Unmute

• 

  Final logo in both light and dark

• 

“Unmute” was born from a simple but powerful question: How can we give a true voice to people who cannot speak for themselves?

For autistic and nonverbal individuals, existing AAC (Augmentative and Alternative Communication) tools are rigid, static, and often feel clinical. We wanted to design something more human — a system that listens, learns, and adapts.

The tagline says it all:

Because everyone deserves to be heard.

📚 What We Learned

Working on Unmute taught us several key lessons:

AI must be adaptive, not prescriptive.

Traditional AAC tools map buttons → phrases. But human expression is nonlinear.

With gpt-oss, we could model intent as a probability distribution:

𝑃 ( phrase ∣ context , user_history ) P(phrase∣context,user_history)

This means the model can anticipate phrases more naturally.

Accessibility design is human-first.

We tested colors, typography, and iconography to ensure calmness and clarity.

Blue gradients and microphone imagery conveyed trust + communication.

Offline AI agents matter.

Not everyone has reliable internet access.

By experimenting with a local GPT-OSS agent, we learned to make Unmute useful in offline-first contexts.

🛠️ How We Built It

The build came together in three phases:

Prototype Backend

Node.js app with dotenv for safe key management.

GPT-OSS for phrase prediction.

NVIDIA Speech AI for natural-sounding voice synthesis.

Frontend + Branding

React Native app mockups for mobile.

Logo + banner design (microphone + soundwaves + blue underglow).

Tagline integrated everywhere to keep mission front and center.

Testing + Packaging

Created .env.example and .gitignore for open-source hygiene.

README structured for judges with install/run instructions.

Asset pack with logos + banners for repo polish.

⚔️ Challenges We Faced

Time pressure ⏳: We had to prioritize core features vs. polish.

API key safety 🔐: Protecting secrets while keeping repo public.

Model fine-tuning 🧠: GPT-OSS is powerful, but tailoring it for AAC-specific communication is an ongoing challenge.

Balancing empathy + tech ❤️🤖: It’s easy to over-engineer; harder to make tools that feel truly human.

🚀 Next Steps

Fine-tune GPT-OSS on AAC datasets to make predictions more accurate.

Expand emotion detection so Unmute doesn’t just say words, it says them with the right tone.

What's Next Partner with accessibility orgs (like KultureCity) to test in real communities.

Built With

• gpt-os
• nativex
• nvidia
• react