An Innovative Approach to Labeling Ground Truth in Speech
summary
VersaTag GUI replaces CLAN and feeds audio through a speech activity detector, presents one acoustic “utterance” at a time, and forces categorical dropdowns for noise, language and speaker metadata so transcribers stay consistent. Two-transcriber comparisons show miss detections drop from ~7% and ~14% to ~5% each and false alarms fall from ~12% and ~6% to ~4% and ~5%. Iteration with student taggers and a wearable called Starling target first a word-count release then quality metrics.
AI-written summary · about
chapters (7)
- 0:00Introduction and motivation
- 2:16Need for ground truth labeling
- 3:39Initial labeling challenges
- 7:55Developing Bertha Tag
- 10:40Results and improvements
- 12:48Future goals and wearable device
- 14:19Q&A session
related talks
31mCreate text classifiers in a few hours using the open-source Label Sleuth system
22mDeep learning module recognizes video sounds
29mLabeling Tools are Great, but What About Quality Checks?
29mAnnotating data the right way
34mScaling Trust
01h 07mDecoding Dolphin Speech
53mArtificial vs Natural Intelligence in Annotation Tasks
40mHow Not to Develop an "umm" Detector Using Create ML