Speech models that understand
what you mean
The next generation speech lab building models that read tone, intent, and subtext so machines finally respond to what people actually mean.
Evaluations
State of the art at understanding people
Oruk models are purpose built for the human layer of speech. Here is how Resonance compares to frontier general purpose systems on two public benchmarks.
EmotionBench
Affect recognition accuracyIdentifying emotion, sarcasm, and concealed intent across 12k human labeled utterances.
Prosody Bench
Prosodic feature F1Decomposing intonation, intensity, rhythm, and stress against expert phonetician annotations.
Powering voice teams at
Understanding, not transcription
Same words. Different meaning.
Transcription tells you what was said. Oruk tells you what was meant. Select a phrase to see how the model reads beneath the surface.
Literal
Positive sentiment. The speaker is pleased about an update.
What they meant
Sarcastic frustration. The user is annoyed and likely overwhelmed.
Acoustic signal
Flat pitch, drawn out vowels, downward final contour.
What we understand
We hear the
when the words mean their opposite
Capabilities
A complete model of how humans really speak
Emotion & affect
Detect joy, doubt, anger, warmth and the subtle states in between across speakers and cultures.
Sarcasm & subtext
Read the gap between words and intent, the hallmark of human conversation.
Real time intent
Sub 200ms understanding so agents can respond in the rhythm of natural speech.
Context memory
Meaning that carries across a conversation through references, mood shifts, and history.
Prosody modeling
Pitch, pace, pauses and stress decoded as first class signal, not noise.
Private by design
On device options and strict data controls. Your voice never trains what you didn’t allow.
The lineup
One lab. A model for every conversation.
Our most expressive model. Full emotional and contextual understanding for production voice agents.
- Real time streaming
- 40+ languages
- Context window: full call
Low latency intent and sentiment for high volume routing, triage, and live assistance.
- <120ms latency
- Intent + sentiment
- On device ready
Frontier model for deep affect research, fine grained prosody, and cross cultural study.
- Fine grained affect
- Prosody decomposition
- Research preview
- 98.4%
- Intent accuracy on spontaneous speech
- <120ms
- Time to first understanding
- 40+
- Languages and dialects
- 12M
- Hours of expressive speech modeled