Ai Podcast Transcription | AI Final Year Project Report

Complete Technical Paper Content — Ai Podcast Transcription

1. Executive Abstract, Problem Formulation & Background

🎙️ Podcast AI Intelligence Platform Podcast AI Intelligence Platform Transform long podcast audio files into clean, interactive transcripts, semantic chapters, and deep AI insights. 📖 Overview The Podcast AI Intelligence Platform is a state-of-the-art, commercial-grade AI software designed to digest hours of raw audio and render structured, easily readable summaries, interactive transcripts, and thematic breakdowns. ❓ The Problem & Why It Exists With the explosion of long-form audio podcasts, video conferences, and webinars, consuming hours of speech has become a major time bottleneck. Searching for specific details, extracting action items, or indexing key topics is manual and inefficient. This platform leverages modern Speech-to-Text models (OpenAI Whisper) and Advanced Natural Language Processing (NLP) pipelines to structure audio semantically, making speech instantly searchable, conversational, and actionable. 👥 Who Can Use It & Real-World Use Cases Content Creators & Podcasters : Automatically generate show notes, timestamps, SEO keywords, and social summaries. Researchers & Students : Read through lecture or interview transcripts rapidly, run semantic queries, and extract direct quotes. Enterprise Teams : Turn team meetings, workshops, and product briefings into bulleted action items and indexable knowledge repositories. SaaS Integrators : Power search indexation engines with structured JSON exports representing full timeline timelines. ✨ Features 🎙️ Audio Processing Broad Format Support : Seamlessly ingest .mp3 , .wav , .m4a , .ogg , .flac , and .webm files. Drag-and-Drop Uploader : Highly intuitive drag-and-drop dashboard supporting large file streams up to 500MB. Inline Playback : Built-in media player for monitoring and listening to source files. 📝 AI Transcription OpenAI Whisper Engine : Sub-word level alignment for accurate word-by-word transcription. Speaker Identification : Segments speaker blocks to distinguish conversational flows. Interactive Search : Real-time keyword filtering directly inside the transcript. Time-Sync Seeking : Clicking any transcript timestamp jumps the audio player to that exact time. 📚 Semantic Chapters Automated Topic Segmentation : Advanced sentence clustering algorithms divide the timeline into distinct chapters. Dynamic Cards : Collapsible cards containing localized segment summaries, keyword tags, and duration metrics. Read More Toggle : Clean summary layouts that collapse long paragraphs to maintain page readability. 🧠 AI Insights Executive Summary : Generates a high-level overview of the entire episode. Key Takeaways & Action Items : Bulleted key concepts and actionable next-steps for rapid comprehension. Important Quotes : Auto-extracts noteworthy quotes with accurate citations. Entity Mapping : Lists people, organizations, and technologies mentioned in the audio. 💬 Interactive AI Chat ChatGPT-Style Assistant : Context-aware assistant powered by LLM models that references the transcript database. Suggested Prompts : Single-click prompts to quickly index insights. Markdown & Code Support : Full markdown formatting rendering with distinct code blocks and styling. 📦 Enterprise Exports PDF Report : Download polished, print-ready reports with professional formatting. JSON Payload : Access raw developer payloads containing timeline vectors, keywords, and structural meta tags. Markdown & TXT : Export structured summaries ready for Obsidian, Notion, or plain text indexers. 