Afrispeech Semantics - Consistency (AfriSpeech-200): leaderboard

Metric: Macro-F1 (%; binary consistent versus inconsistent judgments of text hypotheses against a spoken premise, computed over all premise-hypothesis pairs of the AfriSpeech-200 set; zero-shot generative audio language model, free-form answers mapped to labels by a lightweight language model). Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked.

Top models

#ModelScore
1Qwen2-Audio-7B-Instruct84.46
2SALMONN-13B73.75
3Qwen2.5-Omni-7B68.8

Interactive version: theaggregate.ai/benchmark?slug=afrispeech-semantics-consistency-afrispeech-200 · How It Works · Data refreshed daily, snapshot 2026-09-26.