How Much Physics Can Vision Learn from Audio? reports audio embeddings achieve a Spearman rank correlation of 0.73 for quality factor ordering, while a single video frame encoded by CLIP reaches only 0.28. The paper also shows distilling audio into a four-frame visual model raises correlation from 0.234 to 0.338 on a leave-one-material-out test.
Key points
Audio embeddings reach ρ = 0.73 for ordering impacts by quality factor Q.
Distillation with four-frame input improves correlation from 0.234 to 0.338 on held-out materials.
Sources
Frame-Wise Streaming Speech Compression, BD-Rate cuts
Frame-Wise Streaming for Efficient Learned Speech Compression introduces the StreamLSC model with BD-Rate reductions of –51.37% (ViSQOL) and –53.99% (PESQ). The model uses Boundary-Cached Causal Convolution, Intra-Frame Layer Normalization, and RNN-mode RWKV, and adds a streamable entropy quantizer. Experiments report state-of-the-art rate-distortion results and subjective gains of 0.2154 BD-ViSQOL and 0.5156 BD-PESQ.
Key points
StreamLSC combines Boundary-Cached Causal Convolution, Intra-Frame Layer Normalization, and RNN-mode RWKV.
Achieves –51.37% ViSQOL and –53.99% PESQ BD-Rate savings on average.
Sources
CF-AVSS ties first on Track 1, second on Track 2 in AVSE Challenge
CF-AVSS is a coarse-to-fine two-stage audio-visual speech separation framework that achieved tied first place on Track 1 and second place on Track 2 of the ISCSLP 2026 Real-World AVSE Challenge. It recorded the lowest character error rates of 12.31% and 14.79% on the two tracks.
Key points
Tied for first place on Track 1 of the ISCSLP 2026 Real-World AVSE Challenge
Achieved CER of 12.31% on Track 1 and 14.79% on Track 2, the lowest on both tracks
Sources
From Signals to Evidence: Framework and Recognition released
From Signals to Evidence: Interpretable Underwater Acoustic Target Sensing with Large Language Models - Part I presents a new framework that couples physics-aware acoustic encoding with a pretrained LLM. The paper introduces the Underwater Acoustic Physics-aware Transformer (UPT) and a physics-parameterized data engine, and reports accurate vessel recognition and reliable estimation of shaft frequency, blade-passing frequency, and machinery-related attributes on simulated recordings of four vessels.
Key points
Physics-aware acoustic encoding integrated with a pretrained LLM in a unified feature-evidence-decision chain.
Underwater Acoustic Physics-aware Transformer (UPT) tokenizes complementary acoustic representations and learns task-disentangled physical representations via four physics-oriented learning tasks.
Physics-parameterized data engine provides traceable supervision at two levels of granularity.
Experiments on simulated recordings of four vessels demonstrate accurate vessel recognition and reliable estimation of shaft frequency, blade-passing frequency, and machinery-related attributes.
Sources
Songhunt AI music player launches for 1,500+ PC games
Songhunt is an AI-powered music player that automatically adapts the soundtrack to gameplay. It supports more than 1,500 PC games on Overwolf.
Key points
Feature: automatically adapts soundtrack to gameplay moments
Availability: works with 1,500+ PC games on Overwolf
Sources
Freebeat AI lyric video competition offers $8,000 prize
Freebeat AI lets you turn music into a lyric video and compete for a share of an $8,000 prize pool. Your creativity decides the visual style.
Key points
Turn music into a lyric video.
Compete for part of an $8,000 prize pool.
Sources
Soundverse AI allows Suno songs import and metadata preservation
Soundverse AI lets creators import Suno songs, keep artwork and metadata, and convert tracks into new versions. One library supports more creative possibilities.
Key points
Import: Suno songs can be brought to Soundverse.
Preservation: Artwork and metadata are retained.
Sources
Bias Mitigation in Audio-Enabled LLMs – gender unlearning results
Bias Mitigation in Audio-Enabled LLMs through Adversarial Training and Distillation Methods reports gender-unlearning using adversarial methods without loss on other tasks. The authors also distil a gender-agnostic teacher to create a model that perceives gender but follows the teacher unless explicitly required, with satisfactory gender-unlearning and partially satisfactory distilling results on three AuLLMs.
Key points
Gender perception can be unlearned using adversarial methods without a loss in performance on other benchmarked tasks.
Applied the proposed methods to three AuLLMs, obtaining satisfactory gender-unlearning and partially satisfactory distilling outcomes.