VideoCaptioner — AI Subtitle Pipeline

What it is

VideoCaptioner is an open-source desktop application that automates the video subtitle workflow. It chains Whisper-based speech recognition, LLM-powered text correction, translation into 99 languages, and styled subtitle export (SRT, ASS, VTT) into a single pipeline. The tool provides a GUI for configuring each stage.

VideoCaptioner targets content creators, video editors, and localization teams who need accurate subtitles across languages. It handles the full lifecycle from raw audio to production-ready subtitle files without manual transcription or separate translation tools.

How it saves time or tokens

Manual subtitle creation involves transcribing audio, correcting recognition errors, translating to target languages, and formatting subtitle files. VideoCaptioner automates all four stages. The LLM correction step catches Whisper's common errors (proper nouns, technical terms, homophones) without human review. Batch processing handles multiple videos sequentially.

How to use

1. Download the latest release from GitHub or clone the repo and install dependencies with pip install -r requirements.txt.
2. Run python main.py to open the GUI. Configure your Whisper model and LLM API settings.
3. Load a video file, select source and target languages, and start the pipeline.

Example

# Clone and set up
git clone https://github.com/WEIFENG2333/VideoCaptioner.git
cd VideoCaptioner
pip install -r requirements.txt
python main.py

# Pipeline stages:
# 1. Whisper transcribes audio to text
# 2. LLM corrects transcription errors
# 3. Translation to selected languages
# 4. Export as SRT, ASS, or VTT with styling

Related on TokRepo

• Video AI Tools — AI-powered video production tools
• Content AI Tools — Content creation and processing tools

Common pitfalls

• Whisper accuracy depends heavily on audio quality. Background music, overlapping speakers, and low-quality microphones reduce transcription accuracy.
• LLM correction requires API access (OpenAI or compatible). Without it, you get raw Whisper output which may contain errors for domain-specific vocabulary.
• Translation quality varies by language pair. Common pairs (English-Spanish, English-Chinese) produce better results than less common language combinations.