SEA-SpeechBench - Speaker Recognition (SEA Prompt): leaderboard
Speaker verification in SEA-SpeechBench (Liao et al., 2026): deciding whether two clips come from the same speaker, on constructed pairs in 11 Southeast Asian languages, with the instruction written in the native Southeast Asian language. Macro-F1 in percent, with refusals counted as errors; higher is better.
Metric: Macro-F1 (%; whether two clips come from the same speaker, on self-constructed same and different speaker pairs, refusals count as errors; short clips of at most 30 s from curated public Southeast Asian speech corpora, up to 1,000 sampled items per dataset; instruction prompt in the native Southeast Asian language). Source: arxiv.org. Saturation forecast: Around October 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 51.54 | #452 |
| 2 | Qwen3 Omni 30B A3B Instruct | 44.28 | #902 |
| 3 | Voxtral-Mini-3B-2507 | 38.67 | #3702 |
| 4 | MERaLiON-2-10B | 37.07 | #1646 |
| 5 | Gemma 3n E4B (IT) | 34.13 | #2220 |
| 6 | Phi-4 Multimodal Instruct | 32.73 | #3914 |
| 7 | Gemma 3n E2B (IT) | 32.07 | #3758 |
| 8 | MERaLiON-2-3B | 29.91 | #6258 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=sea-speechbench-speaker-recognition-sea-prompt · How It Works · Data refreshed daily, snapshot 2026-11-12.