Build Your Own Dreams: Browser-based custom narrative generation

[jeremymanning]

Overview

Enable users to create personalized dream narratives that run entirely in the browser, with no server dependencies.

User Stories

1. As a user, I want to describe my ideal dream scenario in a text box, so I can experience a personalized guided dream.
2. As a user, I want my custom dreams saved locally, so I can reuse them without regenerating.
3. As a user, I want the generation to work offline after initial load, so I can create dreams anywhere.

Technical Architecture

Client-Side Stack (All in-browser)

• LLM: SmolLM or similar lightweight model via WebLLM/Transformers.js
• TTS: Web Speech API or Piper TTS (WASM)
• Music: Existing procedural synthesis (already client-side capable)
• Storage: IndexedDB for saving generated dreams

Generation Pipeline

User Input → LLM (narrative sections) → Add [PAUSE] markers → TTS → Mix with music → Save to IndexedDB

Implementation Phases

Phase 1: MVP - Text-to-Narrative

• Integrate SmolLM via WebLLM for narrative generation
• Create prompt templates for dream narrative style
• UI: Simple text input + "Generate" button
• Generate narrative in sections (intro, scenes, transitions, outro)
• Auto-insert [PAUSE] markers between sections
• Display generated narrative for review/editing

Phase 2: Audio Generation

• Integrate browser TTS (Web Speech API as fallback, Piper WASM for quality)
• Port music generation to client-side (already uses numpy-like operations)
• Combine narration + music in browser (Web Audio API)
• Export as playable audio blob

Phase 3: Persistence & Polish

• Save custom dreams to IndexedDB
• Show custom dreams in library alongside pre-made dreams
• Allow editing/regenerating sections
• Share custom dreams (export/import JSON)

Key Technical Decisions

Decision	Choice	Rationale
LLM Runtime	WebLLM	Best performance for in-browser LLMs
Model	SmolLM-360M or similar	Small enough for mobile, good enough for narratives
TTS	Piper WASM (primary), Web Speech API (fallback)	Quality vs. compatibility tradeoff
Storage	IndexedDB	Large audio file support, offline access

Complexity Estimate

• Phase 1: Medium (2-3 weeks) - LLM integration is the main challenge
• Phase 2: High (3-4 weeks) - Audio pipeline in browser is complex
• Phase 3: Low (1 week) - Standard CRUD operations

References

• SmolLM demo: https://context-lab.com/llm-course/demos/chatbot-evolution/
• WebLLM: https://github.com/mlc-ai/web-llm
• Piper TTS: https://github.com/rhasspy/piper

---

Split from #115 - targeted dream content (pre-made narratives completed)