Interviews
Anton Dvorkovich, СEO and Founder at Dubformer – Interview Series

Key PointsAI dubbing still fails at emotionally authentic performances even when speech sounds…
View more
- AI dubbing still fails at emotionally authentic performances even when speech sounds human.
- VOX-DUB's open benchmark forces companies to prove dubbing quality claims with measurable evidence.
- AI localization will let independent creators reach global audiences once reserved for big studios.
Anton Dvorkovich, CEO and founder at Dubformer, has spent over a decade shaping the intersection of AI, language, and speech technologies. Before launching Dubformer in 2023, he led major machine learning and NLP initiatives at Yandex, overseeing products such as Yandex Translate, Yandex Keyboard, and the voice assistant Alice, each serving millions of users. At Nebius, he served as Head of AI, advancing large-scale infrastructure for intelligent systems. His experience managing cross-functional teams of over 100 engineers, developers, and researchers provided the foundation for Dubformer’s vision—bridging language divides through advanced AI dubbing.
Dubformer is an AI-powered dubbing startup headquartered in Amsterdam that has already processed over one million minutes of localized video content. The company specializes in translating and dubbing videos using synthetic voices that preserve emotional tone, timing, and speaker individuality, helping creators, studios, and enterprises make their content globally accessible. Its platform leverages deep learning models for speech recognition, voice cloning, and translation, offering natural-sounding results that far surpass traditional dubbing tools.
After leading NLP and speech teams at Yandex and serving as Head of AI at Nebius, what motivated you to build a dedicated AI dubbing company, and how did those experiences shape your vision?
Working at Yandex as Head of NLP, I developed a passion towards breaking down language barriers with technology. I got to witness (and to participate in) the AI revolution that happened to machine translation – and then I realized that a big gap that still remains largely unsolved is media translation (audio & video). When you see text translation becoming nearly perfect, you start asking: what about the stories that actually move people? What about the performances that make you laugh or cry? So we decided to take on the challenge of building AI dubbing technology, and to go beyond media translation that is just comprehensible, to the one that is actually enjoyable. This is why we launched Dubformer.
What specific gap in media localization did you identify, and how did your background in neural machine translation and voice technologies influence your approach to solving it?
Media localization today is done almost entirely with voice actors, and that has several drawbacks: it is quite expensive, the turnaround times are slow, and scaling it is very hard and leads to huge overheads in people management. Think about it: traditional dubbing craftsmanship is incredible, but those constraints mean lots of projects never get dubbed at all. So, while traditional dubbing is quite good, there are lots of projects that are not being dubbed at the moment because of the aforementioned drawbacks. In fact, 95% of content is only available in the original language. That’s the gap we’re addressing.
The VOX-DUB benchmark is an important milestone for evaluating dubbing quality. What insights did participation provide about where the industry currently stands?
The VOX-DUB results revealed something important: the industry has already solved the problem of making AI voiceovers sound ‘human-like’ for content where the emotional component isn’t critical. However, dubbing content where emotions truly matter remains an unsolved challenge, even though it might seem that everything is already figured out. Many systems can produce understandable speech, but very few can create performances that feel authentic and emotionally aligned with the original. We’ve mastered reading text aloud, but we’re still learning how to actually perform scenes.
You’ve spoken about a trade-off between voice fidelity and overall speech quality. How do you approach balancing these factors in your systems?
In reality, the trade-off often isn’t between voice fidelity and quality, but between the depth of emotional, intonational, and expressive transfer and the stability of the result. The more you try to transfer, the higher the risk of carrying something over incorrectly, especially since emotions sound and are expressed differently across languages. What sounds passionate in one language might sound unnatural or even jarring in another. We aim to find a balance between emotional accuracy and naturalness, ensuring the output remains expressive yet comfortable to listen to.
In your view, what makes AI dubbing fundamentally different from traditional text-to-speech—both technically and artistically?
Traditional TTS focuses on simply reading text, usually in a neutral tone. AI dubbing, especially in the media industry, is a far more complex task, where the goal is not just to convey meaning but to create a sense of presence and genuine performance. Emotions, intonations, and natural pauses all play a crucial role, along with challenges such as mixing, acoustic adaptation, and synchronization with lip movements and scene rhythm. Artistically, AI dubbing is much closer to acting than to reading; the system has to perform the line, not just pronounce it.
