Voice Search Using NLP & CoreML
summary
A working on-device voice search for Kiwi.com that turns live speech into structured queries using Apple's Speech framework, NSDataDetector for dates, the Natural Language tagger and a tiny custom model built with Create ML and exported to CoreML. The surprising claim: a usable prototype needs only about six training sentences (the app uses ~30 sentences in practice) to tag places, dates and direction; ambiguous origin/destination is solved by direction tags and simple heuristics, and English transcription works reliably on device.
AI-written summary · about
chapters (10)
- 0:00Introduction to voice search
- 1:15Motivation for voice search
- 4:33Speech recognition setup
- 8:57Parsing destination from text
- 13:02Extracting dates from speech
- 15:57Identifying origin and destination
- 18:11Introducing machine learning
- 23:39Training the CoreML model
- 26:27Expanding model capabilities
- 28:20Q&A session
related talks
31mPower-Up Your App With Apple’s Machine Learning Frameworks
17mExploring Natural Language Processing
36mVoice Enabling Mobile Apps and Custom Hardware Devices
47mBuilding smarter apps with Machine Learning, from magic to reality
12mMachine learning for Accessibility
38mML Kit: Machine Learning for Mobile with Firebase
40mKeynote: Open-source Multimodal AI in the Wild
27mApplying ML at scale for New User Experiences