📸 Screenshots Dashboard View Transcript View Semantic Chapters AI Chat Assistant 🛠️ Tech Stack Layer Component / Tool Details Frontend React 19, TypeScript, Vite Fast, typed, component-driven client. Styling & UI Vanilla CSS, Lucide Icons, Framer Motion Smooth 200ms transitions, premium dark layout. Backend FastAPI, Python 3.10 High-performance async API backend. Speech Engine OpenAI Whisper High-accuracy speech-to-text transcription. NLP Pipeline Sentence Transformers, KeyBERT, NLTK, spaCy Sentence embedding extraction, keyword modeling, and topic clustering. Inference API Groq Llama-3 API Low-latency inference for chat and summaries. Utilities FFmpeg, ReportLab Audio transcoding and professional PDF generation. 🏗️ System Architecture graph TD A[Frontend React Client] -->|Upload Audio File / Link| B(FastAPI Server) B -->|Transcode / Downsample| C[FFmpeg Preprocessor] C -->|Output Mono 16kHz WAV| D[OpenAI Whisper Transcription] D -->|Raw Sentences & Timestamps| E[NLP Analysis Pipeline] E -->|1. Sentence Embeddings| F[Sentence-Transformers] E -->|2. Keyword Extraction| G[KeyBERT / spaCy] F -->|Clustering & Vector Shift| H[Topic Segmentation] H -->|Semantic Chapter Blocks| I[Groq LLM Llama3] I -->|Chapter Summaries & Key Insights| J[Export Engine] J -->|PDF / JSON / Markdown Output| A Loading 📂 Folder Structure Podcast-AI-Intelligence-Platform/ │ ├── core/ # Core Python Backend Pipeline │ ├── audio_loader.py # YouTube audio downloader (yt-dlp integration) │ ├── preprocess.py # Audio conversion, chunking, and FFmpeg interface │ ├── transcription.py # OpenAI Whisper API or local model transcriber │ ├── embeddings.py # Sentence-Transformers vectorizer │ ├── topic_segmentation.py # Text tiling, topic clustering, and keyphrase extractor │ └── exporter.py # PDF & JSON data report builders │ ├── frontend/ # React TypeScript SPA (Vite) │ ├── src/ │ │ ├── api/ # API client connections (Axios endpoints) │ │ ├── components/ # Shared components (AudioPlayer, EmptyState) │ │ ├── contexts/ # Application state providers (Analysis, Audio, Settings) │ │ ├── layouts/ # Page shell layouts (AppLayout, Sidebar) │ │ ├── pages/ # View pages (Dashboard, Transcript, Chapters, Chat, Exports) │ │ ├── types/ # TypeScript interface definitions │ │ ├── utils/ # Formatter and array helpers │ │ ├── App.tsx # Main router routing setup │ │ ├── index.css # Global premium dark theme CSS configuration │ │ └── main.tsx # Application mounting file │ ├── package.json # Frontend package scripts and dependencies │ └── vite.config.ts # Vite bundling settings │ ├── outputs/ # Exported PDFs and JSON datasets ├── temp_uploads/ # Temporary uploads storage directory ├── main.py # API launch configuration (FastAPI server) ├── app.py # Deprecated Streamlit setup (Alternative UI) ├── requirements.txt # Python environment packages └── README.md # System documentation ⚙️ Installation & Setup Prerequisites Python 3.9 or higher Node.js v18 or higher FFmpeg installed and added to your System PATH variables. 1. Clone the Repository git clone https://github.com/Anjal08/Podcast-AI-Intelligence-Platform.git cd Podcast-AI-Intelligence-Platform 2. Backend Setup Create and activate a Python virtual environment: python -m venv venv # Windows venv \S cripts \a ctivate # macOS/Linux source venv/bin/activate Install the backend Python packages: pip install -r requirements.txt Set up your .env configuration file (see Environment Variables below). Launch the FastAPI server: python main.py The backend will run on http://localhost:8000 . 3. Frontend Setup Open a new terminal window and navigate to the frontend directory: cd frontend Install npm dependencies: npm install Launch the Vite development server: npm run dev The application will be accessible at http://localhost:5173 . 