Cars were on the road for 70 years before anyone published a crash rating. Credit ran for a century before FICO. The measurement layer always shows up late. After the product, after the money, after everyone has spent years trusting the thing on gut feel. Voice AI is in that gap right now. I've sat with research teams evaluating voice models this year. Real matrices, weighted scores. Almost all of them measure two things: latency and word error rate. Not because that's what matters. Because that's what's easy to measure. Whether the model hears the hesitation before "fine, whatever." Whether it catches that the third "yeah" meant no. Nobody has a column for that. They listen to a demo and trust their gut. We're vibing it. And vibing voice quality doesn't scale. So today Hume is releasing Real World VoiceEQ, a benchmark for the human quality of voice AI. 40+ models, proprietary and open source. 1,000,000+ human judgments from real people rating real conversations. Different accents, different ages, the noisy kitchens people actually call from. It's built on a decade of Hume research into how emotion actually shows up in the human voice. That work is the reason we can measure this at all. There is no best model. Some are expressive and miss emotional cues. Some nail a booking code and sound like a kiosk doing it. "Which model won" is becoming the wrong question. Seventy years for cars. A century for credit. Voice just got its measuring stick. See the comment section for links!

Ten days ago I said our researchers had been showing me what was next and I'd been waiting a long time to talk about it. Today I can. Expression Measurement is live. Real-time streaming and file processing. New models, new API spec, docs, SDKs, a playground. Every voice AI company on earth can tell you what was said. Almost none can tell you how it was said. 48 emotions. 600+ descriptors. 50+ languages. Audio and video. In production, in real time, grounded in more real human judgment than anyone else has collected. Generating a voice is on its way to free. Understanding one is not. Really proud of this one. Heroic efforts from the entire team Hume AI. We've been saying we're the data and evaluation layer for voice AI. Today it's a product, and the start of the infrastructure to improve it. Sounding human or understanding humans. Which one do you think decides the next two years

The White House accord on super intelligence has teeth because of one word: auditable. Internal controls. External auditors. A board committee that receives the reports and cannot disregard them. D&O insurance that gets cancelled if they do. D&O insurance gets cancelled if boards ignore the reports. That's the enforcement mechanism. When an external auditor hands a report to the board, it stops being optional. The board has fiduciary duty to act on it. That is not self-policing. That is governance with real consequences attached. The three things every company deploying super intelligence now needs: End-to-end traceability: you have to know what your system is doing at the point of deployment, not after something goes wrong. Policy-to-risk mapping: your standard operating procedures exist, but you need to know exactly how your AI builds on top of them and where the gaps are. Auditable evidence: when a lawyer, a regulator, or a customer asks how a decision was made, you need something to hand them. Not a slide deck. A record. Companies claim safety without being able to show the work. The claim is not the proof. The audit is the proof. This is what private evals actually do. Not a public benchmark that every model has already seen. Not a leaderboard score that tells you how a model performs on the test it trained on. A blind, human-in-the-loop evaluation . Against the specific use case. The specific domain. The specific population of users you are actually deploying to. The companies that built that evidence before anyone demanded it are in a very different position right now than the ones who are going to build it in response to the first subpoena. The accord moved fast. The audit infrastructure has to move faster. Let's chat as it is what we are building.

Google just took #1 on our Real-World VoiceEQ leaderboard. Gemini 3.8 Flash TTS came out on top overall, out of 34 models. Flash Lite came in right behind it at #2. So we added a new Frontier Metric that scores models on both at once. Gemini 3.8 Flash is the first model we've seen sit at the top of expressivity while also staying in the upper tier on reliability. 1. The most expressive models break down on reliability. 2. The most reliable ones sound flat. You could pick one. The biggest jump was in long form stability, meaning how well a voice holds up over minutes of speech instead of one sentence. Almost nothing anyone ships is a single sentence. Extended long form scores more than doubled over the previous Gemini generation. Public leaderboards tell you who's winning. Private evals tell you why, and what to fix next. At the frontier, models rarely fail across a whole category. They fail in specific places: a regional German dialect, background chatter instead of background music, Asian languages while European ones score well. You won't see that in a public average. That's the work we do with our process, methodology and platforms: • Evaluations tailored to your domain, modality, and use case • Blind human ratings from vetted evaluators, plus audio measurement and regression tracking • Held out prompts and references, so the benchmark can't be trained against or gamed hume.ai/blog/newly-released-…

A researcher on our team showed me a clip last month that stuck with me. A speaker on a European Parliament recording clearly says "Thank you, Mr. President." 6 of the 11 leading ASR models we tested wrote just "Mr President." Not because they misheard it. Because that's what the benchmark's transcript says. The benchmark has an error and the models learned to repeat it instead of transcribing the actual audio. That one clip turned into a research project. Public voice AI benchmarks say models are near human-level accuracy. But if models are also learning to recognize which benchmark they're being tested on, and quietly matching its expectations instead of the audio, those scores stop meaning what we think they mean. So we built three tests: ** Do models reproduce known errors in benchmark transcripts? ** Do they "transcribe" numbers we silenced out of the audio entirely? ** Do they switch spelling conventions ("Mr." vs "Mister") depending on which dataset they think they're in? Across 11 widely used open ASR models, the answer was yes, consistently. On fresh audio the models had never seen scored against, most of this behavior faded. We published the full research with @huggingface about the methodology, failure cases, and what it means for anyone choosing a voice AI model. I am very proud of our team as we continue to publish insightful and meaningful research on what it means to truly evaluate Voice AI.  We remain convinced that Voice AI has a listening problem as right now it only reads.  See the research and post hume.ai/blog/measuring-bench…

Looking forward to joining you!

Technical Paper: cdn.sanity.io/files/xqnc2for… Release Blog: hume.ai/blog/introducing-rea…

As voice becomes the default interface for a new generation of AI apps and devices, we’re honored that Hume AI has been named one of Fast Company Most Innovative Companies of 2026. We're grateful to our team, customers and partners who made this possible. We aren’t predicting the future of Voice AI, we are creating it. Read the full list here: fastcompany.com/most-innovat… P.S. - We are hiring so if your passion is the future of Voice please reach out as many roles are not even posted yet!