Ad
Ad
Ad
Short

NVIDIA NeMo, an open-source conversational AI toolkit, has released Parakeet, a set of automatic speech recognition (ASR) models. Developed in partnership with Suno.ai, the four Parakeet models, ranging from 0.6 to 1.1 billion parameters, can transcribe spoken English and are available for commercial use under the CC BY 4.0 license. The models were trained on 64,000 hours of audio data covering different accents, ranges, and sound conditions. According to the developers, the models are robust to non-speech segments such as music and silence, and outperform OpenAI's Whisper v3 in benchmarks. They also offer user-friendly integration into projects through pre-trained control points. A demo of the 1.1 billion parameter model is available here.

Ad
Ad
Short

Microsoft 365 business customers get access to Copilot, but with a catch: they must purchase a minimum of 300 individual licenses at $30 per user per month. This means that AI for Word, Excel, and PowerPoint currently costs at least $108,000 per year. The offer appears to be targeted at large enterprises, as only about 77,000 of the more than 18 million U.S. businesses have more than 250 employees (via naics.com). Microsoft has indicated that the 300-user minimum may be a temporary restriction in the beta phase, Heise Online reports. Educational institutions gained access to Copilot earlier this year, but it is unclear whether the same minimums apply. Microsoft also promises educational institutions "commercial data protection" (data will not be stored or used for training purposes) for users over the age of 18. What this means for underage users is not clear from the announcement.

Short

The National Institute of Standards and Technology (NIST) has released a comprehensive report on adversarial machine learning (AML) that provides a taxonomy of concepts, terminology, and mitigation methods for AI security. The report, authored by experts from NIST, Northeastern University, and Robust Intelligence, reviewed the AML literature and organized the major types of ML methods, attacker goals, and capabilities into a conceptual hierarchy. It also provides methods for mitigating and managing the consequences of attacks, and highlights open challenges in the lifecycle of AI systems. With a glossary for non-expert readers, the report aims to establish a common language for future AI security standards and best practices. The full 106-page report is very detailed and references real-world attacks such as Prompt Injections. It's available for free.

Ad
Ad
Google News