🔑 Environment Variables Create a file named .env in the root workspace directory matching the variables in this .env.example : # Server Configuration HOST = 127.0.0.1 PORT = 8000 # API Credentials GROQ_API_KEY = gsk_your_groq_api_key_goes_here # Paths Configuration UPLOAD_FOLDER = temp_uploads OUTPUT_FOLDER = outputs # Processing Configuration MODEL_NAME = sentence-transformers/all-MiniLM-L6-v2 📡 API Endpoints FastAPI exposes the following swagger endpoints. Documentations can be viewed live at http://localhost:8000/docs . Method Endpoint Description POST /api/analyze Accepts multipart form data audio file, creates analytical task. GET /api/status/{task_id} Polls processing stage index, progress logs, and final transcript results. POST /api/chat/{task_id} Processes conversational chat queries relative to transcript context. GET /api/download/pdf/{task_id} Downloads compiled executive report and transcript timeline as PDF. GET /api/download/json/{task_id} Exports vector-processed JSON metadata. GET /api/health Returns backend online connection and service checks. ⚙️ How It Works (The AI Pipeline) The system transforms raw audio into semantic chapters through an automated multi-step computational pipeline: [Audio Upload] ──> [FFmpeg Preprocess] ──> [Whisper AI Transcription] ──> [Sentence Tokenization] │ ▼ [PDF/JSON Export] <── [Groq LLM Synthesis] <── [Topic Segmentation] <── [Vector Embeddings] Upload & Preprocessing : Audio is sent to the backend where FFmpeg normalizes the signal to a mono 16kHz WAV format (the optimal rate for Speech-to-Text). Transcription (Whisper) : OpenAI's Whisper engine parses audio frames, yielding sentences, words, speaker segments, and millisecond timestamps. Sentence Embeddings : Tokenized sentences are converted into high-dimensional vector representations using Sentence-Transformers . Semantic Clustering : The system calculates cosine similarity shifts between consecutive sentences to identify "boundaries" where topics drift, segmenting the timeline into logical chapters. Insights Generation : Segment blocks and relevant keywords are synthesized by Llama-3 models via Groq API, producing clean chapter names, takeaways, action items, and summaries. 📈 Performance Low-Latency Inference : Integrates Groq Llama-3 processing to generate AI highlights and summaries in milliseconds. Efficient UI Rendering : Optimized React layout structures with clean states preventing unnecessary renders. Fast Transcription : Optimized audio chunking pipelines feeding Whisper engines cleanly. 🔮 Future Improvements Speaker Diarization : Integrate PyAnnote to tag speaker voices with custom speaker names automatically. Vector Database : Integrate Qdrant or Milvus to search across multiple historic podcasts semantically. Multi-language Support : Transcribe non-English podcasts and auto-translate outputs. User Authentication : Add secure login layers, session keys, and persistent user profile accounts. Live Audio Recording : Transcribe and analyze microphone inputs in real-time. 🚀 Deployment Frontend (Vercel) Install Vercel CLI or link your repository to Vercel . Set the build command to npm run build and output folder to dist . Set the environment variable VITE_BACKEND_URL to point to your live hosted FastAPI address. Backend (Render) Create a new Web Service on Render . Use Python environment, set build command to pip install -r requirements.txt , and start command to uvicorn main:app --host 0.0.0.0 --port $PORT . Add the GROQ_API_KEY to the environment settings. Ensure Render has system FFmpeg binaries loaded (via custom Docker builds if using default templates). 🤝 Contributing Contributions make the open-source community an amazing place to learn, inspire, and create. Fork the project. Create your feature branch ( git checkout -b feature/AmazingFeature ). Commit your changes ( git commit -m 'Add some AmazingFeature' ). Push to the branch ( git push origin feature/AmazingFeature ). Open a Pull Request. 📄 License Distributed under the MIT License. See LICENSE for more information. ✍️ Author Anjali Patel GitHub: @Anjal08 Specialization: AI, NLP & Full-Stack Development 🙏 Acknowledgements OpenAI Whisper FastAPI Framework Groq AI Engine Sentence Transformers KeyBERT React & Vite Ecosystem

