Ambiguous Speech Emotion - MSP-Podcast: leaderboard

Metric: Jensen-Shannon divergence (0-1; between the emotion distribution the model predicts for an utterance and the distribution of the human annotators' labels, mean over utterances; lower is better; MSP-Podcast, 12,955 naturalistic podcast segments over eight emotions). Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked.

Top models

#ModelScore
1Gemini 2.5 Pro0.37
2GPT-4o Audio0.37
3Gemini 2.0 Flash0.4
4Qwen2-Audio-7B-Instruct0.41
5Qwen2.5-Omni-7B0.5

Interactive version: theaggregate.ai/benchmark?slug=ambiguous-speech-emotion-msp-podcast · How It Works · Data refreshed daily, snapshot 2026-09-26.