AI subtitle generation model for Korean educational content at Hiconsy

Advanced speech recognition system fine-tuned for Korean educational content using Whisper Large-v3 with LoRA

Korean Educational ASR System Development

This repository is kept private due to Hiconsy’s confidentiality requirements.
However, you can review the project results and documentation via my guidelines for subtitle reviewers here.

Project Documentation

📄 Interim Report - Model Comparison Test (PDF)

Overview

Led the end-to-end development of a Korean Automatic Speech Recognition (ASR) system for educational content of 시대인재 at Hiconsy, diving deep into the entire ML pipeline from model fine-tuning to deployment. This project became my gateway into advanced AI engineering and sparked my passion for MLOps.

What I Built

  • Fine-tuned Whisper Model: Adapted OpenAI Whisper Large-v3 for Korean educational lectures using LoRA techniques
  • Custom Tokenization Pipeline: Developed domain-specific 8,000-vocabulary SentencePiece tokenizer optimized for Korean educational terminology
  • 4-Stage RAG System: Implemented Retrieval-Augmented Generation pipeline for intelligent post-processing
  • Distributed Training Setup: Configured multi-GPU environment with DeepSpeed ZeRO-2 across 6x RTX 4090 GPUs

Key Learnings & Results

Through this project, I gained hands-on experience in:

  • Model Engineering: PyTorch, Transformers, PEFT, LoRA implementation
  • Infrastructure Management: DeepSpeed optimization, Flash Attention configuration
  • NLP Engineering: Korean embeddings, FAISS vector search, custom tokenizers
  • Automation & MLOps: Google Apps Script for evaluation pipelines, Tensorboard monitoring
  • Project Leadership: Coordinating review processes and managing development workflows

Measurable Impact:

  • 275% improvement in segmentation precision (833 → 2,288 segments)
  • 55.6% efficiency gain with custom tokenizer
  • 2.3x faster processing through domain optimization
  • 68.3% reduction in average segment length (5.02s → 1.59s)

Technologies Explored

Core Stack: PyTorch, Transformers, PEFT, LoRA, DeepSpeed
NLP Tools: Korean sentence transformers, FAISS, SentencePiece
Infrastructure: Multi-GPU training, Flash Attention 2.5.8
Post-processing: Several LLM models like OpenAI’s GPT, Google Gemini.

What This Project Taught Me

This experience transformed my understanding of AI engineering and ignited my curiosity about MLOps, distributed systems, and production-ready ML pipelines. It showed me how theoretical knowledge translates into solving real-world problems – exactly the kind of work I want to pursue as I continue growing in this field.