Transform long podcast audio files into clean, interactive transcripts, semantic chapters, and deep AI insights.

The Podcast AI Intelligence Platform is a state-of-the-art, commercial-grade AI software designed to digest hours of raw audio and render structured, easily readable summaries, interactive transcripts, and thematic breakdowns.

With the explosion of long-form audio podcasts, video conferences, and webinars, consuming hours of speech has become a major time bottleneck. Searching for specific details, extracting action items, or indexing key topics is manual and inefficient.

This platform leverages modern Speech-to-Text models (OpenAI Whisper) and Advanced Natural Language Processing (NLP) pipelines to structure audio semantically, making speech instantly searchable, conversational, and actionable.

Podcast-AI-Intelligence-Platform/ │ ├── core/ # Core Python Backend Pipeline │ ├── audio_loader.py # YouTube audio downloader (yt-dlp integration) │ ├── preprocess.py # Audio conversion, chunking, and FFmpeg interface │ ├── transcription.py # OpenAI Whisper API or local model transcriber │ ├── embeddings.py # Sentence-Transformers vectorizer │ ├── topic_segmentation.py # Text tiling, topic clustering, and keyphrase extractor │ └── exporter.py # PDF & JSON data report builders │ ├── frontend/ # React TypeScript SPA (Vite) │ ├── src/ │ │ ├── api/ # API client connections (Axios endpoints) │ │ ├── components/ # Shared components (AudioPlayer, EmptyState) │ │ ├── contexts/ # Application state providers (Analysis, Audio, Settings) │ │ ├── layouts/ # Page shell layouts (AppLayout, Sidebar) │ │ ├── pages/ # View pages (Dashboard, Transcript, Chapters, Chat, Exports) │ │ ├── types/ # TypeScript interface definitions │ │ ├── utils/ # Formatter and array helpers │ │ ├── App.tsx # Main router routing setup │ │ ├── index.css # Global premium dark theme CSS configuration │ │ └── main.tsx # Application mounting file │ ├── package.json # Frontend package scripts and dependencies │ └── vite.config.ts # Vite bundling settings │ ├── outputs/ # Exported PDFs and JSON datasets ├── temp_uploads/ # Temporary uploads storage directory ├── main.py # API launch configuration (FastAPI server) ├── app.py # Deprecated Streamlit setup (Alternative UI) ├── requirements.txt # Python environment packages └── README.md # System documentation

git clone https://github.com/Anjal08/Podcast-AI-Intelligence-Platform.git cd Podcast-AI-Intelligence-Platform

2. Related Work, Literature Survey & Algorithmic Foundations

python -m venv venv # Windows venv \S cripts \a ctivate # macOS/Linux source venv/bin/activate

Create a file named .env in the root workspace directory matching the variables in this .env.example :

# Server Configuration HOST = 127.0.0.1 PORT = 8000 # API Credentials GROQ_API_KEY = gsk_your_groq_api_key_goes_here # Paths Configuration UPLOAD_FOLDER = temp_uploads OUTPUT_FOLDER = outputs # Processing Configuration MODEL_NAME = sentence-transformers/all-MiniLM-L6-v2

FastAPI exposes the following swagger endpoints. Documentations can be viewed live at http://localhost:8000/docs .

The system transforms raw audio into semantic chapters through an automated multi-step computational pipeline:

[Audio Upload] ──> [FFmpeg Preprocess] ──> [Whisper AI Transcription] ──> [Sentence Tokenization] │ ▼ [PDF/JSON Export] <── [Groq LLM Synthesis] <── [Topic Segmentation] <── [Vector Embeddings]

Contributions make the open-source community an amazing place to learn, inspire, and create.

