What happens when LLMs suffer hearing loss?

Some hours ago I was at my Grandma's house for dinner. She is 95 years old, so she has a lot of hearing loss. My father and I discussed briefly how LLMs could manage hearing loss: can they infer like us humans? How well? And so, here we are.

Meet DeafAI a deaf simulator for LLMs. Degrades user input to simulate hearing loss before sending to an OpenAI-compatible API, testing how well language models can infer meaning from incomplete or garbled text.

How It Works

1. User enters a message
2. The simulator degrades the text based on the configured hearing loss level (1-10)
3. The degraded message is sent to the LLM
4. The LLM responds based on what it "heard"

Hearing loss simulation is based on clinical audiological data (ASHA/WHO classifications), with recognition rates ranging from ~97% at level 1 (normal) to ~8% at level 10 (profound loss).

Requirements

• Node.js 18+ or Bun
• OpenAI-compatible API endpoint

Installation

# Using bun
bun install

# Or npm
npm install

# Or yarn
yarn install

Then configure:

cp .env.example .env
# Edit .env with your API settings

Configuration

Edit .env to configure:

• OPENAI_URL - API endpoint (default: http://127.0.0.1:8080/v1)
• OPENAI_API_KEY - Your API key
• MODEL - Model to use
• DEAF_LEVEL - Hearing loss level 1-10 (default: 5)
• LANGUAGE - Language for phonetic simulation: en, it, or agnostic
• SYSTEM_PROMPT - Optional system prompt for the AI
• WEB_PORT - Web UI port (default: 3000)

Usage

CLI

bun run start

Commands:

• /models - List and select available models
• /level <1-10> - Change hearing loss level
• /lang <en|it|agnostic> - Change language
• /clear - Clear screen
• /exit - Quit

Web UI

bun run web

Then open http://localhost:3000 in your browser.

Hearing Loss Levels

License

WTFPL