototype something in the most obscene way, and my engineers rebuild it from scratch.

Do you charge per minute, or based on the outcomes customers see?

A combination. For streaming, it’s per minute. The content platform will be monthly subscriptions. And we have some clients where the bulk of our money is made through success.

Is this an independent company or something that belongs inside a larger AI company?

If the right offer came up, and it was substantial enough, and I thought it made sense, who wouldn’t entertain a conversation? But right now my plan is to raise a significant amount. I don’t think we need to own data centers, but we need enough reserve compute. I wouldn’t think I’d be selling it anytime soon, but who knows what the future brings.

Paramount+ released a documentary about you this summer. Did you watch it?

I didn’t. Believe it or not, I don’t watch or read anything about me, good or bad. I never read a comment. When you’re well-known, you have to live by who you are inside. I go by my own internal compass.

When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

Connie Loizos

Editor in Chief & General Manager

Loizos has been reporting on Silicon Valley since the late ’90s, when she joined the original Red Herring magazine. Previously the Silicon Valley Editor of TechCrunch, she was named Editor in Chief and General Manager of TechCrunch in September 2023. She’s also the founder of StrictlyVC, a daily e-newsletter and lecture series acquired by Yahoo in August 2023 and now operated as a sub brand of TechCrunch.

You can contact or verify outreach from Connie by emailing [email protected] or [email protected], or via encrypted message at ConnieLoizos.53 on Signal.

Loading the next article

原文为英文,正文由自动抓取并保留原文。

本文由采集器自动同步

原文发布于 TechCrunch AI(Jordan Belfort’s new AI voice startup, Tone, wants to sell like the ‘Wolf of Wall Street’),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