This project provides a Chrome extension for searching video content by transcribing audio to text, enabling users to find specific words or phrases within videos. It is targeted at content creators, researchers, and anyone who needs to quickly locate information within video lectures, interviews, or other spoken-word content. The primary benefit is efficient content retrieval, saving significant time compared to manual searching.

How It Works

The system leverages OpenAI's Whisper model (via faster-whisper for speed improvements) to perform audio-to-text transcription. A Flask API runs locally, utilizing the user's GPU for processing. The extension then allows users to input search queries, highlighting exact matches in green and phonetically similar matches (using Levenshtein distance) in purple on the YouTube timeline. Results are stored in JSON for faster subsequent searches on the same URL.

Quick Start & Requirements

• Install by loading the extension folder into Chrome using the "load unpacked" function.
• Run python app.py in the pythons folder to start the Flask API.
• Requires a GPU (e.g., RTX 3090 for ~16s/5min video).
• Python environment with necessary dependencies (e.g., faster-whisper).

Highlighted Details

• Utilizes OpenAI Whisper for accurate audio transcription.
• GPU acceleration via a local Flask API for faster processing.
• Highlights exact and phonetically similar matches on the YouTube timeline.
• Caches search results for immediate retrieval on subsequent searches.

Maintenance & Community

The project is presented as a personal side project by the author, Evan Wildenhain, with an open invitation for others to build upon it. No specific community channels or active maintenance plans are detailed.

Licensing & Compatibility

The README does not explicitly state a license. The project's open nature suggests a permissive intent, but the absence of a LICENSE file requires clarification for commercial or closed-source use.

Limitations & Caveats

Currently, the extension only supports YouTube videos. The initial search may not highlight the timebar while the model is processing; a second search will display cached results instantly. Phonetic similarity matching is noted as "pretty much useless" due to Whisper's effectiveness.