MV Subtitle Generator

Landing Page (light theme)
MP3/MP4 File Upload Page
Subtitle Anomaly Review
Inline editing of subtitles
GIF
Subtitle Review and Edit in Two Ways

Inspiration

A smart transcriber for audio and video files that can not only generate text transcripts/subtitles, but also reviews and corrects text, so that the subtitles are free of any errors - homophones, accent misrecognition, incorrect punctuation, and more. Correct subtitles are crucial for accessibility as well as accurate audio/video summarization.

What it does

This application generates subtitles in SRT (SubRip Subtitle) format for MP3 and MP4 files of size up to 25MB using Gemini 3 models. Auto-generated subtitles often are incorrect, with errors like use of homophones, accent misrecognition, incorrect punctuation, and more. Therefore, this application:

Reviews the generated subtitles and auto-corrects any errors that it has a high confidence of correcting. It refers to the optional context, if provided by user, for the review and auto-correction. It also enables users to review those auto-corrections.
Walks users through the potential errors it is not confident to correct, providing correction suggestions.
Enables users to manually correct subtitle text, including auto-corrected text, in an interactive review synced to audio/video play.

How we built it

With Antigravity:

Frontend: React, TanStack Query, ShadCN UI, Tailwind CSS.
Backend: Node.js (Express).
AI/ML: Google Gemini 3 API (gemini-3-flash-preview)

Challenges we ran into

Model rate limits
Frequent model unavailability

Accomplishments that we're proud of

Smart transcription generation, auto-correction, and user review
Enterprise-grade UX
No user data retained

What we learned

AI Model Resilience - Implement robust retry logic early. Added a 3-attempt retry loop with a 10-second countdown. We also expanded error catching to include generic TypeError and "fetch" related messages, which significantly increased the application's perceived stability.
LLM Behavior Control - The LLM initially made stylistic "improvements" (changing valid text) instead of just fixing errors, and its explanation format was inconsistent. It requires control by explicit prompts to "not make stylistic changes" and enforcing a strict format for auto-correction details Negative constraints are as important as positive ones to achieved predictable, high-quality results.
Decouple build logic - Separating the build into a dedicated script/build.ts and refactoring vite.config.ts to be more modular made the Docker image stable and compatible across different hosting environments.
Security - Documenting how the GOOGLE_API_KEY is handled securely (server-side only) builds user trust.