GPT-4o hit 70% on spectrogram classification two years ago, near human experts' 72.5%

TuhinChakr · x · 2026-09-10

Chris Donahue points out that VLM-based audio spectrogram classification isn't new: in the paper "Vision Language Models Are Few-Shot Audio Spectrogram Classifiers," GPT-4o scored 70% on 10-way spectrogram classification versus human experts at 72.5% — something possible for at least a couple of years. He jokes he's entering the "we tried that years ago" phase of his research career.

Context: Greg Brockman was showing Astra doing spectrogram-to-sound identification.

Original post →

More from Models

Models channel →