A Human-in-the-Loop System for Sound Event Detection and Annotation
summary
A human-in-the-loop system splits long audio into fixed-length segments matching an example, extracts features (handcrafted or embeddings), ranks segments by similarity and presents small batches (typically five) for listen/label/ boundary-adjust feedback; labels feed back to re-rank. A 20-participant study on 12-minute files (11 classes, 18 events/class) shows the interface lets users label target events about twice as fast as manual annotation, but logs reveal substantial interaction cost — estimated 4.68 mean interaction overhead and an idealized up-to-4x possible speedup with zero overhead.
AI-written summary · about
chapters (17)
- 0:00Intro
- 1:11What Kind of Sound Event
- 1:38Manual Annotation
- 2:24Why not using automated system
- 3:36Proposed approach
- 4:06Two goals
- 4:47System overview
- 5:57User feedback
- 8:21Interface
- 10:30Evaluation
- 11:13Results
- 12:53Interaction
- 13:35Future Work
- 16:32Survey Results
- 16:50Demonstration
- 18:17Assumptions
- 19:04Target Audio
related talks
10mVOICE 2021 | Daria Baidakova, Toloka
12mIncreasing the Speed and Accuracy of Data Labeling Through an AI Assisted Interface
32mEvaluating multi-turn conversations
49mMachine Learning 2.0 With Hugging Face
13mDxHF: Providing High-Quality Human Feedback for LLM Alignment with Interactive Decomposition
17mImproving Automated Email Tagging with Implicit Feedback
17mThe Voice-First AI Overlay: Designing Conversational Co-Pilots
20mEmploying AI to extract actionable insights from qualitative user...