AI Voice Synthesis Goes Enterprise as ElevenLabs Hits $22B

Voice stops being a demo and starts being a business

For most of the generative AI boom, synthetic voice was treated as a feature. It narrated YouTube videos and dubbed trailers, and every so often it showed up in a deepfake scandal. The events of the past two weeks suggest that period is over. ElevenLabs, the London- and New York-based voice company, completed a $300 million employee tender offer at a $22 billion valuation. That is double the $11 billion it was valued at in its February Series D.36 Around the same time it released Eleven v4 and v4 Turbo, text-to-speech models that the company says can speak, listen and translate across more than 90 languages.8 Over the same stretch, Apple disclosed a deal to hire staff from personalized-audio startup Huxe and license its technology, and Automation Anywhere agreed to buy conversational-voice vendor Boost.ai.1112

Taken together, these events show that the money in voice AI no longer comes from making voices sound human. It comes from voices that finish tasks inside companies. The wider push toward multimodal models, where video generators now produce synchronized dialogue in a single pass, points the same way.

The ElevenLabs numbers, and what they actually measure

The outlets covering the tender offer report the core facts the same way. Wellington Management and T. Rowe Price co-led the transaction. Goldman Sachs, GIC, EQT, Ontario Teachers' Pension Plan, Sapphire Ventures and BDT & MSD joined as new investors, and Andreessen Horowitz, ICONIQ, Lightspeed and others took part as existing backers.28 This is the company's second secondary sale for staff. The first was a $100 million tender at a $6.6 billion valuation in September 2025.6

The valuation needs some care. A tender offer lets employees sell vested shares. It does not put new money on the company's balance sheet. One commentary called it a retention tool in the competition for voice-AI researchers.5 Another outlet described it as part of a broader pattern of AI startups offering staff liquidity so they don't leave for rivals.6 That reading looks right. The headline number reflects what late-stage institutional buyers will pay for a slice of the company. It is not a priced primary round.

The operating figures may matter more. ElevenLabs says enterprise customers now bring in 55% of its revenue.27 Its ElevenAgents platform handles more than 15 million conversations a week, roughly triple the February level, and the platform's annual recurring revenue has more than tripled over the same period.27 The company also cites internal analysis claiming its voice agents resolve issues 31% faster on average than chat-based alternatives.2 That figure comes from the vendor and has not been independently verified.

Coverage differs mainly in which customers it highlights. Fintech and insurance outlets stress that half of the world's ten largest insurers use the technology, and name Stripe, Customers Bank and Admiral as clients.210 Seoul Economic Daily says five of the ten largest technology companies and four telecom carriers are customers, and lists Deutsche Telekom and DoorDash.7 Another account adds the governments of Ukraine and Greece.5 Timing is also reported unevenly. Most outlets dated the announcement October 1, while the Korean report says it came on the 13th.37 The underlying facts hold up either way.

India as a test of multilingual voice at scale

The most telling detail may have come from Bengaluru. At its first India Summit on October 6, ElevenLabs said India is now its second-largest market after the United States. Indian businesses generated 100 million ElevenAgents conversations over the past year, and more than 70% of them were in Hindi, Kannada, Tamil and Telugu.4 Eleven v4 supports 14 Indian languages, and the company plans to cover all 22 official ones.4 The deployments are already sizeable. CARS24 reportedly runs more than 400 voice agents that have automated three million minutes of calls, and Razorpay has handled more than 35,000 Hinglish merchant calls through the platform.4

This matters because it tests whether voice synthesis can serve people who were never well served by text-first interfaces. Code-switching between English and a regional language is hard for a speech model. If that works commercially, ElevenLabs has a stronger claim to being the default voice layer than any benchmark could give it.

