How much faster can the same speech model run?
Faster speech inference with the original acoustic weights.
Speech-to-text, emotion, intent, tone of voice, meaning, sarcasm, and understanding.
Oruk is a research lab training speech foundation models to understand words, tone, and emotion directly from audio. Our API helps developers build AI that can recognize a hesitant answer or respond to a frustrated customer.
Backed by
Research backgrounds
try resonance-1
r u ok? understand how things are said
Loading the recording…
01 / Voice agents
Give your agent a better ear.
Let your agent pick up on hesitation, frustration, or excitement, then use that context to shape its response.
- Phrase-level emotion
- Conversation review
- Response design
Caller
“I’m not sure what to do next.”
worriedhesitant
Make room for a clearer explanation.
Illustrative workflow
Research highlights
All researchOrukeet: a new shape for speech recognition
An open speech model for 25 languages, with fitted Gabor filters.
How a quantum memory learns what to connect
Past questions help a quantum memory choose which qubits to connect.
We taught a fruit fly to hear human emotion
A speech model built from 499 fruit fly neurons.
Speech models learn physics
Six speech systems learn the same filter shape, predicted by signal theory.
The shape of a voice
Speech models organize voices into shapes you can explore in 3-D.
Universal, with an accent
Emotion structure transfers across languages, with a measurable gap.
Resonance-2: hear how it was said
Continuous emotion and speaking-style scores, alongside the words.
What a smile does to a voice
Mouth shape changes speech. Model happy tags need their own controls.
Color is shared. Meaning has an accent.
Shared color associations vary with language, climate, and context.
More audio. Less spend.
Prices in USD. Included minutes reset monthly.
Hobby
$9/ month
Billed monthly
- Speech understanding
- 250 min/month
- Transcription · Orukeet
- 20,000 min/month
7-day free trial
Start 7-day free trial of Hobby$0 today. Then $9/month. Cancel before the trial ends.
Builder
Recommended$49/ month
Billed monthly
- Speech understanding
- 2,500 min/month
- Transcription · Orukeet
- ≈ 109,000 min/month
7-day free trial
Start 7-day free trial of Builder$0 today. Then $49/month. Cancel before the trial ends.
Production
$199/ month
Billed monthly
- Speech understanding
- 20,000 min/month
- Transcription · Orukeet
- ≈ 442,000 min/month
7-day free trial
Start 7-day free trial of Production$0 today. Then $199/month. Cancel before the trial ends.
Enterprise
Contact us
Tailored to your team
- Speech understanding
- Custom
- Transcription · Orukeet
- Custom
How the two allowances work
Resonance, Fourier, Realtime, and Proficiency share the speech understanding allowance. Orukeet transcription has a separate allowance; optional tasks and request rounding reduce its estimated minutes. Allowances reset monthly. Trials include 25%; extra usage shares one spending cap.
Now serving: Resonance 2 — emotion and style, same pricing and existing plans.
Benchmarks
Full results & methodologySpeech emotion
Accuracy ↑+9.1 pts vs next best
oruk Resonance 177.8%
95% confidence interval 77.5 to 78.1emotion2vec+ seed68.7%
95% confidence interval 68.3 to 69.1EmotionThinker60.5%
95% confidence interval 59.1 to 61.8SenseVoice Small55.7%
95% confidence interval 55.3 to 56.1Hume AI legacy prosody49.6%
95% confidence interval 48.2 to 51.0Gemini 3 Flash Preview46.0%
95% confidence interval 44.6 to 47.4GPT-Audio 1.543.3%
95% confidence interval 41.9 to 44.7Qwen2.5-Omni 7B31.3%
95% confidence interval 30.0 to 32.6Oruk trained in-distribution; alternatives zero-shot. Open models: 64,384 clips; API/audio LLMs: 5,000. Hume is the legacy prosody endpoint.
Sarcasm
Macro-F1 × 100 ↑Vocal context vs transcript sentiment
Standard 5-fold
Oruk Resonance 156.2
95% confidence interval 52.4 to 60.1Deepgram nova-351.7
95% confidence interval 47.8 to 55.3Speaker-independent
Oruk Resonance 151.0
95% confidence interval 47.1 to 54.8Deepgram nova-339.5
95% confidence interval 35.8 to 43.1Cross-show
Oruk Resonance 146.2
95% confidence interval 40.9 to 51.5Deepgram nova-329.9
95% confidence interval 27.3 to 32.2Two identical research probes on MUStARD, not standalone API classifiers. Standard-split confidence intervals overlap.
Protocols & splitsResonance 1 · July–August 2026 evaluation. Whiskers show 95% confidence intervals. Results describe the evaluated checkpoint.
Work with us
Talk to the people
behind the models.
Bring us your audio and your use case. We’ll help you find the right model and integration.
Build with Oruk
Send audio. Get a transcript, speaker turns, and vocal context.
Common questions
What exactly does Oruk do?
Oruk turns speech into transcripts, speaker turns, and labels for emotion and delivery. Choose a model for recorded audio or live conversations.
How is it different from speech-to-text?
Speech-to-text tells you what was said. Oruk also analyzes how it sounded: frustrated, excited, hesitant, sarcastic, and more.
What can I build with it?
Voice agents with more context, searchable call reviews, expressive captions, research tools, and other products that work with speech.
Can I try my own audio?
Yes. Open the full demo to use your microphone or upload a short recording. You can try it without an account.
Try your own audioDoes it work in real time?
Use Resonance for a complete English recording: it returns a transcript, emotion and speaking-style labels, and timed segments. Use Realtime for live multilingual transcription with phrase-level emotion. Realtime is in preview and does not return the full file-analysis label set. Orukeet is an option for short English recordings or streamed utterances, with final text after commit.
Which languages and formats are supported?
File analysis supports English audio in WAV, FLAC, MP3, M4A, OGG, and WebM. Realtime supports 32 locales with automatic language detection. Language support for transcription does not imply that every emotion or style task is available in that language.
What happens to my recordings?
Oruk’s inference services discard audio and outputs after responding. Optional speaker diarization retains uploaded audio for up to 48 hours and speaker-label results for up to 24 hours. Request metadata is retained for billing, security, and support. Oruk does not use customer audio or outputs to train, fine-tune, or evaluate models without your explicit written agreement.
Data handlingHow accurate are the results?
Performance varies with the model, language, recording, and task. The benchmark report includes dated results, confidence intervals, and evaluation conditions. Emotion scores describe vocal expression, not someone’s inner state.
Read the evaluationsHow does pricing work?
Every plan includes all speech models. From $9/month with a 7-day free trial.
See pricing