Oruk

oruk is a file-based Speech API that accepts prerecorded English audio and returns transcription, emotion detection, speaking-style classification, or a unified response combining all three. The core workflow is a POST request against one of three endpoints — transcription, affect, or unified analysis — with the model tier specified in the request. The unified endpoint returns a transcript, calibrated multilabel scores for emotion and style, time-local segments, and tagged text in a single response, which means teams building downstream annotation or review tooling get everything in one round trip rather than chaining separate calls.

The differentiating feature is multilabel calibration on held-out audio. Rather than outputting a single dominant emotion class, the API returns scores for up to 15 emotion labels simultaneously — a clip can carry fear and surprise at the same time, and the scores are calibrated against out-of-sample data rather than raw softmax outputs. The vendor’s July 2026 benchmark report shows per-class F1 scores across all seven mapped emotion classes, with the Spectra model posting results materially above Gemini 3 Flash Preview on disgust, fear, surprise, and anger specifically. These figures come from the vendor’s own evaluation methodology; the report documents its limitations and the dataset sources used.

Two models are actively serving traffic: Resonance, described as the flagship for highest-accuracy transcription and unified analysis, and Spectra 1, positioned for efficient affect workloads where transcript output can be skipped. A third model, Spectra 2, is listed in the catalog as a preview tier but the docs state it is not yet serving inference traffic. The scope is deliberately narrow — English, file-based, prerecorded audio only. No self-hosted deployment is available, and the vendor page shows no streaming contract in the current API v1 specification.