The Future of Voice AI: Agents, Dubbing, and Real-Time Translation with ElevenLabs Co-Founder Mati Staniszewski

Mati reveals ElevenLabs has reached $300M ARR with 350 employees globally, split 50/50 between self-serve creative platform (5M monthly actives) and enterprise agents platform (thousands of customers including Fortune 500s). He addresses the counterintuitive nature of building both research and multi-segment products simultaneously, explaining how the initial insight came from Poland's poor dubbing experience.

• •Revenue split: 50% self-serve subscriptions/creators, 50% enterprise sales-led agents platform
• •Serves 5M+ monthly active users on creative side, thousands of enterprise customers including Fortune 500s
• •Started remote-first from Europe with hubs in London, NYC, Warsaw, SF, Tokyo, and Brazil
• •Original insight: Poland's single-voice dubbing for all movies created terrible experience, revealing global content localization opportunity
• •Counterintuitive strategy of combining foundational research with multiple product segments proved successful despite investor skepticism

Mati explains ElevenLabs' organizational structure of creating specialized 'labs' (voice lab, agent lab, music lab) that combine researchers, engineers, and operators around specific problems. He discusses how they sequence research first, then build product layers, and the critical decision framework of whether to wait for research breakthroughs or ship product improvements (3-month rule).

• •Organized into problem-specific labs: voice lab (narration), agent lab (interactive), music lab (licensed music creation)
• •Each lab combines ~5 researchers, engineers, and operators working on both research and product
• •Rule of thumb: if research breakthrough expected in <3 months, wait; if >3 months, build product solution now
• •Initial models from existing players were too robotic, requiring custom foundational model development
• •Research provides 6-12 month head start, but long-term value comes from product ecosystem, integrations, and workflows

Discussion of how voice selection is highly subjective and context-dependent, requiring ElevenLabs to employ 'voice sommeliers' who help enterprises choose appropriate voices for their brand and use cases. Mati reveals that benchmarking audio quality is fundamentally harder than text/image because voice preference varies dramatically by use case, language, and even customer demographics.

• •Employ voice coaches/sommeliers to help enterprises select voices matching their brand and customer base
• •Voice selection can be dynamic: Japanese/Korean company uses different voices for young vs elderly customers
• •Benchmarking challenge: switching voices between models can make lower-quality model seem better due to voice preference
• •Audio data labeling requires specialized skills to describe 'how' something was said, not just transcription
• •Future vision: fully personalized voice selection based on user context, time of day, and interaction patterns

Mati outlines the evolution of voice agents from reactive customer support to proactive assistants that enhance entire customer journeys. Examples include Meesho (India's largest e-commerce) using agents for product discovery and checkout, Square enabling voice ordering, and Epic Games bringing Darth Vader's voice to Fortnite for interactive gameplay.

• •Customer support is fastest-growing use case, working with Cisco, Twilio, Telus Digital
• •Shift from reactive (refunds, tracking) to proactive (product discovery, navigation, checkout assistance)
• •Meesho deployment: voice widget helps users find gifts, navigate catalog, complete purchases through conversation
• •Immersive media: Epic Games integrated Darth Vader voice in Fortnite for millions of interactive player experiences
• •Square partnership enables small businesses to deploy voice ordering and discovery experiences

Mati identifies education as the most exciting future application, with examples of Chess.com offering lessons from Hikaru Nakamura and Magnus Carlsen, and Masterclass enabling practice negotiations with FBI negotiator Chris Voss. He envisions a future where students have personalized AI tutors on-demand while maintaining dedicated human interaction time.

• •Chess.com partnership: learn from grandmasters like Hikaru Nakamura, Magnus Carlsen, Botez sisters through voice
• •Masterclass integration: practice FBI hostage negotiation tactics with Chris Voss interactively
• •Vision: personal tutor on headphones/speakers available on-demand, understanding individual learning patterns
• •Balanced approach: significant time with AI tutors, but explicit percentage of time reserved for human-to-human learning
• •Ukraine's Ministry of Transformation building 'agentic government' with education as key pillar alongside citizen services

Mati describes Ukraine's ambitious project to transform all government ministries using AI agents, combining customer support (benefits, employment, travel processes), proactive citizen communication, and education through a digital app. Engineering leaders in each ministry coordinate with central digital transformation team.

• •Ukraine building first 'agentic government' across all ministries despite ongoing conflict
• •Three pillars: government customer support, proactive citizen information, personal tutoring for education
• •Structure: engineering leaders in each ministry coordinate with central digital transformation team
• •Use cases include benefits inquiries, employment processes, travel documentation through conversational interface
• •Demonstrates how voice agents can transform large-scale institutional operations beyond commercial applications

Mati clarifies ElevenLabs' positioning against consulting firms (Palantir), point solution providers (Sierra), and foundation model companies (OpenAI). He argues that foundation labs won't focus on product layers or audio-specific research, while ElevenLabs excels at multi-use-case deployments with international support and forward-deployed engineering.

• •Not ideal for single point solution; best for deploying across multiple use cases (support, training, sales)
• •Consulting firms have broader digital transformation resources; ElevenLabs focuses specifically on conversational agents
• •Competitive advantage: international language support, geographic coverage, broader integration ecosystem
• •Foundation labs (OpenAI, Google) won't prioritize audio product layer or audio-specific architectural breakthroughs
• •Audio requires architectural innovation over scale: ~50-100 top audio researchers globally, ElevenLabs has ~10 of them

Mati predicts base models will commoditize within 2-4 years, making product layer and fine-tuning critical. He distinguishes between narration (where open source quality is good), real-time interaction (still 1+ year from passing Turing test), and real-time dubbing (2 years away). The key differentiator is controllability, not raw quality.

• •Internal belief: models will commoditize or differences become negligible in 2-4 years for most use cases
• •Narration: open source and commercial models achieving similar out-of-box quality, but controllability remains differentiator
• •Real-time interaction: cascaded approach (speech-to-text → LLM → text-to-speech) best for enterprise reliability currently
• •Fused speech-to-speech models emerging for more expressive use cases that can tolerate hallucination
• •Real-time dubbing/translation (Babel fish vision) still ~2 years from production quality

Mati reveals ElevenLabs just released Scribe 2 speech-to-text model with sub-150ms latency and 93.5% accuracy across 30 languages. Upcoming releases include new orchestration mechanisms that add emotional context to conversations, and investments in parallelized speech-to-speech approaches for more natural interactions.

• •Scribe 2 model: <150ms latency, 93.5% accuracy across top 30 languages, beating all benchmarks
• •Covers nearly 100 languages including low-resource languages for global accessibility
• •Upcoming: orchestration mechanism lowering end-to-end latency while adding emotional conversation context
• •Investing in fused speech-to-speech approach alongside cascaded systems for different use case requirements
• •Creative side: combining audio models with visual modalities, fully licensed music model, cross-modal integration

Mati shares his vision for voice as the primary interface, preferring 'Jarvis-style' super assistants over social companions. He predicts education will be transformed by on-demand AI tutors (potentially voiced by Richard Feynman or Einstein), while maintaining explicit human-to-human learning time. Voice will be critical for the coming 'decade of robots' following the 'decade of agents.'

• •More excited about Jarvis-style super assistants than social companion use cases for loneliness
• •Vision: start day with AI that understands context, controls environment (blinds, music), delivers personalized briefing
• •Education transformation: AI tutors on headphones/speakers with explicit percentage of time reserved for human interaction
• •Future content: learn from historical figures like Richard Feynman or Albert Einstein delivering lectures
• •After 'decade of agents,' expects 'decade of robots' where voice becomes primary human-robot interface