Marimba: Open-Source Speech Decision Model Built on Frozen Qwen3-ASR-0.6B
solanotodesch · reddit · 2026-11-16
Marimba is a System One model for audio—like Jev, but for speech—designed to be fast and lightweight while returning typed decisions: yes/no, a choice, or a score.
It runs on a frozen Qwen3-ASR-0.6B. Word and language questions use zero-shot scores from the ASR decoder's hypotheses, followed by calibration. Small readouts on the frozen encoder handle voice activity detection, turn endings, speaker changes, and speaker count.
Fully open under Apache-2.0, installable via pip install marimba-audio, with a 60-second demo video, live Hugging Face Space, a benchmark dataset, and an arXiv paper coming this week.
More from Multimodal
- Dad builds a Pokémon-style math game for kids in one weekend with Fable and Scenario — AIandDesign · 2026-11-16
- Text-to-image models keep drawing an odd bulge in women's underwear, user seeks fix — jiroq · 2026-11-16
- 3D scene built with three.js + Fable 5.2 + Blender, full source released — gizmoo007 · 2026-11-16
- AI-generated GTA-style 'GrampTheftAuto' intro with self-written concept lyrics — No_Seaworthiness_979 · 2026-11-16
- Dead Star Trek actors sing again: AI-powered parody video stuns complexity blogger — fortnow · 2026-11-16
- Midjourney sref nails crayon-textured LOTR characters: Frodo, Legolas, Aragorn — sergeantsref · 2026-11-16