An end-to-end GenAI pipeline for podcast transcription, topic segmentation, and RAG-based Q&A

3. Dataset Ingestion, Preprocessing & Stochastic Augmentations

Automated Podcast Transcription and Topic Segmentation 🧩 Problem Statement Podcasts are increasingly popular, but long audio episodes are difficult to navigate. Manual transcription and topic-wise timestamp creation is time-consuming and inefficient. This project builds an automated system that converts podcast audio into structured, searchable, topic-wise knowledge using GenAI . 🎯 Project Objective To design and implement an end-to-end automated pipeline that transforms podcast audio into: Accurate text transcription Semantically segmented topics with timestamps Human-like chapter titles for easy navigation Context-aware question answering from podcast transcripts 🧩 Solution Overview The system is a Django-based backend application that provides a one-click automated pipeline. It performs audio preprocessing, speech-to-text transcription, semantic topic segmentation, and question answering , executed using Groq-powered GenAI . User Workflow: Upload podcast audio (any format supported; .wav recommended) Audio preprocessing (noise reduction, silence removal, VAD) Speech-to-text transcription using Whisper Semantic topic segmentation with AI-generated titles (Groq API) Ask natural language questions from the generated transcription Download final outputs: transcription.txt topic_segmented_timestamps.txt 🏗️ System Architecture The system follows a modular, pipeline-based architecture: Frontend: Audio upload, question input, and result download Backend: Django-based orchestration Processing Modules: Audio preprocessing Transcription Topic segmentation Question answering Evaluation 📌 The complete workflow diagram is available in the assets/ folder. ⚙️ Tech Stack Layer Technology Backend Django ASR OpenAI Whisper (Small) NLP Sentence Transformers GenAI Groq API (llama-3.3-70b-versatile) Audio Processing Librosa, Silero VAD Evaluation JiWER, Scikit-learn 🧩 Core Modules Module Responsibility preprocessing/ Audio cleaning, noise reduction, silence removal, VAD transcription/ Speech-to-text transcription segmentation/ Semantic topic segmentation & title generation (Groq-based) question.py Question answering from podcast transcription pipeline.py End-to-end pipeline orchestration evaluation.py Quality and performance evaluation 📄 Outputs / Results The system generates two user-downloadable files: transcription.txt – Complete podcast transcript topic_segmented_timestamps.txt – Topic-wise timestamps with titles Additionally, users can interactively query the podcast content using natural language questions. 📊 Evaluation Metrics 🔹 Transcription Quality WER: 0.561 CER: 0.462 🔹 Topic Segmentation Performance Topic Coherence: 0.849 Semantic Boundary Alignment: 1.0 Total Predicted Topics: 8 🔹 GenAI Usage ASR Model: Whisper Small LLM Model: llama-3.3-70b-versatile (via Groq API) Purpose: ASR output evaluation Topic title generation 🔹 Safety Handling Hybrid execution (Local ASR + Cloud LLM) No user data storage Cloud API used: Groq API 🔹 Cost Awareness ASR Model: Whisper Small Reason: Low GPU usage, zero ASR API cost Inference Cost: Low (local ASR) + API-based (LLM) 🔹 Code Quality Modular pipeline design Reusable components Readable and maintainable code 🔹 Documentation & Explainability Clear docstrings Inline comments Stepwise, explainable pipeline Transparent evaluation metrics 🚫 Not in Scope Real-time transcription Speaker diarization Live podcast streaming 🌱 Stretch Goals Multilingual transcription support Advanced Whisper models (medium / large) Search functionality inside transcripts 🔐 Safety & Privacy Secure API-based LLM usage No persistent storage of user data API keys managed via environment variables 📘 Project Resources All project resources are available in the assets/ folder , including: 📘 Full Project Report (PDF): View Report 📊 Project Presentation (PPT): View Presentation 📊 Project Workflow : View Presentation 📎 Project Video Explanation Google Drive (Demo & Resources): https://drive.google.com/file/d/1RAtWC6xAEqP-cBFJroqcZYl9zd2NP0ZY/view?usp=drivesdk 👥 Contributor Ismail Sk ML / NLP / Backend Developer 📜 Internship Context This project was developed as an individual project under the Infosys Springboard Internship 6.0 program. ⭐ If you find this project useful, consider starring the repository!

