Voices Built a Marketplace for Humans. Now It Has to Make AI Pay Them.

The decisive moment in the history of Voices was not an audition. It was a phone call. In the early 2000s, audio engineer David Ciccarelli and vocalist and marketer Stephanie Ciccarelli had built a plain online directory of performers around their London, Ontario recording work. Actors asked to be listed. Then a visitor called and asked the useful question: how do I hire one of them? A list of profiles had quietly become a marketplace.

Two decades later, Voices is still answering that question, only the pronouns have become slippery. A buyer may want a human narrator for an e-learning course, 28,000 video clips for a data project, or one actor whose performance can be cloned into a car, a chatbot or five localized ads. Who is being hired - the performer, the recording, the reusable model, or the rights to all three? Voices now makes money by turning that ambiguity into a managed process.

The company says its network includes more than 4 million professionals in 185 countries, that it has completed more than 500,000 projects and that more than 75 Fortune 100 companies trust it. Those figures describe supply and scale. The more revealing business is the machinery between them: briefs, auditions, matching, contracts, protected payments, directed recording, quality assurance and delivery. The product is not merely a nice voice. It is less uncertainty per voice.

Act IThe marketplace found the transaction

The early playbook was beautifully ordinary. Businesses posted a project. Actors uploaded demos and auditioned. Voices used its VoiceMatch system to recommend suitable talent, held project money through SurePay and kept files and communication in one place. Small buyers avoided a slow circuit of agents and demo reels. Freelancers found clients beyond their geography. Producers could hear dozens of interpretations before choosing one.

The founders first charged talent subscriptions for visibility. The model later added transaction fees and managed production for larger jobs. Today it remains a hybrid: a self-serve marketplace for buyers who know what they need, paid memberships and fees on the talent side, payment processing, and negotiated professional-service charges for complicated enterprise work. Public pricing can explain a card fee; it cannot price a five-year exclusive AI voice license for a global product. That requires a conversation.

One bold early expense helped the marketplace look inevitable before it was. The company began as Interactive Voices, a name that sounded like an instruction manual and confined the business to “interactive” media. The owners of Voices.com wanted $50,000. The Ciccarellis countered at $30,000, paid in six $5,000 installments. David has said the first installment came from a credit-card cash advance; before later payments, he paused Google ads and banked the spend. Annual revenue was still under $100,000.

It worked because the asset did three jobs at once: it shortened the pitch, widened the category and improved discovery. David later said traffic doubled around the move and credited the domain with a 20-fold growth period. The copyable lesson is not “buy an expensive dot-com.” It is to pay for a bottleneck that compresses customer understanding. The same gamble would fail with weak retention, no organic demand or a name that merely looks prestigious.

The less cinematic failure was managerial. David has named his early reluctance to confront underperformance as a mistake; when hard conversations were postponed, poor performance became the accepted standard. Years later, BDC advisers found a company growing at the expense of profitability, with leadership and financing gaps. The practical fixes included restructuring professional services so the team spent less effort on large jobs before a customer committed, clarifying strategy and preparing leadership for institutional due diligence.

That changed the financing logic. The founders had preferred debt - successive loans let them preserve ownership - including C$2 million from BDC Capital in 2015. But expansion required a larger balance sheet. In July 2017, Morgan Stanley Expansion Capital invested US$18 million for a minority stake. A month later, Voices bought Voicebank.net, a casting workflow used by agents and union talent. The deal broadened access, but it also drew public concern from SAG-AFTRA members, agents and casting directors wary of consolidation and the buyer's mostly non-union marketplace. Scale had acquired politics.

Act IIAI made the contract part of the product

Generative voice could have made a casting marketplace less relevant. If one synthetic speaker can read every script instantly, why search four million profiles? Voices' answer is to move up the stack. Its Branded AI Voice service starts with brand calibration and casting, then handles consent, exclusivity, recording, usage rights and governance. The customer can bring a preferred cloning model. Voices supplies the person, performance and paper trail.

BMW shows why this is more than compliance theater. For next-generation in-car assistants, its team wanted voices that felt natural, supportive and premium without sounding smug. Voices supplied a shortlist, the team narrowed it to three and listened inside a real car. It also had to consider that the same gender and tone can land differently in Saudi Arabia and Canada. The work was casting, interface design, localization and risk management folded into one purchase.

Cresta offers another condition: a stock text-to-speech voice may sound fine alone and flatten under the emotional work of a live support call. Voices found an Australian actor and captured source audio designed for cloning, with a bounded use case. The goal was not an all-purpose digital double. It was a voice able to move from upbeat to empathetic and back during one customer conversation.

The SuperBloom case is more revealing because parts of the technology did not work. The agency wanted one English performance to anchor a Deel campaign in five additional languages. The actor was paid the equivalent of five localized bookings, the clone was restricted to the agency's account, and native producers joined sessions. Some accents were still too nuanced for the model. The fallback was refreshingly un-disruptive: cast native human voices. The team concluded that 30 minutes of clean audio could produce a result, but roughly two hours was a better target.

This is where Voices sits in the market. Voice123 and Bodalgo compete for online casting. Talent agencies offer curation and relationships. Fiverr and Upwork sell broad freelance access. ElevenLabs, Resemble, WellSaid, Murf and cloud platforms sell generated speech. Voices occupies the middle: more managed than a listing site, more human than a stock voice library and less attached to one model than a generation vendor. The distinction matters most when a voice becomes brand property rather than a disposable file.

The stealBuild the boring chain before the magical demo

The reusable playbook is a sequence. First, define the actual context: a luxury cabin, a frustrated support call, a game character or a six-language ad. Second, audition for that context rather than abstract pleasantness. Third, write down geography, channels, duration, exclusivity, retraining and prohibited uses. Fourth, capture more clean, directed audio than the minimum. Fifth, test where the voice will live. Finally, keep the performer available and keep a human fallback when the model cannot carry cultural nuance.

What did it cost? The famous early answer is $30,000 for the domain, paid painfully. The growth answer is C$2 million in BDC working capital and US$18 million from Morgan Stanley. The AI answer is not public because every engagement bundles different talent, territories, recording and rights. SuperBloom gives a better unit of thought than a price sheet: if a clone replaces five localized bookings, compensate with those five bookings in mind.

That final condition is the uncomfortable one. Voices has faced years of actor criticism over memberships, marketplace fees, managed-service opacity and the fear that AI jobs can turn performances into cheap, permanent substitutes. Its consent-and-compensation standard is therefore not a decorative ethics page. It is a repair mechanism for the supply side of the marketplace. If actors believe the contracts are fair, Voices gains scarce, licensable performances. If they do not, the four-million-person number becomes a noisy database rather than a moat.

The company is now pushing into voice data, customer experience and games, where character rights may persist across sequels and synthetic dialogue can multiply. That expansion makes governance more valuable and mistakes more expensive. A bad narration can be rerecorded. A poorly licensed franchise voice can become litigation, reputational damage and a character that suddenly needs a new throat.

Voices began by making humans searchable. Its next business is making their digital echoes accountable. The insight worth copying is modest: when technology makes creation abundant, build around the new scarcity. In AI voice, that scarcity is not sound. It is a performance people remember, permission people can prove and a relationship that survives the next line of dialogue.

Keep listeningVoices around the web