Research · NewsShorts-Bench · Stage 6: Active speaker detection
Active speaker detection on split screens: specialists near chance
AVA-trained detectors (TalkNet, LR-ASD, Light-ASD, LRX-ASD) pick the right half 55-60% of the time, against 50% chance. Gemini 4 Argon gets 94.8%, and 7 of 7 when the same person is on both sides.
- Data
- 116 pairs
- Compute
- 1x H100 80GB
- Chance
- 50%
- Metric
- Pairs right %
Split-screen pairs with the right speaker (%) · higher is better · dot = measured · bar = 95% CI
Reframing a 16:9 news shot for a 9:16 phone means choosing which face to keep, and the obvious choice is the person talking. Active speaker detectors do exactly that, and the well-known ones score 92–95% mAP on the AVA-ActiveSpeaker benchmark. So we built the case news actually throws at them: a split screen with two people, both moving their lips, and one voice. The detectors pick the right person 55–60% of the time. A coin gets 50%.
Read the full study on Every Local AI