DEV Community

Anubhav Batham
Anubhav Batham

Posted on

AI Ghost Story Generator: Bringing Spooky Lore to Life with Open-Weight LLMs & Horror Voice Narration 👻

#devchallenge #hf26challenge

## What I Built
AI Ghost Story Generator is an open-source, full-sensory storytelling application that transforms ideas, nightmares, and folklore into spine-chilling ghost stories in both English and Hindi.
For the "Ghost in the Machine" theme, I wanted to interpret the prompt both literally and artistically: rather than treating AI as a sterile text-completion box, we breathed an eerie, theatrical spirit directly into the machine.
When you share a premise—such as a forgotten railway station in Shimla, an old mansion on a cliff, or an ancient heirloom curse—the application delivers a complete haunted experience:

  1. Open-Weight AI Generation: Leverages local open-weight models (llama3.2:3b via Ollama) to craft atmospheric horror tailored across 7 distinct categories (Haunted House, Ancient Curse, Urban Legend, Psychological Horror, Folk Ghost Story, Haunted Forest, Emotional Tale).
  2. Horror Voice Narration: Reads the generated story aloud in an ominous, low-pitch horror cadence with sentence-level playback controls (Play, Pause, Resume, Stop, Replay) and authentic Hindi (Devanagari) pronunciation.
  3. Haunted Atmosphere Soundscape: Procedurally generates subtle howling wind and a 55Hz sub-bass drone directly in the browser via the Web Audio API—with 0 external audio assets.

4. Living Haunted Scenery: A fully animated gothic world featuring floating spirits, a walking ghost pacing across the screen with synchronized footsteps, a glowing full moon, drifting fog, and arcane summoner runes.

## Demo

- 💻 Local Full-Fidelity Experience: Run locally with Ollama for 100% offline, private AI generation.

## Code

GitHub logo anubhav9236 / ai-ghost-story-generator

An AI-powered ghost story generator with spooky animations, floating ghosts, and open-source AI using Flask and Ollama.

AI Ghost Story Generator 👻

"Turn Your Imagination into Terrifying Ghost Stories!"

License: MIT Python: 3.10+ Flask: 3.x Ollama: llama3.2:3b Tests: Passing

AI Ghost Story Generator is a complete, full-stack open-source web application designed for creative storytelling and AI hackathons. It transforms simple seeds, eerie locations, nightmares, or strange childhood memories into rich, original ghost stories using open-weight AI models executed locally through Ollama (llama3.2:3b), framed inside a cinematic gothic user interface with visible, animated ghost characters.

📖 Table of Contents

  1. Project Overview & Problem Statement
  2. Key Features
  3. Haunted Design & Ghost Animations
  4. Technology Stack
  5. How AI Generation Works
  6. Why Open-Weight AI Matters & Privacy
  7. Screenshots & Visuals
  8. Prerequisites
  9. Installation & Setup
  10. Running the Application
  11. Running Automated Tests
  12. Troubleshooting Guide
  13. Project Structure
  14. Deployment Considerations
  15. Contributing & Future Enhancements
  16. License

🌌 Project Overview & Problem Statement

The Problem

Most AI writing applications either rely on expensive, proprietary cloud APIs with recurring subscription fees, store…

View on GitHub

GitHub Repository: https://github.com/anubhav9236/ai-ghost-story-generator

_- **License: _**MIT Open Source

## How I Built It**_
The project was architected from the ground up around open-source AI and browser-native multimedia standards:

1. Open-Source AI Engine (Ollama & Llama 3.2 3B)

  • Local Open-Weight Inference: Powered by Meta's open-weight llama3.2:3b running through Ollama via its local REST API (http://127.0.0.1:11434/api/generate).
  • Bilingual Horror Prompt Engineering: The Flask backend provides fine-tuned system instructions tailored to horror tropes in both English and Hindi. The Hindi prompt enforces authentic cultural vernacular and Devanagari script without transliteration artifacts.
  • Dynamic Token Budgeting: Token limits (num_predict) dynamically scale between Short (~200 words), Medium (~400 words), and Long (~700 words) to guarantee responsive generation times even on CPU inference. ### 2. Horror Voice Narration Engine (Web Speech API)
  • Built on the browser's native window.speechSynthesis API.
  • Sentence-Level Chunking Queue: Chrome and WebKit browsers have a known quirk where speech synthesis silently cuts off on utterances exceeding ~250 characters. We implemented a queue worker that splits stories at sentence boundaries ([.!?।\n]), maintaining continuous playback with instant Pause/Resume/Replay support.
  • Atmospheric Modulation: Narration pitch is lowered (pitch: 0.72) and rate moderated (rate: 0.85) to evoke the cadence of an old-school campfire storyteller. ### 3. Procedural Haunted Atmosphere (Web Audio API)
  • To avoid heavy copyrighted sound files, the eerie audio is generated 100% procedurally:

_ - **Howling Wind:_** Pink noise synthesized into an AudioBufferSourceNode, filtered through a dynamic BiquadFilterNode modulated by an LFO (Low-Frequency Oscillator) to create the rise and fall of winter gales.

  • Sub-Bass Horror Drone: A pure 55Hz sine-wave oscillator combined with a 110Hz warm undertone that introduces subtle infrasonic tension into headphones. ### 4. Spooky Visual & CSS Animation System
    • Walking Ghost: A dual-animation keyframe system (walkingGhostJourney + walkingGhostStep) that traverses the viewport horizontally while bobbing vertically with authentic stride physics.

    - Floating Spirits & Scenery: Pure SVG ghosts with glowing halos, blinking pupils, a moonlit horizon, drifting multi-layered mist, and rotating arcane summoning rings.

## Why Does Open Innovation Matter?
Open innovation is what made this project possible, and it represents the future of creative artificial intelligence:

  1. No Token Tolls or API Paywalls: Proprietary LLM APIs charge per token and impose rate limits. Running open-weight models locally democratizes immersive digital storytelling for students, writers, and independent creators without a credit card.
  2. Uncensored Creative Horror Fiction: Closed commercial APIs frequently trigger overzealous safety filters when handling horror themes, supernatural death, curses, or psychological tension. Open-weight models like Llama 3.2 allow authentic, evocative gothic fiction while respecting artistic nuance.
  3. Absolute Privacy: Your nightmare ideas, folklore prompts, and creative drafts never leave your device. There is no telemetry, no tracking, and no data harvesting.

4. Interchangeability & Longevity: Because the backend communicates with Ollama's standardized interface, developers can freely swap llama3.2:3b for mistral, qwen2.5, or custom fine-tuned horror models with zero code rewrites.

## My Agent Session
This project was developed through an agentic pair-programming workflow:

  • Architected the full-stack Flask and vanilla CSS/JS pipeline.
  • Solved Windows UTF-8 console output challenges for Hindi Devanagari text.
  • Overcame the Web Speech API 250-character cutoff limitation through a custom queue worker.

- Synthesized browser-native Web Audio procedural atmospheric effects to achieve 0 KB audio bundle overhead.

## The Spirit of It

"उस रात हवेली की खिड़कियों से कोई रोशनी नहीं आ रही थी, फिर भी अंदर से कदमों की आवाज़ सुनाई दे रही थी..."

The moment this project truly landed was testing the Hindi horror generation late at night. Hearing the machine narrate an eerie folklore story in Hindi, accompanied by the gentle 55Hz sub-bass drone and the howling wind whistling in the background while little ghosts floated across the screen, felt genuinely magical. It proved that open-source AI isn't just about spreadsheets and code completion—it can evoke genuine wonder, nostalgia, and a good old-fashioned spine shiver.

## Prize Categories

  • Ghost in the Machine (Primary Track)
  • Best Use of Open-Source Models / Local AI (Ollama + Llama 3.2)
  • Most Creative / Atmospheric User Experience

Top comments (0)