Cardigan：面向 PBS 媒体管理的智能转录与元数据生成工具

Cardigan is an open-source tool designed specifically for PBS Media Manager, combining large language model (LLM) capabilities in automatic transcription, intelligent keyword extraction, and smart editing workflows. It addresses the bottlenecks of manual metadata management in PBS, improving content processing efficiency and ensuring compliance with PBS's strict metadata standards.

In PBS and other large media institutions, manual transcription and metadata annotation are time-consuming, labor-intensive, and prone to inconsistencies. As video content volume explodes, this manual approach becomes a critical bottleneck. Cardigan was created to solve this pain point by integrating AI-assisted workflows to enhance efficiency and quality.

Cardigan's core functions include three modules:

1. Auto-transcription: Optimized for broadcast content (professional terms, multi-speaker scenarios) with timestamp annotations.
2. Smart keyword extraction: Uses LLM semantic understanding to extract explicit and implicit keywords (e.g., "greenhouse effect" from climate change content even if not directly mentioned).
3. Agent-Assistant workflow: AI provides initial results, while human editors retain final decision-making. It supports incremental editing and learns editor preferences for better future recommendations.

Cardigan integrates deeply with PBS Media Manager. Its generated metadata directly meets PBS's standards (title, description, keywords, categories), eliminating manual data conversion steps. Content teams can import metadata directly into PBS Media Manager without extra formatting, especially beneficial for batch processing large content volumes.

Technical architecture:

• Modular design: Core transcription engine supports multiple backends (open-source or cloud APIs). Keyword extraction uses LLM embedding vectors for semantic similarity.
• Workflow engine: State-machine driven task management (tracks status: pending, transcribing, reviewing, completed) with concurrency and exception handling.
• Web UI: Real-time collaboration, keyboard shortcuts, batch operations, and version history comparison for efficient editing.

Application scenarios: News (fast transcription for timely release), documentaries/education (enhanced discoverability via metadata). Value: Reduces hours of work to minutes; shifts editors from repetitive tasks to quality control; improves accessibility (subtitles, audio descriptions) and SEO for content.

Cardigan is open-source on GitHub with a permissive license, allowing customization for other media institutions. Future plans: Multi-language support, real-time transcription for live content, advanced AI features (auto-summary, sentiment analysis, entity linking).

Cardigan exemplifies successful AI application in media content management. It combines LLM capabilities with industry-specific workflows to create a practical solution. For media organizations facing similar challenges, Cardigan serves as a valuable reference implementation.