Marimba: a frozen 0.6B ASR backbone beats 30B Qwen3-Omni on 6 of 9 audio tests
solanotodesch · reddit · 2026-11-16
Author released the Marimba paper (arXiv:2611.18330): a "System One" speech model built on a frozen Qwen3-ASR-0.6B encoder, answering many typed questions about an 8s audio window from a single forward pass with calibrated probabilities.
- Method: term spotting scores greedy + 4-beam decoder hypotheses with Platt calibration (zero-shot); language ID reuses ASR's own scores; turn-end, speaker change and count use small readouts on frozen layers 18/12.
- Results: matches or beats Qwen3-Omni (30B-A3B) on 6 of 9 tests over the same 1,000 public questions; 100% on 300 FLEURS language clips; speaker-change AUROC 0.966–0.980 beats ECAPA; speaker count beats pyannote 3.1.
- Weaknesses: turn-end AUROC 0.930 vs Smart Turn v3.2's 0.982 on its own benchmark; heads don't generalize to unseen tasks (CochlScene at chance).
- Open source: code, Apache-2.0 weights, HF demo, pip install marimba-audio.
Related event: Marimba: Open-Source 0.6B Speech Decision Model Beats 30B Rivals(2 posts)→
More from Multimodal
- A $250 robot hands you tools, fly-around World Series home runs: 10 wild AI visual demos this week — minchoi · 2026-11-17
- Meituan's LongCat launches Avatar 1.6: 10+ min consistent avatar video, MIT licensed — AdinaYakup · 2026-11-17
- Guide: use Claude Code to write Modular Diffusers integrations for unsupported models — linoy_tsaban · 2026-11-16
- Runway partners with INK Studio on AI-animated short films, debut 'Clumsy Cowboy' — tlakomy · 2026-11-16
- MOSS Image ComfyUI workflow: layered Base/SFT blending with 8-step acceleration — wjc_5 · 2026-11-16
- Testing MiniMax H3 image model in ComfyUI at a massive 16K resolution — tisch_eins · 2026-11-16