Try it live — clicktap and speak
Powering speech understanding at
Speech emotion
oruk Spectra · Accuracy — higher is better
speech-emotion-bench: 64,384 held-out clips, 7 emotion classes, ~20 languages, one identical scoring pipeline for all 64 systems.
Transcription
oruk Spectra · WER — lower is better
LibriSpeech clean/other · Common Voice · TED-LIUM
Per-emotion F1
oruk Spectra against Gemini 3 Flash Preview, the strongest frontier multimodal API in the July 2026 report, on all seven emotion classes.
The lineup
Highest-accuracy transcription and unified analysis.
oruk-resonanceAccess the APIThe efficient choice for emotion and style.
oruk-spectra-1Access the APIA preview tier that is not yet serving traffic.
oruk-spectra-2Access the APIScope
English transcription. Transcribes prerecorded English audio files via POST /v1/audio/transcriptions.
Multilabel emotion detection. 15 emotion labels with scores calibrated on held-out audio; a clip can carry several labels at once.
Speaking-style classification. 16 speaking-style labels (calibrated, multilabel) describing how something was said.
Combined affect. Emotion and speaking style in one call, skipping transcript output when it is not needed.
Unified analysis. One request returns transcript, calibrated labels, time-local segments, and tagged text.
No streaming or real-time API — the public contract is file-based (prerecorded audio uploads).
No production multilingual support — API v1 processes English audio.
No intent or deception detection. The sarcastic speaking-style label scores delivery — how speech sounds — not what the speaker meant.
No claims about a speaker’s inner mental state — outputs are calibrated acoustic annotations of how speech sounds.
API v1 · scope statement updated 2026-07-13. Before July 2026, oruk.ai hosted an unrelated broadcast/news-intelligence product. That product is retired. oruk is a speech lab; the current product is the file-based Speech API described here. Full details on the capabilities page.
Questions
oruk API v1 processes English audio files for transcription, emotion detection, speaking-style classification, or unified analysis. Unified responses include a transcript, calibrated labels, time-local segments, and tagged text. The current public contract is file based, not streaming.
The API serves two tiers today: Spectra 1 for efficient affect workloads and Resonance, the flagship, for the strongest transcription and unified analysis. Spectra 2 is listed in the catalog as a preview tier but is not yet serving traffic.
The API scores 15 emotion labels against thresholds calibrated on held-out audio. A clip may return several emotions. If no emotion crosses its threshold, the highest-scoring emotion is returned as a fallback. Long audio also includes segment-level labels so applications can follow changes over time.
Speaking style is a separate 16-label multilabel output. A clip may have multiple styles or no selected style. Emotion and style can be requested independently, together through the affect endpoint, or alongside a transcript through unified analysis.
Yes. Resonance is the recommended model for English transcription. The same API can also return acoustic emotion and speaking-style labels, which preserves useful information that a transcript alone does not contain.
Developers create a verified account, receive organization approval, and generate a production API key in the developer portal. The v1 API uses bearer authentication and multipart file uploads. Keys are displayed only once.
API v1 supports English. Inputs may be mono or stereo and use WAV, FLAC, MP3, M4A, OGG, or WebM containers with common sample rates. oruk does not currently advertise multilingual production support for this API.
Inference is priced per audio minute and metered by measured second, with a one-second minimum. Each response includes the billable duration, rate, estimated cost, and pricing version. The developer portal shows the immutable credit ledger and request-level usage history.