Podcasts are increasingly popular, but long audio episodes are difficult to navigate. Manual transcription and topic-wise timestamp creation is time-consuming and inefficient. This project builds an automated system that converts podcast audio into structured, searchable, topic-wise knowledge using GenAI .

To design and implement an end-to-end automated pipeline that transforms podcast audio into:

The system is a Django-based backend application that provides a one-click automated pipeline. It performs audio preprocessing, speech-to-text transcription, semantic topic segmentation, and question answering , executed using Groq-powered GenAI .

Additionally, users can interactively query the podcast content using natural language questions.

📎 Project Video Explanation Google Drive (Demo & Resources): https://drive.google.com/file/d/1RAtWC6xAEqP-cBFJroqcZYl9zd2NP0ZY/view?usp=drivesdk

This project was developed as an individual project under the Infosys Springboard Internship 6.0 program.

Podcasts are increasingly popular, but long audio episodes are difficult to navigate. Manual transcription and topic-wise timestamp creation is time-consuming and inefficient. This project builds an **automated system that converts podcast audio into structured, searchable, topic-wise knowledge using GenAI**.

4. Neural Network Backbone, Layer Specifications & Loss Formulations

For 43 years, I struggled with bad credit due to my own poor decisions, and my credit score was around 490. When my girlfriend and I decided to buy a house, a mortgage broker informed us that it would be impossible to secure a mortgage with my credit score. As a result, she referred me to a company called HACK MAVENS CREDIT SPECIALIST, assuring me of their professionalism and ability to assist with credit improvement. Upon contacting them, I was impressed by their professionalism and they assured me that they could help. In less than 6 days, my credit score skyrocketed to 785, and they also successfully resolved issues in my credit report, including the bankruptcy. I am incredibly satisfied with their service and would highly recommend HACK MAVENS CREDIT SPECIALIST for reliable credit repairs. You can reach them at H A C K M A V E N S 5 [AT] G M A I L [DOT] COM or at [+] [1] [2 0 9] [4 1 7] – [1 9 5 7]. Thanks to their help, my girlfriend and I are now proud homeowners.

