VKontakte Speech Recognition Engines
summary
Describes VKontakte’s speech recognition engines, comparing online and offline systems by naming the neural network architectures used and explaining how they are trained and adapted for products. Covers practical enhancements and extra modules that can be added to improve recognition.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:29ASR use cases
- 3:57Buy vs build
- 5:49ASR engine components
- 8:42Linguistic models
- 10:31Training ASR systems
- 14:05Online vs offline recognition
- 16:57Quality improvement approaches
- 18:03Acoustic events and confidence
- 22:41Punctuation placement
- 23:44Pipeline performance
- 24:48Model selection tips
- 25:10Q&A
- 37:13Conclusion
related talks
42mSpeech recognition for subtitles in VK Video
45mA Pipeline to Transcribe Speech in Millions of Videos per Day
46mFeed and Recommendation Architecture at VKontakte
36mVoice Enabling Mobile Apps and Custom Hardware Devices
22mOn-Device Speech Recognition for Mobile Devices and Custom Hardware
56mSpeech Recognition: How to Build Speech-to-Text Yourself
06mSensing to Hear: Speech Enhancement for Mobile Devices Using Acoustic Signals
46mIntroduction to Speech Recognition Systems Through a Tester’s Eyes