A Pipeline to Transcribe Speech in Millions of Videos per Day
summary
Description of a pipeline used at VK to transcribe speech from millions of videos per day, covering how the team scaled speech recognition from a voice-message feature to other services. The talk discusses handling differing infrastructures, requirements and approaches across services while improving recognition quality and deploying the transcription system at large scale.
AI-written summary · about
chapters (23)
- 0:00Introduction
- 1:08Original task and goals
- 2:34Existing infrastructure
- 4:04Subtitling challenges
- 5:51Deployment issues
- 7:14Pipeline design differences
- 8:18Queue for millions
- 10:03Processing cluster setup
- 12:11Preprocessing audio
- 13:15Speech recognition
- 16:28Subtitle generation
- 17:31Subtitle formats
- 18:35Model deployment
- 19:41Pipeline launch results
- 20:22Testing real content
- 21:48Confidence filtering
- 23:54Performance issues
- 25:20Caching subtitle links
- 27:48Current metrics
- 28:53Future plans
- 29:36Public API
- 30:37Key takeaways
- 31:03Q&A
related talks
31m“Sorry, Could you repeat that again?” - Speech Recognition with Python
28mEnhance speech recognition accuracy using custom language models
35mTricks Learned from Scaling WhisperSpeech Models to 80k+ Hours of Speech
01h 34mReinventing Speech-to-Text Transcriptions with Go and Whisper
33mMaximize lifetime value of media content
24mTowards Large-scale Speech Analytics Systems
20mMeta's AI Translations Inference
29mUsing Amazon Transcribe to make content searchable and accessible