Open-weights SUPlime beats pyannote's commercial diarization with 15.86 DER
solyarisoftware · x · 2026-09-12
Re:WayAI has open-sourced SUPlime, a speaker diarization model packaged for pyannote.audio 4.x. It achieves 15.86 macro DER on pyannote's 8-corpus benchmark, beating pyannoteAI's commercial precision-2 (16.06), and 20.94 across all 12 corpora.
- Two variants: SUPlime on a WavLM-Base+ backbone (114M params) and SUPlime-L on WavLM-Large (349M)
- SUPlime-L is stronger on meeting/conversational audio; the base model wins on far-field recordings (CHiME-6, DiPCo, NOTSOFAR-1)
- Architecture: Conformer segmentation + SimAM-ResNet34 embeddings; weights and training code are open
- Alpha CEPHEI notes strong results on non-English telephony, close to Diarizen Large V2
More from Models
- antirez uploads DeepSeek V4.1 Flash GGUF quantizations to Hugging Face — Queasy_Asparagus69 · 2026-09-12
- Qwen3.8-Max scores 19 on OpenRouter vs 26 on Alibaba — effort level bug suspected — PawelHuryn · 2026-09-12
- Pipecat v1.9 adds Meta's Muse Voice Transcribe, the lowest semantic-WER STT model tested — solyarisoftware · 2026-09-12
- Rumor: Claude Opus 5 reportedly available to Google engineers in their IDE — unironictechbro · 2026-09-12
- MiniCPM5-2B lands on Hugging Face as author claims Llama architecture beats Qwen3.5 — solyarisoftware · 2026-09-12
- DeepSeek V4.1 model card shows same model scores wildly differently across agent harnesses — solyarisoftware · 2026-09-12