AUTOMATED PODCAST TRANSCRIPTION AND TOPIC SEGMENTATION USING AI 1. Introduction Podcasts are a popular way to share information and knowledge, but they are often very long. It can be difficult and time-consuming for users to find specific topics in a podcast. This project helps by automatically converting podcast audio into text, identifying different topics, generating summaries and keywords, and making it easier for users to find the information they need quickly. 2. Requirement Gathering People often find it difficult to locate specific information in long podcast episodes. Most podcast platforms only allow users to listen to the audio and do not provide easy ways to search or navigate topics. This project solves the problem by converting podcast audio into text, dividing it into topics, generating summaries, and making the content easy to search and access. 3. Objective Definition Main Objective To develop an AI-powered system that automatically transcribes podcast audio recordings and segments them into meaningful topics for efficient navigation and content discovery. Specific Objectives To convert podcast audio into text using speech recognition technology. To preprocess audio files by reducing noise and improving audio quality. To identify topic boundaries within the generated transcript using NLP techniques. To extract important keywords from each topic segment. To generate concise summaries for identified topics. To provide a search mechanism for locating specific information within transcripts. To develop an interactive dashboard for displaying transcripts, topics, keywords, and summaries. To improve accessibility and user experience by making podcast content searchable and easier to understand. 4. User & Module Identification 1. User Management Module This module allows users to register, log in, and securely access the system. 2. Audio Upload Module This module enables users to upload podcast audio files in MP3 and WAV formats. 3. Audio Preprocessing Module This module improves audio quality by reducing noise and normalizing audio levels. 4. Speech-to-Text Module This module converts podcast audio into text using Whisper AI. 5. Topic Segmentation Module This module identifies topic boundaries and divides transcripts into meaningful sections. 6. Keyword Extraction Module This module extracts important keywords from each topic segment. 7. Summary Generation Module This module generates concise summaries for identified topics. 8. Search and Navigation Module This module allows users to search transcript content and navigate directly to specific topics. 9. Dashboard and Visualization Module This module displays transcripts, summaries, keywords, and topics through an interactive dashboard. 10. Database Management Module This module stores and manages all project-related information. 5. Database Requirement Analysis The system must support the following operations: User Management Users can register and log in. Store user details: user_id (Primary Key) name email (Unique) password (Encrypted) Podcast Management Users can upload podcasts. Each podcast contains: podcast_id (Primary Key) title audio_file duration upload_date Transcript Generation Each podcast generates a transcript. Transcript includes: transcript_id (Primary Key) transcript_text Relationship: One Podcast → One Transcript Relationship Handling One User can upload Multiple Podcasts (1:M) One Podcast generates One Transcript (1:1) Integrity Requirements Primary Keys user_id podcast_id transcript_id topic_id keyword_id Unique Constraints email must be unique Foreign Keys podcast.user_id → user.user_id transcript.podcast_id → podcast.podcast_id topic.transcript_id → transcript.transcript_id keyword.topic_id → topic.topic_id Relationship Constraints User → Podcast = One-to-Many (1:M) Podcast → Transcript = One-to-One (1:1) Transcript → Topic = One-to-Many (1:M) Topic → Keyword = One-to-Many (1:M) 6. ER Diagram Design Insert the ER Diagram image here. docs/ERDiagram.png Technology Stack Frontend ReactJS HTML CSS Bootstrap Backend Spring Boot Java Database MySQL AI & NLP Python Whisper AI NLTK spaCy Development Tools VS Code IntelliJ IDEA GitHub Postman Project Features User Registration and Login Podcast Audio Upload Audio Preprocessing Speech-to-Text Conversion Topic Segmentation Keyword Extraction Summary Generation Transcript Search Interactive Dashboard Transcript Download Project Workflow User Login/Register ↓ Upload Podcast Audio ↓ Audio Preprocessing ↓ Speech-to-Text Conversion ↓ Topic Segmentation ↓ Keyword Extraction ↓ Summary Generation ↓ Dashboard Visualization ↓ Search and Download Transcript Current Progress Author DEVADHARSHINI S Register Number: 922524148016 Department: Computer Science and Engineering Institution: V.S.B. Engineering College, Karur

Podcasts are a popular way to share information and knowledge, but they are often very long. It can be difficult and time-consuming for users to find specific topics in a podcast. This project helps by automatically converting podcast audio into text, identifying different topics, generating summaries and keywords, and making it easier for users to find the information they need quickly.

People often find it difficult to locate specific information in long podcast episodes. Most podcast platforms only allow users to listen to the audio and do not provide easy ways to search or navigate topics. This project solves the problem by converting podcast audio into text, dividing it into topics, generating summaries, and making the content easy to search and access.

To develop an AI-powered system that automatically transcribes podcast audio recordings and segments them into meaningful topics for efficient navigation and content discovery.

This module identifies topic boundaries and divides transcripts into meaningful sections.

This module allows users to search transcript content and navigate directly to specific topics.

This module displays transcripts, summaries, keywords, and topics through an interactive dashboard.

5. Training Protocol, Convergence Behavior & Experimental Evaluation

This project aims to automatically transcribe podcast audio recordings, identify topic boundaries, generate summaries, and provide topic-wise navigation using AI and NLP techniques.

