The world of voice AI, with Mati Staniszewski of ElevenLabs
Summary
ElevenLabs reported $350 million of ARR at the end of 2025 and then added $100 million of net-new ARR in one quarter, putting John Collison’s rough arithmetic in the $450 million range. More than 50% of the business is now sales-led enterprise, with attractive entry pricing driving a “land and expand” motion from one workload into multiple departments.
The technical moat is a combination of model architecture and unusually rich data about how speech is delivered, not merely what was said. ElevenLabs built specialist labelling operations because public audio rarely identifies speakers, accents, emotions or prosody correctly; its internal transcription work subsequently became a customer product. The result is a model where even “Britishness is an emergent property,” rather than a hard-coded switch.
Voice AI’s near-term bottleneck is deployment and orchestration, not raw vocal quality. Staniszewski argued that production-quality asynchronous narration emerged three years ago, real-time speech two years ago and connected, context-aware applications only recently. Customer support can already pass a scoped “voice Turing test,” but open-ended gaming and general assistants remain far harder; ElevenLabs hopes the broader conversational-agent problem improves “in the next year or so.”
ElevenLabs is betting on cascaded agents—speech-to-text, LLM, then text-to-speech—because enterprises value reliability, visibility and control more than minimum latency. Direct speech-to-speech models are quicker but Staniszewski called them “definitely dumber”; they also lose visibility into intermediate pipeline steps and can be less reliable. Staniszewski expects a hybrid future: direct models for low-complexity interaction and cascades for higher-complexity actions, authentication and database retrieval.
The product opportunity extends well beyond synthetic voices into the full interaction layer between businesses and customers. ElevenLabs wants to supply models, knowledge connections, telephony, integrations, monitoring and safeguards across support, sales and marketing, while leaving deeply domain-specific applications to vertical vendors. Fixing voice also supplies the text layer, but complex multitouch reasoning and financial analysis are explicitly outside the company’s optimization target.
Voice models have structurally lighter training economics than frontier LLMs, but distribution and inference still matter. Leading voice models are “a few billion to tens of billion” parameters versus hundreds of billions for leading LLMs; customers pay mainly per token or minute, with annual commitments and volume discounts. ElevenLabs often offers a new, costlier model near cost so customers reveal its use cases and failure modes—turning adoption into both distribution and research feedback.
The next product frontier is individualized understanding and controllable delivery. ElevenLabs expects person-specific transcription in a forthcoming version, “hopefully in the next month,” supplementing diarization, noise reduction and keyword dictionaries; V3 already adds instructions such as slowing down or inserting dramatic pauses. Its expressive mode can also detect stress and generate a reassuring response in both language and tone.
ElevenLabs’ operating model is as important to the growth thesis as its models: 470 employees, product and research initiatives generally below 10 people and more than 15 direct reports for each founder. Nontechnical functions receive technical leads who automate workflows, while self-serve access exposes the newest technology to developers before it is enterprise-ready. Staniszewski’s cultural filter is “agency”; Collison’s sharper formulation was that high-agency people win from AI while low-agency people inside organizations lose out.
Deep dive
1. Speech became a prediction problem, but context made it human
Staniszewski traced synthesis from Wolfgang von Kempelen’s mechanical vocal tract through Bell Labs’ structured signals and libraries that concatenated phonemes—the smaller human sounds beneath syllables—to neural systems predicting the next sound from preceding audio and text.
A conventional pipeline moved from text to a mel spectrogram, a visual representation of pitch and energy, and then into waveform audio. Co-founder Piotr’s architectural contribution was to abstract and improve those encoding and decoding stages while bringing transformer and diffusion ideas into speech.
The model receives text plus a speaker embedding: the reference that carries accent, style, prosody and other vocal characteristics. It must simultaneously infer how the sentence is constructed and how earlier and later context should alter its delivery—whether a line has a particular dialect, is happy or is part of a streaming exchange.
Earlier systems exposed hard-coded controls for pitch, energy or categories such as British and Polish. ElevenLabs instead lets the model discover those dimensions, including enthusiasm and sadness; Collison’s summary—“Britishness is an emergent property”—earned Staniszewski’s “Exactly.”
2. Proprietary annotation turned a research necessity into a product portfolio
Staniszewski reduced model-building to architecture, compute and data, then emphasized audio’s unusual data problem: recordings are abundant, but annotations normally capture the “what,” not the “how.” They may omit speaker identity and the emotion, accent and delivery choices that create humanness.
ElevenLabs therefore built its own voice-trained labelling team and combined manual work with semi-automatic techniques. Its speech-to-text model began as internal infrastructure because available systems could not annotate the training corpus well enough, then became a standalone customer product.
Today Staniszewski describes ElevenLabs as a “research and product deployment company”: foundational text-to-speech, speech-to-text across more than 100 languages, conversational and music models, plus the production layer connecting them to knowledge bases, telephony, actions, evaluation, monitoring and safeguards.
The boundary is horizontal capability rather than deep vertical specificity. ElevenLabs will serve common enterprise workflows directly, while specialized application companies own domain-heavy products; captioning can cover “99.9%” of ordinary cases, while healthcare customers may still require custom models and technical dictionaries can supply domain context.
3. Voice quality arrived before the products capable of using it
Collison’s challenge was experiential: his phone could not reliably read a PDF while he drove, car voice controls remained poor, and leading consumer assistants felt “10 years ago” despite capable underlying models. Staniszewski accepted a deployment gap but disputed how long the full technology had actually existed.
His timeline was compressed: good asynchronous narration arrived roughly three years ago, real-time generation two years ago, production readiness about a year ago, and context-connected interaction only over 2025. He expects cloud-connected automotive use cases “this year,” but on-device operation without connectivity is probably another two or three years away.
