aiqbench analysis
EchoDistill introduces a post-training self-distillation technique that enhances Large Audio Language Models' resilience against acoustic noise without incurring additional inference costs. By leveraging clean audio as privileged information during training, the method aligns noisy-input generation with clean-conditioned semantics, yielding measurable accuracy gains across multiple backbones and domains. This advancement addresses a critical vulnerability in multimodal AI systems where environmental interference previously compromised task-relevant evidence.
EchoDistill: Robust Large Audio Language Models via Noisy-to-Clean Self-Distillation
arXiv:2605.23954v2 Announce Type: replace Abstract: Large Audio Language Models (LALMs) remain vulnerable to acoustic noise, which can obscure task-relevant evidence and produce unreliable responses. We propose EchoDistill, a noisy-to-clean self-distillation framework that uses clean audio as privileged information during post-training. A noisy-input student samples candidate responses reflecting its inference-time behavior, while a frozen copy of the same backbone processes the corresponding clean audio. EchoDistill combines masked response-token distillation, task-gated consistency shaping, and teacher-referenced group-relative optimization to align noisy-input generation with clean-conditioned semantics. Only the student is retained at inference time, introducing no additional inference cost. Across three LALM backbones and three audio domains at -10dB, EchoDistill improves average noisy-input accuracy by 1.63 percentage points over the strongest baseline. On Qwen2.5-Omni, it raises noisy-input accuracy from 59.33% to 62.94%, while clean-audio accuracy increases from 76.56% to 77.56%. Replacing matched audio with random, shuffled, or silent inputs reduces accuracy by 3.08-6.42 points, confirming that matched acoustic evidence contributes to its predictions. Additional evaluations show improvements on held-out additive noises and external benchmarks, while revealing that these gains do not reliably extend to non-additive distortions. These results demonstrate robust post-training improvements under severe additive noise without sacrificing clean-audio capability across diverse tasks.
Abstract:Large Audio Language Models (LALMs) remain vulnerable to acoustic noise, which can obscure task-relevant evidence and produce unreliable responses. We propose EchoDistill, a noisy-to-clean self-distillation framework that uses clean audio as privileged information during post-training. A noisy-input student samples candidate responses reflecting its inference-time behavior, while a frozen copy of the same backbone processes the corresponding clean audio. EchoDistill combines masked response-token distillation, task-gated consistency shaping, and teacher-referenced group-relative optimization to align noisy-input generation with clean-conditioned semantics. Only the student is retained at inference time, introducing no additional inference cost. Across three LALM backbones and three audio domains at -10dB, EchoDistill improves average noisy-input accuracy by 1.63 percentage points over the strongest baseline. On Qwen2.5-Omni, it raises noisy-input accuracy from 59.33% to 62.94%, while clean-audio accuracy increases from 76.56% to 77.56%. Replacing matched audio with random, shuffled, or silent inputs reduces accuracy by 3.08-6.42 points, confirming that matched acoustic evidence contributes to its predictions. Additional evaluations show improvements on held-out additive noises and external benchmarks, while revealing that these gains do not reliably extend to non-additive distortions. These results demonstrate robust post-training improvements under severe additive noise without sacrificing clean-audio capability across diverse tasks.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Sound (cs.SD)
arXiv:2605.23954 [cs.CL]
(or arXiv:2605.23954v2 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2605.23954 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 292 words
Related coverage
Links #3: 2026/06 Part 1
LessWrong · Jun 15, 2026
A kernel-centric path to real-time video generation on Trainium
Amazon Science · Sep 25, 2026
Best Open Speech Recognition (ASR) Models in 2026: WER, Languages, Latency, and License Compared
MarkTechPost · Jul 23, 2026
Leveraging Audio-LLMs to Filter Speech-to-Speech Training Data
arXiv · Jun 12, 2026