TalkLess: Blending Extractive and Abstractive Speech Summarization for Editing Speech to Preserve...
summary
TalkLess combines extractive cuts and LLM-driven abstraction to shorten speech while keeping voice style, generating 25 candidates per segment and scoring them for compression, edits, insertions, and content preservation before a speech editing model applies audio edits. Evaluation on four target compressions (15, 25, 50, 75%) shows higher content coverage and more disfluency removal than a state-of-the-art extractive baseline; a user study (N=12) reports reduced cognitive load and editing effort. The UI exposes a compression pane, outline view, and an edit-types visualization to focus attention on high-risk content changes.
AI-written summary · about
chapters (13)
- 0:00Introduction to TalkLess
- 0:27Problem with speech editing
- 1:32Limitations of extractive summarization
- 2:16Abstractive summarization challenges
- 2:59TalkLess approach overview
- 3:41Automatic speech editing pipeline
- 5:05Scoring and selecting best candidates
- 5:26Editing interface features
- 6:49User study results
- 7:52Editor user study findings
- 8:55Creator preference study
- 9:17Future work directions
- 9:58Conclusion and Q&A
related talks
38mThe Art of Compression: Crafting Insightful Summaries with LLMs
07mJust Speak It: Minimize Cognitive Load for Text Editing with a Smart Voice Assistant
17mBuilding Effective Voice Agents
05mUser Preferences for Automated Curation of Snackable Content full strip notes
49mAutomatic Summarization of 10K Daily Meetings: from Requirements to Production
20mDemo Session
21mA Human-in-the-Loop System for Sound Event Detection and Annotation
19mUnlocking the Full Potential of LLMs: Advanced Prompt Engineering Techniques