We’ve tested our Audio Decisions API on iconic performances. All kinds of iconic.
This rooftop scene swings from peak fury to casual pleasantries in seconds, and the model tracks every pivot. Emotion detection on hard mode.
Not a product, just a glimpse of what this model makes possible.
Our team’s so good at optimization, they even cut the time between performance breakthroughs.
Two weeks since launch, the Audio Decisions API is faster across the board. Responses now average ~160ms, with 95% back in about 200ms, and our slowest responses are down ~60%.
Try it yourself, no account required. 250,000 free tokens (~50 min of audio).
Hume AI has been named to CB Insights and Accenture’s NYC Emerging AI 50, which recognizes the most promising AI startups based in New York.
We’re listed under AI infrastructure, and that’s exactly where we think voice understanding belongs: underneath the products, helping every system that listens to people go beyond the transcript and actually listen.
Thank you to @CBinsights and @Accenture for the recognition, and to the incredible team here whose work earned it.
Experimental demo: Scene Analysis
In one of his earliest explorations, research engineer Hoon Shin asked: what happens when our Audio Decision API listens to a scene from a movie?
Each orb is a line of dialogue. Its color shows the emotional mix, its size and pulse show intensity, and a red flare means someone’s raising their voice. Threads trace speaker changes, and the gauge tracks tension across the scene,
All from the audio alone.
Not a product, just a glimpse of what this model makes possible.
Research engineer Hoon Shin built this liquid metal soundscape with our Audio Decision API. Using the model’s span feature, it picks out and identifies each sound, then renders it as an object morphing in and out of existence.
Voices are read for age and language, while laughter and cheers are read for emotion and intensity, and the louder the room gets, the more the sea moves.
Not a product, just a glimpse of what this model makes possible.
When AI can actually hear, things get a lot more interesting.
just saw this and i'm still shocked.
@hume_ai trained a model on millions of human judgments to rate voice AI. tone, timing, accents, noise.
it matched a panel of human listeners more closely than a single human rater did.
now labs can test every checkpoint.
🤝 Paid partnership
Introducing the Hume Audio Decision API.
Ask anything: How many speakers? Is anyone laughing? Is the caller frustrated?
What used to take 10+ models is now one call.
Trained on 10M+ human judgments, so it answers the way human listeners would.
Realtime, or across hours of recordings.
Request access below.
Emily Beimfohr retweeted
When your girlfriend says "I'm fine," you know you're in trouble.
When an AI agent says "everything's fine," you also know you're in trouble.
Because tone is half of every convo, and AI has been ignoring it this whole time. Most voice agents would read a transcript, see the word "fine," and move on like nothing happened.
And every guy reading this right now knows exactly how ignoring the signs ends for them.
Hume just launched their Decision API so AI can now finally hear the difference.
Send it audio, ask any question in plain English like "does this person sound frustrated?" and it gives you back a probability for each answer your app, your business or your life can actually act on.
It doesn't need a transcript either. It picks up the sighs, the pauses, and the "I'm fine" when your users very much aren't.
Labs have spent years making AI voices sound human, but as it turns out the bigger gap was that agents were reading transcripts, never listening.
AI is officially better at reading the room than some of my friends now. Holy shit.
🤝 Paid partnership
Introducing the Hume Audio Decision API.
Ask anything: How many speakers? Is anyone laughing? Is the caller frustrated?
What used to take 10+ models is now one call.
Trained on 10M+ human judgments, so it answers the way human listeners would.
Realtime, or across hours of recordings.
Request access below.
Introducing the Hume Audio Decision API.
Ask anything: How many speakers? Is anyone laughing? Is the caller frustrated?
What used to take 10+ models is now one call.
Trained on 10M+ human judgments, so it answers the way human listeners would.
Realtime, or across hours of recordings.
Request access below.
Someone says "I'm fine." You know they aren't.
You heard it - the pause, the flat tone. Voice AI doesn't. It all gets dropped at transcription.
So we built the Expression Measurement API. 48+ emotions, 600+ descriptors, 50+ languages, audio and video, real-time.
Conversation was always more than words.
6/ The board on Hugging Face, with the full per-factor breakdown:
huggingface.co/spaces/HumeAI…
Same data on our site, with every factor tab:
hume.ai/rw-voice-eq
We asked 11 voice cloning models to copy the same 25 voices, then asked real people: is that the same person?
The most natural-sounding clone ranked 8th of 11 on sounding like the person.
New from Hume AI: the Voice Replication Leaderboard. Live now on @huggingface, links below.
Sounding human is the third most important factor in voice AI. Not the first.
Our CEO Andrew Ettinger (@ahe23) joined the MTS podcast (@MTSlive) to break down CLEAR, the framework we use internally to grade voice AI:
1. Can it respond appropriately? If not, nothing else matters.
2. Can it actually understand you? If not, it's just guessing well.
3. Does it sound real? If not, you won't keep talking to it.
4. Can it do all three on interaction #1,000?
The bar isn't a voice demo. It's whether your parents trust it with their doctor's office.
Emily Beimfohr retweeted
Today, we are launching Tinker grants of up to $50,000 in credits for safety research on open-weight models. We share some project ideas that excite us below; if you’re working on a safety project that could be accelerated by additional Tinker credits, we want to hear from you!
For anyone who saved the paper and hasn't opened it :) huggingface.co/blog/asr-benc…
We silenced a number in the benchmark audio. It was never spoken.
Some of the highest-scoring speech recognition models wrote it down anyway, in 30–40% of cases.
New research with @huggingface on measuring when a model is optimizing for the test instead of the audio.
Emily Beimfohr retweeted
That growth wasn't evenly distributed. It came from people the big models don't serve well.
English sounds different everywhere, from Singaporean and Malaysian to Indian, Australian, and British. When cloning a voice, our models preserve what makes each accent distinct instead of flattening it into a generic American voice. And we go beyond English, we also support Mandarin, Japanese, Korean, Spanish, Arabic, German, French, and Italian.
Our creator community tell us when a voice sounds wrong, and we tune emotional nuance from that feedback. No lab dataset gives you that.
Emily Beimfohr retweeted
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!










