Your guide to getting started with ZoomMate
Everything you need to start using ZoomMate: notes, search, workflows, and AI tools — all inside Zoom Workplace.
Select a language
Every AI conversation begins with a machine trying to understand a person. I have spent most of my career on that problem, and I still find it the most honest test of whether a system is ready for real work.
6 min read
Published on September 4, 2026
Xuedong Huang is the Chief Technology Officer (CTO). Prior to Zoom, he was at Microsoft where he served as Azure AI CTO and Technical Fellow. His career is illustrious in the AI space: he began Microsoft’s speech technology group in 1993, led Microsoft’s AI teams to achieve several of the industry’s first human parity milestones in speech recognition, machine translation, natural language understanding, and computer vision, is an IEEE and ACM Fellow and an elected member of the National Academy of Engineering and the American Academy of Arts and Sciences.
Xuedong received his Ph.D. in EE from the University of Edinburgh in 1989 (sponsored by the British ORS and Edinburgh University Scholarship), his MS in CS from Tsinghua University in 1984, and BS in CS from Hunan University in 1982.
Every AI conversation begins with a machine trying to understand a person. I have spent most of my career on that problem, and I still find it the most honest test of whether a system is ready for real work. AI grows more capable every month. Capability is the easy part — the harder question is whether that capability survives contact with real work: a clinician's note, a support interaction, a team decision, an accessibility tool, a voice agent helping someone finish a task.
At Zoom, one principle keeps that question front and center for Zoom AI Services (ZAS) — our platform that brings AI capabilities like speech recognition, translation, and intelligent agents to enterprises, creators, and developers:
ZAS Success = Accuracy × Ease of Use Inference COGS
Accuracy earns trust. Ease of use turns capability into adoption. Inference COGS — the compute cost of serving each request — decides whether that adoption can be sustained.
Neglect any one of them, and a promising demo fails as a product.
Everything in conversational AI begins with listening. If a transcript is wrong, every layer above it inherits the error. A summary can be elegantly written and still be false. An agent can act on it with complete confidence. A workflow meant to save effort can cost more of it.
That is why accuracy comes first.
The Speech Accessibility Project works to improve speech technology for people with disabilities and speech patterns that conventional systems have handled poorly. We ran the production Zoom Scribe API on its public dataset, exactly as it reaches customers, and achieved a word error rate of 5.72%. The next-lowest result in our comparison was 11.99%. The number needs context: no single dataset captures every language, speaker, or environment, and an internal comparison is not an independent certification. It is evidence of progress, not proof that speech recognition is solved.
The Open ASR Leaderboard (lower scores indicate better performance)
Our commitment to accuracy is ongoing. The latest Scribe API achieves over 10% reduction in relative word error rate compared to the previous generation. Accuracy isn't a one-time milestone — it's a continuous improvement process. ZAS Scribe API Pro has now reached state-of-the-art performance in production, as demonstrated by its ranking on the Open ASR Leaderboard. The accuracy alone does not earn trust. That also requires privacy, security, transparency, appropriate safeguards, and human judgment where the consequences demand it.
The most sophisticated technology usually produces the simplest experience.
Customers should not have to understand what sits beneath an AI capability in order to use it. A developer building clinical documentation, a voice agent, or a contact-center application should be able to focus on the experience they want to create — not on stitching together speech models, GPU infrastructure, orchestration systems, and post-processing pipelines.
Zoom Scribe provides Live, Fast Sync, and Batch transcription, along with capabilities such as speaker diarization. Each depends on difficult work across models, data, infrastructure, and operations. That complexity belongs inside the platform, not in the customer's way.
This is what ease of use means to us. Not fewer capabilities. Fewer barriers between a customer's intent and the outcome they want.
The best infrastructure does not ask for attention. It turns complexity into a dependable building block.
One dimension of AI draws less attention than model capability but ultimately determines how widely AI can be deployed: inference economics.
Every transcription consumes compute. So does every generated answer, translated sentence, synthesized voice, and agentic action. A small difference in cost or latency looks trivial in a prototype. At production scale, it decides whether a feature reaches every customer or only a few high-value ones.
So our teams work across the inference stack: model architecture, quantization, memory movement, batching, GPU utilization, and kernel optimization. The goal is not optimization for its own sake. It is to make better intelligence affordable enough to use everywhere.
