Cortex makes your CLI agent write and read persistent, structured memory about any substantial work performed or facts gathered within a project. Information is preserved over weeks and months, essentially providing an unlimited context window. Cortex also makes it easy to specify new agents, assign them specific roles and permissions, and schedule how often they wake up. It comes with predefined agent roles, such as the Watch agent, which observes your deep learning runs and sends you a Telegram message if anything goes wrong. To learn more about the Cortex framework, watch the talk I recently gave in Sony CSL’s seminar series and/or visit the repository: youtube.com/watch?v=QL0LDzBR… github.com/SonyCSLParis/cort… #AgenticFramework #AIAgents #CodingAgents #ResearchEngineering #DeveloperTools @SonyCSLMusic

We just made Cortex public! 🚀 Cortex is a lightweight, open-source framework that turns AI coding agents into a persistent, coordinated workforce with continuity across chat sessions. It practically yields an unlimited context window, maintaining structured logs, cheat sheets and shortcuts. It’s built for researchers and engineers whose work spans multiple projects, machines, and long-running experiments—and who want agent activity to remain understandable and auditable. It gives agents durable, structured, human-readable project memory, shared task tracking, remote-server coordination, experiment bookkeeping, monitoring, specialized roles, and carefully scoped permissions. One conductor remains your main interface, while node agents and specialized workers can inspect systems, run experiments, maintain projects, back up configured work, safely package approved changes into commits, and preserve useful context across sessions. Cortex is practical: plain files, Git, SSH, and the provider CLIs you already use. We’re excited to share it publicly and curious to see what you build with it! github.com/SonyCSLParis/cort… #AgenticFramework #AIAgents #CodingAgents #ResearchEngineering #DeveloperTools, @SonyCSLParis @SonyCSLMusic @SonyCSL

💫 New paper at C&C '26 (ACM Creativity & Cognition): Adversarial Aesthetics: Using Spectrogram Steganography to Interrogate AI Ethics in Electronic Music. 🕵️ Hide information inside the spectrograms of electronic music and use that act of concealment as a way to probe the ethical questions AI raises for the field. Steganography becomes a critical tool rather than just a technical trick, turning the spectrogram into a site where questions about data, authorship, and consent in AI music can be made audible and visible. Great work by Alexander James Williams & Yutian Hu! 📄 Paper: doi.org/10.1145/3803784.3809… @SonyCSLMusic @SonyCSL #MusicAI #CreativityAndCognition #AIEthics #ElectronicMusic #MIR

Excited to share LiveBand, the first tool that enables robust, real-time jamming with an AI audio generator! 💫 🎸 LiveBand runs locally on a Mac, responds instantly to a musician’s input, and can be guided using either audio references or text prompts. The result is a highly versatile musical partner that you can actually jam with. 🚀 It opens up new possibilities for live performance, creative experimentation, and practicing your improvisation skills. We believe that this is a real game changer for human–AI musical interaction! 📜 Paper: arxiv.org/abs/2606.03803v2 🎶 Audio Samples: sonycslparis.github.io/liveb… 📽️ Video: 👇 Made possible by the great Sony CSL Music Team: 👉 @marco_ppasini 👈 (!), @latentspaces , Ben Hayes, @cyranaouameur & @BjareMathias

🔊 New Paper Alert — Training-Free Inference-Time Timbre Transfer Check out our latest #ICASSP paper on timbre transfer in music audio! 🎶 ➡️ Diffusion Timbre Transfer Via Mutual Information Guided Inpainting; Ching Ho Lee, Javier Nistal, Stefan Lattner, Marco Pasini & George Fazekas 📜 Paper: arxiv.org/pdf/2601.01294 🥁 Audio Demos: anon-audio-demo-25.github.io… In this work, we rethink timbre transfer as an inference-time editing problem — and show that you don’t need to retrain or fine-tune heavy models to change the instrumental color of a piece while preserving its musical structure. 🎯 What’s New? Instead of training separate models or adding control modules for each instrument: ✅ We start from a pre-trained latent diffusion model and steer it on the fly using two simple controls: • Mutual-Information guided noise injection: add noise only in latent channels most informative of timbre. • Early-step clamping: “lock in” melody and rhythm by restoring structure-dominant channels during denoising. This lightweight, training-free procedure lets you control timbre without sacrificing the original melody, harmony or rhythm — and works with text or audio conditioning (e.g., CLAP). ✨ Why It Matters 🎵 Practical music production tools for re-orchestration and sound design 🛠️ Efficient editing with no added model training 🔍 A framework that could extend beyond timbre to other label-driven audio edits 📌 Compatible with strong diffusion backbones and generative audio models @SonyCSLMusic @SonyCSLParis

