Pei Yin

We released a new VibeVoice model, VibeVoice ASR, a unified speech-to-text model designed to handle 60-minute long-form audio in a single pass, generating structured transcriptions containing Who (Speaker), When (Timestamps), and What (Content), with support for Customized Hotwords. It redefines ASR in the era of LLMs, and the new ASR will be a foundation to unlock the potential of new AI-Native hardware, memory systems for AI, and many other scenarios. Please have a try at https://lnkd.in/dNCCzQiN Key features: -         60-minute Single-Pass Processing: Unlike conventional ASR models that slice audio into short chunks (often losing global context), VibeVoice ASR accepts up to 60 minutes of continuous audio input within 64K token length. This ensures consistent speaker tracking and semantic coherence across the entire hour. -         Customized Hotwords: Users can provide customized hotwords (e.g., specific names, technical terms, or background info) to guide the recognition process, significantly improving accuracy on domain-specific content. -         Rich Transcription (Who, When, What): The model jointly performs ASR, diarization, and timestamping, producing a structured output that indicates who said what and when. We are continuously pushing the frontier of voice AI. More release on the way. Stay tuned! https://lnkd.in/dATPu9nw

380

8 Comments

⏩️⏩️ AI is evolving very fast. How can we keep up with the trends? Yann LeCun, one of the AI godfathers, has a simple advice for this: 👉 Study fundamentals! I personally find this advice to be solid. I can keep up with the fastly-changing AI trends thanks to the firm understanding of fundamental math like calculus and linear algebra. Let's name and dissect some of the fancy AI terms: ➡️ LLM: it is fundamentally built with calculus and linear algebra ➡️ Vector: it is a concept in linear algebra ➡️ Embedding: it is originally a concept in math Fundamentals may be boring but it makes us easier to understand advanced AI concepts! #AiTrends #AiFundamentals #Calculus #LinearAlgebra

7

Drop 5/14: Introducing Bulbul V3: our latest text-to-speech model, built to redefine how human and how reliable synthetic speech can be. Bulbul V3 sets a new benchmark for naturalness and robustness in real-world voice agent deployments, not just studio demos but production environments where consistency matters. In a blind human listening study conducted by Josh Talks, listeners compared Bulbul V3 with ElevenLabs (v3 alpha and v2.5 flash) and Cartesia Sonic-3. With 20,000+ votes across languages, Bulbul V3 emerged as the highest listener-preferred model for 8kHz audio the backbone of voice agents while also recording the lowest average error rates from tagged failure modes. We further stress-tested it on numerics, technical content, named entities and complex linguistic inputs. Across languages, Bulbul V3 consistently delivered the lowest error rates built to handle messy real-world inputs at scale, from customer support to outbound and multilingual deployments. All developers and enterprises can access Bulbul V3 at no cost for the rest of February. We encourage everyone to build, experiment, and deploy at scale. Play a little game and dive into the details in our blog: https://lnkd.in/gVPMNxFk Try it out here: https://lnkd.in/gGnn6ffV

1,326

40 Comments

Everyone’s talking about preference models in AI—and for good reason. In our latest AAAI Senior Member track paper, Abhinaba Roy, Ph.D. and I explore approaches to align generative music systems with what humans truly want. Generative music AI has achieved incredible fidelity and stylistic diversity, but without preference alignment, these systems risk missing the essence of human musical appreciation. We dive into: - Large-scale preference learning (MusicRL) - Multi-preference frameworks (DiffRhythm+) - Inference-time optimization (Text2midi-InferAlign) Our paper highlights key challenges—temporal coherence, harmonic consistency, and subjective quality—and envisions a future where preference-aligned music generation powers personalized music services and interactive composition tools. 🔗 Read the paper: https://lnkd.in/gK7Er-iA How do you typically integrate user preference alignment in your models? PS: if you are planning to attend this year's AAAI, be sure to stop by the Emerging Technologies for AI Workshop! https://lnkd.in/gY2NnUda Made in the AMAAI Lab at Singapore University of Technology and Design (SUTD) Association for the Advancement of Artificial Intelligence (AAAI) #AAAI #MusicAI #GenerativeAI #PreferenceAlignment #AIResearch #MusicTechnology #ISMIR

71

Thrilled to spotlight the release of Neuroevolution: Harnessing Creativity in AI Agent Design, a new MIT Press book written by long-time contributors to the field – Sebastian Risi, Yujin Tang, David Ha, and Risto Miikkulainen. This book is the first of its kind to distill down decades of research and show how evolutionary computation is becoming a foundation for building adaptive, creative AI systems. The website also features demos, tutorials, exercises, and teaching resources, making it a great starting point for hands-on exploration. Read the open-access edition: https://lnkd.in/gHcW2dTi Read the full blog: https://lnkd.in/gepfPzYB #Neuroevolution #AIAgent #AI

24

Scripps accelerating agentic AI at scale. Hot-swapable LLMs connected to sophisticated embedded prompts, secure proprietary data, tools, MCP connections and many-many time saving agents now available and many more in the bullpen. Incredible team effort and massive shared learning!

19

7 Comments

FSP EP. 12 “Algorithmic Alchemy” is now live on Focus Stacked Podcast. Host Jim Reiman sits down with AI conceptual artist Todd Dobbs to unravel how generative tools can expose hidden biases, rewrite photo history, and turn fake houseplants into haunting cyanotypes. From prompting algorithms to define “Typical Americans” to debating who owns an image when code shares the credit, this conversation is equal parts mind-bending and eye-opening. https://lnkd.in/gGEanSW8

5