Speech Emotion Recognition on Open-Source Orukeet

Taught an open-source speech recognizer to hear how something is said, not just what is said.

Company: FYI.AI
Role: ML engineering
Period: 2026

PyTorchHugging Face TransformersFastAPIDockerNVIDIA H100Next.js

01 / At a glance

77.4%Accuracy on speakers never heard in training

+3.8 ptsAhead of emotion2vec+, the best open-source baseline

0.2 sTo analyse a 10–15 s clip on GPU

460 hEmotional speech from 50+ public corpora

02 / One clip in, three answers out

16Emotions, each with a confidence score

3Delivery scales: energy, positivity, dominance

25Languages transcribed with word timestamps

03 / How it was built

01Data
- 460 h of speech
- 50+ corpora
- ~20 languages
- Automated quality checks
02Train
- Orukeet encoder
- Emotion + delivery together
- Best two generations merged
03Evaluate
- Unseen speakers only
- Public benchmarks
- Leakage audited
04Ship
- GPU API
- 0.2 s per clip
- Dashboard UI

04 / Benchmarks

This project Other models

Emotion benchmark · 5,000 unseen clips

Orukeet v3.177.4%
Orukeet v4 (production)76.6%
emotion2vec+ large73.6%

IEMOCAP · positive / neutral / negative

Orukeet v274.4%
Orukeet v4 (production)73.9%
Orukeet v3.172.0%
Oruk Resonance 171.5%

CREMA-D · 6 emotions, held-out actors

Orukeet v279.6%
Orukeet v3.174.3%
Oruk Resonance-261.7%
Oruk Fourier55.7%

Accuracy, %. Orukeet models are scored on speakers never heard in training. In the 5,000-clip test every model ran on the same clips; the other IEMOCAP and CREMA-D numbers are the vendors' published results.