What aspects of the VOX-DUB evaluation framework do you find most valuable in driving transparency and accountability in AI media localization?
What makes VOX-DUB so valuable for transparency is it forces everyone to be honest about what they can actually deliver. Before this framework, companies could make claims about their dubbing quality without any way to verify them. The crucial breakthrough is that VOX-DUB introduces the first systematic catalog of AI dubbing errors through its comprehensive human-based benchmark. This open-source dataset creates a shared language for quality – original speech fragments with translations, generated audio from multiple systems, and human annotations across five critical aspects: pronunciation, naturalness, sound quality, emotion similarity, and voice similarity. For the first time ever, we have a standard for human evaluation of dubbing with clear guidelines and examples. Now when teams talk about quality, everyone’s actually talking about the same thing. Published results that anyone can see create real accountability. Teams have to back up their marketing claims with performance data across all five quality dimensions. This kind of openness changes how the entire industry operates and builds trust with creators who need reliable, measurable results.
How do you see benchmarks like VOX-DUB influencing the pace of progress in multilingual content creation and distribution?
The beautiful thing about having clear benchmarks is it creates a roadmap for everyone to follow. When teams see what’s achievable through VOX-DUB results, it accelerates their own development timelines. Instead of guessing what quality looks like, they have concrete targets to aim for. I’m watching companies compress their roadmaps because they can see evidence of what works elsewhere. This shared understanding is helping the entire field move faster together, which means more diverse content can reach global audiences sooner.
Your platform achieved strong results in pronunciation and naturalness. What underlying innovations or model choices contributed most to this?
It comes down to understanding that dubbing isn’t just about individual words; it’s about the entire performance ecosystem. Two breakthroughs made the difference. First, we stopped treating pronunciation as purely technical and started treating it as emotional. People don’t speak in perfectly articulated syllables; they slur words when sad, they rush when excited. Capturing that natural imperfection was crucial. Second, we had to think about the entire soundscape. When someone’s speaking in a scene with rain, or in an echoey cathedral, their voice naturally adapts. Our system learned to blend with the original acoustic environment. These weren’t just technical improvements. They were about understanding what makes speech feel human.
How does your platform blend human expertise with automation to deliver professional-grade dubbing at scale?
We work closely with traditional dubbing studios, helping them train their teams in new skills and emerging professions. AI handles translation, synchronization, and voice generation, while professional reviewers, translators and sound engineers, perform the final quality check through our platform. This approach allows us to maintain studio-level quality while reducing turnaround times from weeks to hours. Automation provides scalability, and human expertise ensures artistic expressiveness. It’s the best of both worlds: machine efficiency with human artistry.
With the AI media localization market growing rapidly, what broader role do you see AI playing in reshaping global creative production?
AI localization will make dubbing accessible to a much wider range of creators and storytellers. It will enable them to share their stories with global audiences without losing emotional depth or cultural authenticity. Over time, this will lead to far greater diversity in content, opening the door for niche projects and specialized audiences that previously couldn’t afford professional dubbing. Think about independent filmmakers, documentary makers, educational content creators; they can finally reach global audiences that were previously out of reach.
What upcoming advancements or collaborations are you most excited about as you look toward the next stage of development?
What excites me most is continuing to improve our emotional transfer capabilities. We’ve made good progress, but there’s still so much room to get better at preserving the subtle nuances that make performances feel authentic. On collaborations, we’re building deeper partnerships with traditional dubbing studios to create hybrid workflows that combine the best of both worlds. These studios have decades of expertise that we’re learning from, while we provide the technology to scale their artistry. The most exciting collaborations are with studios that see AI not as a replacement, but as a way to augment their teams and take on more diverse projects they couldn’t handle before. The real breakthrough will be when we can consistently deliver performances that audiences forget were created by AI. That’s where we’re heading.
Thank you for the great interview, readers who wish to learn more should visit Dubformer.