Automated Podcast Transcription & Topic Segmentation Automated Podcast Transcription & Topic Segmentation is an audio analysis system that transforms long-form podcast audio into structured, searchable, and navigable insights . The project focuses on making spoken content easier to understand, explore, and reuse by automatically identifying topics, summaries, keywords, and timestamps from raw audio. Overview Podcasts contain valuable information, but their linear nature makes them difficult to browse. This project addresses that problem by converting audio into organized knowledge units that can be consumed programmatically or through a user interface. At a high level, the system: Converts audio into text Identifies topic boundaries Generates summaries and keywords Aligns insights with timestamps Exposes the results through APIs What This Project Does Transcribes podcast audio into readable text Segments long transcripts into meaningful topic sections Generates high-quality abstractive summaries (BART) for each topic Extracts semantic keywords (KeyBERT) for search and filtering Preserves timestamps to enable audio navigation Stores processed data in a structured and reusable format Use Cases Podcast platforms with chapter-based navigation Audio content analysis and indexing Educational or research-focused podcast exploration AI-powered search over spoken content Backend support for transcript-based user interfaces System Architecture (High Level) Podcast Audio Upload ↓ Backend Upload API (Node.js + Multer) ↓ Python Transcription Pipeline (Vosk ASR) ↓ Transcript Cleaning & Chunking ↓ Topic Segmentation (NLP) ↓ Segment Summarization (BART-Large-CNN) & Keyword Extraction (KeyBERT) ↓ Timestamp Alignment (Start–End per Segment) ↓ Structured JSON Generation ↓ MongoDB Storage (Podcasts & Segments) ↓ REST APIs (Fetch Podcasts, Segments, Search) ↓ Frontend Dashboard (React) ↓ Podcast Details View (Search & Keyword Filter) Output Format Each podcast is converted into a structured representation containing: Topic-wise transcript segments Summaries for quick understanding Keywords for search and filtering Start and end timestamps for each segment This structure makes the data suitable for: Frontend applications Search engines Further AI-based analysis Backend Services The project includes a backend service that: Stores podcast and segment data Provides REST APIs for retrieval Supports keyword-based and text-based filtering Enables easy integration with user interfaces Installation & Setup Node.js ≥ 18 Python ≥ 3.9 (3.11.x Recommended) MongoDB (local or Atlas) FFmpeg (must be available in system PATH) Backend Setup cd backend npm install npm run dev Python Pipeline Setup python -m venv .venv source .venv/bin/activate # Windows: .venv\Scripts\activate pip install -r requirements.txt python scripts/download_vosk_model.py # Download Vosk Model Frontend Setup cd frontend npm install npm run dev Design Principles Modular – Each stage of processing is independent Scalable – Designed to handle long-form audio content Extensible – Easy to add new analysis or models UI-ready – Output designed for direct frontend usage Future Enhancements Interactive transcript viewer Timestamp-based audio navigation Semantic search over podcast content AI-assisted question answering on podcasts Cloud deployment and scaling Acknowledgements This project builds upon modern advancements in: Speech recognition Natural language processing Audio analytics Backend API design

Automated Podcast Transcription & Topic Segmentation is an audio analysis system that transforms long-form podcast audio into structured, searchable, and navigable insights .

The project focuses on making spoken content easier to understand, explore, and reuse by automatically identifying topics, summaries, keywords, and timestamps from raw audio.

Podcasts contain valuable information, but their linear nature makes them difficult to browse. This project addresses that problem by converting audio into organized knowledge units that can be consumed programmatically or through a user interface.

Podcast Audio Upload ↓ Backend Upload API (Node.js + Multer) ↓ Python Transcription Pipeline (Vosk ASR) ↓ Transcript Cleaning & Chunking ↓ Topic Segmentation (NLP) ↓ Segment Summarization (BART-Large-CNN) & Keyword Extraction (KeyBERT) ↓ Timestamp Alignment (Start–End per Segment) ↓ Structured JSON Generation ↓ MongoDB Storage (Podcasts & Segments) ↓ REST APIs (Fetch Podcasts, Segments, Search) ↓ Frontend Dashboard (React) ↓ Podcast Details View (Search & Keyword Filter)

python -m venv .venv source .venv/bin/activate # Windows: .venv\Scripts\activate pip install -r requirements.txt python scripts/download_vosk_model.py # Download Vosk Model

One API to every speech-to-text model worth using. Compare side by side, route to the best, pay by the second.