Gemini: Slashing Subtitle Time & Cost with AI
summary
Claims production workflow that turns 1 hour of video into subtitles in under 5 minutes of processing time and cuts costs about 90% vs speech-to-text APIs and ~99% vs manual transcription while avoiding $1/hr API bills. Uses Gemini 2.5 Flash on extracted audio stored in GCS, splits audio into 5-minute chunks with 30s overlap, asks Gemini to emit structured JSON (start/end/speaker/text), then runs a second pass to break segments over 8s and formats SRT timestamps.
AI-written summary · about
chapters (15)
- 0:00Subtitles for accessibility
- 1:02Traditional subtitle challenges
- 3:12Why speech-to-text APIs weren't enough
- 4:35Generative AI for accuracy
- 5:43Choosing Gemini for subtitling
- 6:44AI subtitle workflow
- 7:50Audio over video for subtitles
- 8:35Overcoming output token limits in LLMs
- 10:25Format and structure fixes
- 11:49Timestamp and repetition fixes
- 14:01Readability with two-pass processing
- 16:31Chunking long subtitles
- 18:40Technical takeaways
- 19:46AI beyond transcription
- 20:08Results: Time and cost savings
related talks
01h 16mI'm in my Gemini Era!
18mIntroduction to Gemini APIs and AI Studio
19mGemini 2.0: Beyond the Basics
01h 17mLet's go Bananas with GenMedia
15mBuilding with the Gemini API and AI Studio
54mGemini, Google's Large Language Model
01h 45mAI Engineering with the Google Gemini 2.5 Model Family
09mDeveloper Keynote recap – Richard Seroter