audi-tory

Notes-to-audio web app built on React and AWS. PDF or plain text in, MP3 out in ~20–30 seconds.

Live App: https://audi-tory.vercel.app

Table of Contents

• Usage
• Quick Start (Local)
• App Overview
• Backend Pipeline

---

Usage

Visit https://audi-tory.vercel.app and:

1. Upload a PDF or paste your notes as plain text
2. Choose a style, length, and voice
3. Hit Generate Audio — takes ~20–30 seconds
4. Play or save the MP3

Styles

Style	What it does
Core Concepts	Identifies the key concepts in your notes and explains each one in depth
Podcast	Rewrites your notes as a natural, conversational audio script
Readback	Reads your notes back clearly and faithfully, no expansion

Lengths

Length	Target	Best for
Short	~200 words	Quick review, sparse notes
Medium	~800 words	Standard study session
Long	~1800 words	Deep dive, dense notes

Voices

Voice
Matthew	Masculine
Stephen	Masculine
Joanna	Feminine
Ruth	Feminine

---

Quick Start (Local)

To run this locally:

Backend

git clone https://github.com/EshanPrakash/audi-tory.git
cd audi-tory/backend
pip install -r requirements.txt

Create a .env file in the project root:

AWS_REGION=us-east-1
S3_BUCKET=your-s3-bucket
BEDROCK_MODEL_ID=us.anthropic.claude-haiku-4-5-20251001-v1:0
LOCAL_OUTPUT_PATH=/path/to/output.mp3

Run the pipeline locally with sample notes:

python test_pipeline.py
python test_pipeline.py --notes black_holes --style podcast --length medium
python test_pipeline.py --pdf /path/to/notes.pdf --style concepts --length short
python test_pipeline.py --voice Joanna

Frontend

cd frontend
npm install
npm run dev

Create frontend/.env:

VITE_API_URL=https://your-api-gateway-url/prod/process

Visit http://localhost:5173.

To deploy the backend (Lambda + API Gateway):

cd backend
./deploy.sh

---

App Overview

What You Can Do

audi-tory takes your notes and turns them into a generated audio file:

• PDF or plain text: upload a PDF or paste notes directly — both go through the same pipeline
• Three output styles: Core Concepts breaks your notes into explained topics, Podcast rewrites them conversationally, Readback reads them faithfully as-is
• Three lengths: Short (~200 words) for quick review, Medium (~800 words) for standard study, Long (~1800 words) for deep coverage
• Four neural voices: Matthew and Stephen (masculine), Joanna and Ruth (feminine)

Performance

• Processing time: ~20–30 seconds end to end (Bedrock + Polly async)
• Audio expiry: presigned S3 URL valid for 1 hour after generation
• Input limits: 5MB max for PDFs, 50,000 characters max for plain text
• Concurrency: serverless — scales automatically with demand

Features

• Fully serverless — no backend server to maintain
• Rate limiting: 2 req/sec, burst of 5
• Auto-deletes generated audio from S3 after 24 hours

---

Backend Pipeline

The core pipeline lives in backend/pipeline.py and runs as four steps:

1. extract_text_from_pdf — extracts raw text from an uploaded PDF using pypdf
2. generate_script — sends the text to Claude Haiku via Bedrock with a style/length prompt, returns a clean spoken-word script
3. synthesize_audio — submits the script to Polly as an async neural TTS task, polls until complete, stores MP3 in S3
4. get_download_url — generates a presigned S3 URL valid for 1 hour

Plain text input skips step 1 and goes directly to step 2.

---

License

MIT License - see LICENSE file.