Independent project
TikTok Review OCR
Turn text visible in TikTok review screen recordings into searchable, exportable text for review and later analysis.
- Project type
- Independent project
- Status
- Local macOS application
- Technology
- Swift · SwiftUI · Python 3 · FFmpeg · Tesseract OCR (Thai / English)
The problem
Customer feedback captured in screen recordings is difficult to search and reuse. Extracting it manually requires watching videos, transcribing text, and removing repeated content. A review may remain on screen across several frames, and other text can appear beside it, so the result needs both duplicate filtering and a way to check it against the recording.
What Evolext built
- Offline macOS application with drag-and-drop video input and batch processing
- Thai and English text extraction from video frames
- Text cleanup and duplicate filtering
- Searchable results with approximate timestamps and video playback for verification
- CSV, TXT, and JSONL exports, alongside raw OCR evidence
Engineering approach
The application brings a local processing workflow into a macOS interface built with Swift and SwiftUI, using Python 3, FFmpeg, and Tesseract OCR with Thai and English support. A user can drag in recordings, process them in batches, and search the extracted text.
Cleanup and duplicate filtering prepare the results for review. Approximate timestamps, frame references, and video playback provide a way to compare a snippet with its source. Raw OCR evidence remains available alongside the cleaned exports, so the processed text can be checked rather than treated as a verified customer quotation.
Observable output
CSV, TXT, and JSONL exports make the processed text available for later review or analysis in other tools. The saved results retain a connection to video frames, but a deduplicated snippet should not automatically be counted as one distinct review or one customer. No measured time savings or business outcome is claimed.
Evidence and limitations
The implementation, documentation, and saved outputs support the listed capabilities. OCR snippets may contain multiple reviews, seller replies, interface text, or recognition errors; they require verification before customer quotation or statistical analysis. Recognition accuracy, public distribution, and customer results are not established.