oruk is a speech understanding API built by a research team out of Stanford and Berkeley. Where most speech tools stop at transcription, oruk reads the acoustic signal underneath the words. One API call returns the transcript plus calibrated emotion analysis across 15 labels and speaking-style analysis across 16 labels, so your product can tell not just what people said but how they said it: intent, tone, hesitation, frustration, sarcasm, urgency, and more.

The models are fast enough for live products. Time-to-first-understanding lands around 120 milliseconds, which means live voice agents can adapt to a caller's tone mid-conversation instead of waiting for a transcript to be analyzed after the fact. For teams working with recorded audio, the same API powers batch analysis of call-center conversations, meetings, interviews, and qualitative research at scale.

Everything is benchmarked in the open. The team publishes speech-emotion-bench, a public evaluation of 64 systems on 64,384 held-out clips spanning roughly 20 languages, all scored through a single pipeline. On that benchmark, oruk Spectra reaches 77.6 percent accuracy, ahead of the best open-weights model at 68.7 percent and far ahead of frontier multimodal APIs such as Gemini 3 Flash Preview at 46.0 percent and GPT-Audio 1.5 at 43.3 percent.

Developers get a versioned REST API, Python and TypeScript SDKs, per-second metering, and a portal with usage and billing. Popular use cases include voice agents that detect frustration and escalate to humans, contact-center quality monitoring, meeting insight tools, and research pipelines that need vocal signals plain text throws away. You can start free with five dollars in trial credit, no credit card required, and make your first API call in minutes.