The product strategy: from voice model to platform

The company's release record shows a firm building well beyond text-to-speech. Over 2026 it shipped Music v2 and v2.5, a Dubbing v2 that keeps the original speaker's delivery across more than 90 languages, and a medical transcription model.9 In September it launched Studio 4.0, an editor that covers video, images, speech, music and sound effects, followed by Image, Video and Template APIs for developers.9 Recent SDK notes show image generation inside its platform running on OpenAI's GPT Image 2.5 models.1 They also show agents that can use third-party language models from OpenAI, Anthropic, Google and DeepSeek.1

The picture is clear. ElevenLabs sees voice as its way into broader multimodal production and enterprise workflow tools, and it plans to own the orchestration layer while using other labs' models where that makes sense. The company has said it wants to connect agents to internal systems so they can handle bookings, contract changes and customer service, not just reply out loud.7

Consolidation and competition from the platforms

The deal activity points to a market that is consolidating. Apple's Huxe arrangement surfaced through the EU's Digital Markets Act disclosure database. It involved job offers to selected staff and a nonexclusive IP license, not an outright purchase, and came after Huxe had already shut down its consumer app in May.11 Apple has not tied the deal to Siri or any specific product.11 Its value seems to lie in expertise in turning personal data into conversational audio briefings, an area where Google's NotebookLM Audio Overviews already compete.11 Earlier this year Google DeepMind acqui-hired Hume, and OVHcloud bought speech company Gladia.14

The big labs are also cutting prices. Google's Gemini 3.8 Flash TTS, released September 22, is priced at $9 per million audio tokens on the paid tier, below OpenAI's gpt-4o-mini-tts at $12.19 OpenAI's full-duplex GPT-Live-1 costs $0.05 a minute, and Gemini 3.8 Live charges fractions of a cent per minute.19 For a specialist like ElevenLabs, the defense is not model quality alone. It is the deployment tooling, language coverage and enterprise relationships that the hyperscalers have been slower to build.

Investors are also funding the defensive side. Modulate raised $25 million for a suite that covers deepfake detection and policy enforcement for voice agents, on the reasoning that cloning has become easy.20 Resemble AI's pricing page now leads with detection plans instead of text-to-speech.17

Video generators are absorbing voice too

The same trend is showing up in AI video. In 2026, native audio has replaced the old routine of generating a silent clip and adding sound afterward.30 Alibaba's Wan 3.0 produces 30-second clips with dialogue, music and effects in a single pass. ByteDance's Seedance 2.5 matches that length.22 MiniMax's open-weight H3 handles dialogue in 11 languages with stereo audio.22 xAI's Grok Imagine Video 1.5 is promoted partly for its voice cloning.21 ShengShu launched a Vidu Q4 preview on October 7 aimed at expressive audio and video, starting at $0.014 per second.24

Rankings of who leads depend on when and where you look. Several roundups put Wan 3.0 first on the Artificial Analysis text-to-video leaderboard, both with and without audio.2122 Others name Google's Gemini Omni Flash as the arena leader.2529 Leaderboards change week to week, and the conflicting rankings show how close the frontier is. They also show that nobody holds a lasting lead in text-to-video. OpenAI has stepped back, closing the Sora app in April and listing September 24 as the API's shutdown date.23

Regulation and labor set the limits

The commercial expansion is happening under tighter rules. Under Article 50 of the EU AI Act, machine-readable marking of AI-generated audio has applied since August 2, 2026. Generators already on the market have until December 2 to comply, and fines can reach €15 million or 3% of global turnover.17 In games, SAG-AFTRA's contract requires written consent for digital replicas of a performer's voice and lets members suspend that consent during strikes.13

The reading

The usual framing is that voice AI is the next interface. It's more accurate to say voice is becoming plumbing: a commodity layer that labs price aggressively and that video models now generate as a by-product. The companies that do well will be the ones that turn synthesis into finished work, such as a renewed policy or a merchant call handled in Hinglish. ElevenLabs' valuation is a bet that it can be one of them. Whether that bet holds depends less on Eleven v4's sound quality than on whether its 15 million weekly conversations keep growing once Google and OpenAI sell the underlying speech at a fraction of the price.