NVIDIA's SOL-ExecBench measures how closely GPU kernels approach hardware-specific speed-of-light bounds across workloads derived from AI models. Our ZAS submission reached state-of-the-art performance on the Nvidia B200 quantization track. We reached 87.8% by combining codified engineering expertise — CUDA, CUTLASS, Triton, and Blackwell-specific optimization — with an agentic kernel-development system that proposes hypotheses, synthesizes implementations, tests correctness, measures performance, and records what it learns. The same system is now working on what comes next, and it starts from everything the previous attempt taught it. This is the practical promise of our Federated AI approach. Accumulated expertise becomes reusable, and the search for improvements becomes a learning loop instead of a sequence of isolated experiments. Engineers set the goals and constraints. AI explores a far larger solution space. Rigorous evaluation decides what is real.
NVIDIA SOL-ExecBench Leaderboard (higher scores indicate better performance)
A benchmark is never the destination. An optimization matters when it lowers latency, lowers cost, adds capacity, or makes a more capable model feasible in production. The hardware result is interesting. The customer outcome is the point.
Benchmarks measure technical progress. Customers reveal whether that progress creates value. athenahealth evaluated Zoom's Scribe Live API for clinical documentation, assessing transcription quality, medical terminology recognition, speaker attribution, latency, and data handling. The evaluation showed strong performance in medical entity recognition while meeting athenahealth's critical transcription requirements.
The number is not the important part. The path is.
Accuracy made the technology worth evaluating. Ease of use made integration practical. Efficient inference made production deployment sustainable.
That is the equation in action.
Trust gets you considered. Simplicity gets you adopted. Economics gets you scaled.
When all three hold, customers can build experiences that were previously too difficult, too expensive, or simply impractical.
Speech recognition is only the beginning.
A conversation carries far more than words — intent, decisions, questions, commitments, knowledge, actions. Historically, most of that disappeared the moment the conversation ended.
AI lets us change that. Over time, ZAS can help developers move from conversation to intelligence, and from intelligence to action: speech recognition, speaker understanding, translation, speech generation, structured extraction, reasoning, agentic experiences.
To an engineer these are separate technologies. To a customer they are one experience:
A person speaks.
The system understands.
Intelligence helps.
Work gets done.
That is where conversational AI is heading. Not toward more AI for people to manage, but toward more intelligence embedded in how people already communicate and work.
Getting there takes more than adding capabilities one at a time. It requires the platform to improve faster than any single piece of it could alone. Which is why we hold to one principle internally:
1P = 3P. What we build for Zoom's own products reaches developers on Zoom AI Services, running on the same underlying technology. And what we learn from operating it at scale improves the platform beneath both.
That creates a compounding loop:
More real-world usage → better understanding → better models and systems → simpler services → more customer value → more real-world usage.
Executed well, each turn makes the next one stronger. This is the advantage of building across products, models, infrastructure, and developer services. Owning more technology is not the objective. Learning faster is — and turning that learning into customer outcomes.
AI will keep changing. Models will keep improving. Hardware will keep getting faster. Today's benchmark leaders will be replaced.
So the standard has to survive any single technology cycle. Ours is simple:
ZAS Success = Accuracy × Ease of Use Inference COGS
Can we understand more people, more accurately?
Can we make increasingly powerful AI dramatically easier for developers to use?
Can we keep lowering the cost of delivering that intelligence?
And are customers getting enough value by choosing Zoom and able to build on it?
We are encouraged by the progress. But we are not finished. The hardest problems in front of us are the ones the field has not yet solved: long-tail languages and dialects, crowded and noisy rooms, and teaching a system to recognize the moments when it should defer to a person rather than answer. We would rather name those than claim them finished.
The direction is clear enough. The future of AI will not be decided by capability alone. AI becomes useful at scale when people can trust it, developers can use it without wrestling with complexity, and the economics let it reach everywhere it needs to be.
Better intelligence. Less complexity. Better economics.
Every conversation should leave something behind. That is what we are building.
Zoom Scribe is available today through Zoom AI Services (ZAS). Documentation, quick start guides, and pricing are here. Build something with it — and tell us where it falls short. That is the part we most want to hear.
