Marimba: Speech System Using Frozen Qwen3-ASR Encoder Matches or Beats Qwen3-Omni on 6 of 9 Tests

·2026.11.17 00:47

Key point

Marimba, a system using a frozen Qwen3-ASR-0.6B encoder, matches or beats the 30B Qwen3-Omni on six of nine public tests.

Details

  • Marimba is a system that answers typed questions about audio windows (yes/no, choice, score) using a single frozen Qwen3-ASR-0.6B encoder pass per 8-second window.
  • On 1,000 public questions, Marimba matches or beats Qwen3-Omni (30B-A3B) on 6 of 9 tests, though Qwen3-Omni remains superior on Speech Commands.
  • Language identification achieves 100% accuracy on 300 FLEURS clips, and speaker change detection yields an AUROC of 0.966-0.980 on AMI and VoxConverse, exceeding ECAPA performance.
  • Speaker count detection outperforms pyannote 3.1, while turn-end detection achieves an AUROC of 0.930 on Smart Turn data and 0.716 on ICSI meetings.
  • The system does not generalize to unseen tasks, performing at chance level on CochlScene scene classification.
  • Code, weights (Apache-2.0), and a demo are available on GitHub, Hugging Face, and via pip install marimba-audio.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.