←
open.intel1,648 posts
open.intel avatar

open.intel

@OpenAtIntel

Everything #opensource at Intel. We have a lot to share and a lot to learn. Join us.

◷ Joined October 2021▧ 373 Photos and videos
open.intel avatar
open.intel@OpenAtIntel·16 Jul 2025

Meet Lucas Melo, Cloud Solutions Architect at Intel and a driving force behind Intel Optimization Cloud Modules for Terraform and Ansible. His work turns @intel’s cloud optimizations into accessible, actionable code for the developers who need it most. This is open source with purpose. #OpenSource #MaintainerSpotlight

Post attachment
open.intel avatar
open.intel@OpenAtIntel·1 Jul 2025

EmoNet isn’t just another benchmark. It’s a new baseline for how we train AI to interpret emotion, not flatten it. It makes space for: Real evaluation, not leaderboard chasing Emotional signals across languages, cultures, and contexts Models that respond like they’re listening, not just parsing This is groundwork for building AI that feels less robotic and more human-aware.

open.intel avatar
open.intel@OpenAtIntel·1 Jul 2025

It gets better: we didn’t stop at classification. With BUD-E Whisper, we're going beyond transcription to emotion captioning. The model detects emotional tone, vocal bursts (sighs, laughter), and speaker traits from raw speech.

open.intel avatar
open.intel@OpenAtIntel·1 Jul 2025

Why synthetic data? No privacy issues Full demographic control Massive scalability More ethical training sets This is how we generated over 203K facial images and 5K hours of speech with real, measurable emotional intent.

open.intel avatar
open.intel@OpenAtIntel·1 Jul 2025

And we built models to match. The Empathic Insight-Face model beats Gemini 2.5 Pro and Hume AI on our facial emotion benchmarks. It tracks closely with human annotators across 40 emotions.

Post attachment
open.intel avatar
open.intel@OpenAtIntel·1 Jul 2025

Emotion is complex. So we didn’t settle for “basic” emotions. EmoNet’s taxonomy includes states like: shame, doubt, pride, fatigue, teasing, even intoxication It’s built on the Handbook of Emotions and refined with psychologists.

open.intel avatar
open.intel@OpenAtIntel·1 Jul 2025

We’ve done it differently with EmoNet-Voice: 5,000+ hours of synthetic voice-acted emotion 40 emotion categories 4 languages Openly licensed Expert-verified benchmark for real evaluation In so many vital ways, a major step forward.

open.intel avatar
open.intel@OpenAtIntel·1 Jul 2025

Let’s start with the problem. Here’s what most speech emotion datasets look like: Tiny datasets Coarse labels ("happy," "sad") Narrow speaker demographics No multilingual support Often closed license, which makes follow-up and iterative improvements difficult to say the least.

Post attachment
open.intel avatar
open.intel@OpenAtIntel·1 Jul 2025

It’s called EmoNet: a new open source benchmark + model suite for emotion recognition in speech and facial expressions. This is emotion recognition designed the right way round for once: starting with nuance, not reduction.

open.intel avatar
open.intel@OpenAtIntel·1 Jul 2025

Emotion isn't binary. It’s messy. Contradictory. Cultural. And we’re finally building AI that can see and hear that.

open.intel avatar
open.intel@OpenAtIntel·30 Jun 2025

OPEA megaservices don’t do it all—they make it all happen. They run GenAI workflows by coordinating microservices with a blueprint. Modular, efficient and built to scale. Learn more: intel.ly/43Ws5zA

Post attachment
open.intel avatar
open.intel@OpenAtIntel·25 Jun 2025

What does it take to turn a research idea into a production-ready open source project? Mona Vij did exactly that with Gramine. At Intel Labs, she leads groundbreaking work in Confidential Computing and trusted execution—from cloud to edge. #OpenSource #MaintainerSpotlight

Post attachment