Nathan Roll

That familiar voice might be AI.

Meet OrukZero, our AI speech detector.

Upload an English voice clip. Get a human-or-AI prediction and segment scores for longer recordings.

Try the demo free. No account needed.
API included with paid Oruk plans.

Links below.

hearing is no longer believing.

Meet OrukZero, our AI speech detector.

Upload an English voice clip. Get a human-or-AI prediction.
Try the demo free. No account needed.

API included with every paid Oruk plan.

Built by @OrukLabs. Link below.

Adding noise can make speech clearer.

A sound is missing from both versions. One leaves a silent gap. The other fills it with noise.

Can you hear the word more clearly in the second?

@OrukLabs

Long-running agents can log thousands of steps per run.

Analyzing every step to find what went wrong is slow and expensive.

@kalmiadev has been experimenting with @typesafeai Jev to help choose which steps matter for our harness to analyze.

Same bugs found on real customer runs, 9x cheaper and 6x faster.

Try now kalmia.dev

Your ears have an accent.

In a classic study, Japanese listeners often heard "ebuzo" where French listeners heard "ebzo". Same recording. An extra vowel.

The language you grow up with can change what you hear. (what do YOU hear?)

@OrukLabs

Can you tell if this voice is AI?

It's hard to know what's human anymore.

We're releasing OrukZero today for free, which distinguishes AI from human voices with 99.7% accuracy.

Try it now @OrukLabs

Also, catch us @ a16z @speedrun Demo Day today.

Introducing Kardashev-0.7, the world’s first trained swarm made of 32 distinct models

Trained with RL for Population Scaling (RLPS), 32 models organically develop specialization & complementary capabilities, delivering frontier performance at:
- 0.007x ~ 0.02x of the inference cost
- 0.03x of the required memory

Civilization advances through different minds specializing and working together. We’re bringing that principle into AI

@BanburyRoadAI, we’re scaling intelligence by model count, toward civilizations of models that learn to build on one another

How would you feel if your AI could read the room better than you?

Meet Oruk Spectra 2: speech recognition with emotion and tone.

Transcription in 25 languages. Emotion scores alongside the words. 0.39s median after audio ends in our streaming test.

you can sound dishonest without telling a lie.

Researchers changed how made-up words were spoken. Listeners judged some versions as more honest.

Here’s an audio illustration. Same voice, same invented word. Which sounds more believable? @OrukLabs

pauses are words too?

These are the same voices saying the same words. We only changed the pauses between replies.

Which version sounds like they click?

An audio illustration of a real experiment. @OrukLabs

You can change a sentence without changing a word.

Say "I asked for the BLUE one." Now stress "I."

The correction moves from a color to a person.

Irene Yi breaks down tone of voice in our new @OrukLabs article. The animation illustrates emphasis.

your brain can turn three whistles into a sentence.

Listen once. Hear the original. Listen again.

The first and last sounds are the exact same audio. @OrukLabs

even your “um” has an accent.

The sound you make while looking for a word has patterns of its own. English, German and Russian speakers don’t hesitate quite the same way.

Three real speakers, each played twice. @OrukLabs

you can hear someone talking with their hands.

In a controlled experiment, listeners matched unseen arm movements using the voice alone.

We animated an original study recording with its accompanying motion trace. @OrukLabs

this is what happens when linguists build a speech model. it actually understands what you mean.

"I told you to buy a bunch of red roses."

change the emphasis and you change the instruction. a plain transcript misses it. @OrukLabs

Unstoppable.

we squished 24 encoder blocks into one and somehow improved ASR performance.

looped Orukeet runs the same block repeatedly, giving it another pass at the audio without adding another set of weights.

@OrukLabs

“it’s eleven o’clock” hits different when you were supposed to be home at ten.

we gave Resonance 2 the same sentence spoken with six different emotions.

here’s what happens inside the model when the words stay the same, but the delivery changes.

@OrukLabs

Your “s” has a social life.

The same acoustic feature can mean different things in different communities. Irene Yi explores how tiny details in speech shape our impressions of a speaker.

Film: a synthetic sound unfolding into its spectrum. @OrukLabs

Your AI tutor should notice when it's losing you.

We're working with Edixir to test speech alongside handwritten math, so a tutor gets more context from the way a student asks for help.

Illustration with synthetic Inworld voices. @OrukLabs