The PDF problem already has a direct answer in ElevenReader. ElevenLabs built it after authors used its 2023 software to generate affordable audiobooks but distributors such as Audible blocked AI content; the resulting app lets users upload PDFs or text and hear them in voices including Sir Michael Caine’s.
General conversation remains an orchestration problem: the system must infer when the user has finished based on silence and context, when to clarify and when to execute. Customer support can pass a domain-specific voice Turing test; a genuinely human-like interactive game cannot, and tool calls or authentication expose the remaining seams.
4. Personalization and controllability are the next accuracy gains
Collison argued that recognizing every utterance with a global model wastes stable information about a known speaker’s accent. Staniszewski agreed: ElevenLabs is combining person-specific recognition with diarization, noise reduction and keyword detection, letting a coffee shop or technical podcast constrain likely vocabulary.
Speaker fine-tuning is “solvable,” with rollout in a forthcoming version “hopefully in the next month.” The highest-stakes examples are an operating room that must isolate one doctor’s command and a home device that should hear its authorized user rather than a family talking nearby.
On generation, V3 changes the model from choosing a performance autonomously to following delivery cues such as “say it in a slower way” or alter a dramatic pause. Expressive mode can detect that the other person is stressed, produce reassuring language and render it with reassuring emotion.
ElevenLabs nevertheless favors cascaded agents: transcription, LLM reasoning and speech generation. Direct speech-to-speech removes text and lowers latency, but sacrifices observability and reliability; Staniszewski sees it flourishing in companions, where hallucinations may matter less and could even be a feature, while complex enterprise work remains cascaded.
5. Speaking changes what users disclose and who can participate
ElevenLabs replaced part of its lead form with a voice agent and observed both higher completion and richer disclosure. People described wider, more complex use cases because speaking was easier than writing, while follow-up questions made the interaction resemble Collison’s “open-ended adventure game.”
Multilingual support widened the funnel globally, and companies from banks to automotive businesses began building voice SDR variants. Staniszewski’s broader aspiration is the “Babel fish”: users speak Polish or English while listeners receive their own native language, collapsing a distribution barrier that poor dubbing once exposed.
The most consequential cases restore rather than automate communication. Staniszewski described people losing speech through ALS or throat cancer, then singled out a woman whose voice was recreated after she lost it before marrying: she could finally repeat the ceremony and speak her vows in her own voice.
Agents can be reactive or proactive. The Guinness-themed “Guinndex,” built with ElevenLabs, called Irish pubs to collect pint prices while also accepting inbound reports; roughly 3,000 entities could contribute, demonstrating voice as an interface to organizations that may lack structured APIs.
6. Smaller models improve the economics, while subsidized launches accelerate learning
Voice models are “significantly” cheaper to train than LLMs or image-video systems because they are smaller—roughly a few billion to tens of billions of parameters. CapEx remains meaningful, however: ElevenLabs had recently raised half a billion dollars at an $11 billion valuation to sustain model research and deployment.
Commercially, text-to-speech is generally charged by text token, while transcription and agents are billed per minute; large businesses sign annual agreements with larger commitments and discounts. Inference cost therefore scales with usage even though training runs are lighter than frontier-LLM runs.
A new model is more expensive and initially less reliable, yet ElevenLabs often offers it to customers near cost. The objective is broad exposure: customers discover what is possible, generate distribution and supply evidence about what must improve before enterprise reliability catches up.
Staniszewski thinks audiobook narration may already be near its useful model-size ceiling. Pure voice components should remain quick and relatively small, while fused systems combining intelligence and speech could reach tens or hundreds of billions of parameters; orchestration, rather than raw scale, remains central to cascaded systems.
7. Enterprise expansion now drives a historically fast revenue curve
ElevenLabs reported $350 million of ARR at the end of 2025 and its first quarter with $100 million of additional net-new ARR. Collison’s rough inference was “in the 450 million range”; Staniszewski did not formally restate that figure, instead pointing to customers including Deutsche Telekom, Revolut, Klarna, Meta and IBM.
More than 50% of the business is now sales-led enterprise. Expansion comes from both rising usage and departmental spread: Deutsche Telekom began with Magenta marketing and podcast generation, moved into support, and then expanded toward an agent serving interactions across the network.
Self-serve remains the discovery engine. Immediate usage tests model quality, exposes failures and lets developers and SMBs explore technology before it meets enterprise requirements for compliance, reliability and scale; high-touch deployed engineers then customize production systems alongside the largest customers.
ElevenLabs is also launching full pay-as-you-go self-serve billing, beyond subscription plans and overages. Collison argued that AI products should let users keep paying after a limit rather than stopping productive work with “Sorry, you’ve hit your usage limit.”
8. A flat, technical organization is ElevenLabs’ second experiment
At 470 employees, ElevenLabs still organizes product and research initiatives into teams generally smaller than 10, with smaller teams also used in parts of go-to-market. Each founder has more than 15 direct reports, roughly double a traditional span; Staniszewski admitted some of this is “definitely a little bit of a startup effect” whose durability will be clearer in five to 10 years.
Technical capability is embedded outside engineering. Talent and operations teams receive technical leads who make pipeline data explorable, automate manual work and identify candidate profiles; sales teams generate tailored pre-reads, case studies and decks, while humans retain responsibility for reviewing and developing the output.
Ukraine provided an external analogue: its Diia citizen platform connected benefits, frontline information, education and healthcare, with technical resources inside each ministry and a central transformation group assembling the services. Staniszewski called it “the most advanced set of work we’ve seen.”
Hiring therefore centers on culture and agency: first-principles thinking, ownership, excellence and humility. Staniszewski’s proudest outcome is that “culture builds the company now,” while Collison concluded that high-agency employees become AI’s winners because they continually explore how to amplify their craft.