Orchestral Pit

Keeping up with MIR is more than having a daily feed of new papers. Labs, communities, events and newsletters matter just as much, but you mostly find them through word of mouth. We’ve compiled an MIR resource guide on GitHub: github.com/The-Orchestral… Contributions are welcome!

📊 This week: representation learning is consolidating around richer input structures; generalist audio generation is accelerating. Ethics and computational musicology were quiet, and no industry preprints surfaced — watch those gaps.

🔍 Spotlight — Cognitive & User-Centered MIR: Aligning MusicLLM to emotion via feedback-driven tuning — affective responsiveness without forgetting. Underrated direction. arxiv.org/abs/2606.24123…

📰 Week of Jun 22 — 14 MIR papers. Here's what moved the field this week. #MIR #MusicInformationRetrieval 🧵

Multimodal framework matches expert judges on singing quality by combining lyric and pitch-rhythm analysis. Handles melisma, vibrato, tempo elasticity. #SingingVoice #MIR arxiv.org/abs/2606.26451…

Dynamic gated attention boosts cross-domain audio representations to 0.836 score. Combine pre-trained models via selective feature routing for better generalization. #AudioRepresentation #CrossDomain arxiv.org/abs/2606.26556…

Unified audio-visual encoder outperforms specialized models with cross-modal prediction. Drop-in frozen model for audio tasks, competitive on video with 10x less data. #AudioVisual #SelfSupervised #MIR arxiv.org/abs/2606.25225…

2D spectrogram SSL outperforms 1D sequence models across MIR tasks. PupuJEPA applies visual JEPA to spectrograms with domain-specific tweaks. Code + checkpoints available. #MusicIR #SelfSupervisedLearning arxiv.org/abs/2606.25713…

Feedback-driven alignment boosts music emotion prediction while preserving MusicLLM capabilities. Emotion prediction now viable in music LLMs. #MusicLLM #EmotionRecognition #MIR arxiv.org/abs/2606.24123…

Headrest speakers boost immersive audio perception in cars without distracting passengers. Standard feature ready for OEM integration. #SpatialAudio #AutomotiveAudio arxiv.org/abs/2606.24146…

Non-invasive acoustic sampling preserves and reactivates museum woodwind instruments digitally. Study extinct techniques and enable historically informed performance without damaging originals. #MusicalAcoustics #CulturalHeritage #MIR arxiv.org/abs/2606.24216…

Multi-stage separation cascades acoustic and semantic clues for spatial sound scene segmentation. Iterative self-guided refinement yields 7dB CAPI-SDRi improvement over baseline. #SoundSceneAnalysis #SourceSeparation arxiv.org/abs/2606.24512…

Two-stage cascade outperforms single-stage music source restoration on all stems. Decouples distribution fitting from signal reconstruction. Reveals signal-semantic trade-offs. #MusicRestoration #SourceSeparation arxiv.org/abs/2606.24127…

Real-time music generation enables live human-AI co-creation. Data-free streaming distillation lets musicians steer generation on the fly. Generative models become playable instruments. #MusicGeneration #InteractiveAI #MIR arxiv.org/abs/2606.24307…

Pluck depth shifts guitar timbre, loudness, and harmonics in measurable ways. Robotic testing across six pick materials reveals design optimization paths. #AcousticGuitar #Timbre #MIR arxiv.org/abs/2606.24356…

Keeping up with MIR is more than having a daily feed of new papers. Labs, communities, events and newsletters matter just as much, but you mostly find them through word of mouth. We’ve compiled an MIR resource guide on GitHub: github.com/The-Orchestral… Contributions are welcome!

LLM-generated playlist captions boost user engagement on Deezer's Daily Mix. Semantic framing shapes perception of unchanged recommendations. #MusicRecommendation #MIR arxiv.org/abs/2606.22460…