🎶 Internship Opportunity — Sony CSL Paris Music Team Focus: Artist-centric music AI Location: Paris, France Duration: 3-6 months Start date: January-June 2026 (flexible) About the Team The Music Team at Sony Computer Science Laboratories (CSL) Paris is working on the future of AI-assisted music creation. We explore how AI can empower creativity rather than replace it. Our research spans human-AI collaboration, interactive audio generation, live performance tools, and music cognition, all centered on artists’ real-world workflows and creative needs. Internship Mission We’re seeking a creative and technically skilled research intern to contribute to the development of artist-centric AI tools. You will work with our research team to design, implement, and test cutting-edge machine learning algorithms, which will form the basis of tools that enhance artistic control and exploration. Candidate Profile Essential: - Strong Python skills - Experience with a machine learning framework (PyTorch strongly preferred) - Comfortable using the command line in Linux environments - Background in at least one of: audio signal processing / acoustics / music theory / music cognition - Experience training modern neural networks Strongly preferred: - Familiar with the design of modern transformers - Familiar with major classes of modern generative models (autoregressive, diffusion/flow-based) - Currently pursuing a Master's or PhD in computer science, AI, EE, music technology, or equivalent Bonus: A strong interest in music, whether you play it, produce it, or listen to a lot of it. Publications in relevant conferences or journals (ISMIR, ICASSP, AES, ICLR, NeurIPS, ICML). 👉 How to Apply Send your CV and a short motivation statement to [email protected].

🎶 New ISMIR 2025 paper! "Estimating Musical Surprisal from Audio in Autoregressive Diffusion Model Noise Spaces" by Mathias Rose Bjare (JKU), Stefan Lattner (Sony CSL), and Gerhard Widmer (JKU/LIT AI Lab). We explore how surprisal — the unexpectedness of musical events — can be modeled directly from audio using autoregressive diffusion models (ADMs). 💡 What we did: - Compared surprisal from diffusion models vs. Generative Infinite-Vocabulary Transformers (GIVT). - Evaluated across tasks: monophonic pitch surprisal (expectation) & segment boundary detection (structural surprise). - Tested surprisal at different noise levels in diffusion processes to see how musical features emerge at multiple granularities. 🔥 Key take-aways: - Diffusion surprisal beats GIVT in modeling pitch expectation & boundary detection. - Mid-level noise surprisal captures pitch-level expectations while suppressing timbre-related variation. - Surprisal curves align with human-like musical segmentation, showing potential as proxies for perceptual surprise. Why it matters: Understanding musical surprisal links computational modeling with human perception and cognition — with applications in AI composition, real-time music interaction, and brain-music studies. 📜 Paper: arxiv.org/abs/2508.05306 💻 Code: github.com/SonyCSLParis/audi… #DiffusionModels #Surprisal #MIR @ISMIRConf @SonyCSLParis @SonyCSLMusic

🎉 New paper announcement: "Assessing the Alignment of Audio Representations with Timbre Similarity Ratings" (accepted at ISMIR Conference 2025) 🎶 We asked a simple question: How close are today’s audio embeddings to the way humans perceive timbre? - We evaluated 18 hand-crafted & learned representations (MFCC, Music2Latent, Encodec, JTFS, CLAP, etc.) against 2 614 human similarity ratings spanning 21 classic “timbre-space” datasets. - We also tested a new sound-matching model and introduced an open-source Python package for easy benchmarking. 👉 Key take-aways: - Style embeddings (Gatys/Huang) extracted from CLAP and our sound-matching model *achieve the highest agreement* with human judgments, topping all other metrics. - Music2latent representations beat most of the others in (non-rank-based) relative distances - MFCCs are still competitive, sometimes beating more complex models. - Results pave the way for perceptually grounded metrics in generative audio, sample retrieval & instrument modeling. Big thanks to @tiianhk (QMUL) & Charalampos Saitis (QMUL) 📜 Read the paper: arxiv.org/abs/2507.07764 (Repo link inside the paper – contributions welcome!) #Timbre #AudioAI #MIR #ISMIR2025 #Research @SonyCSLMusic @SonyCSL

As @TechHubbySony went offline, here is the direct link to DrumGAN: drumgan.csl.sony.fr/