Speech-to-text

Meta & AI

Meta has recently accelerated its AI development, unveiling a new AI model on a weekly basis, with the latest being their groundbreaking Massively Multilingual Speech (MMS) models. These models, remarkable for converting speech to text and vice versa, support an unprecedented 1,100 languages.

This feat was achieved by utilizing the New Testament for training data, ensuring robust performance across both religious and secular domains. Impressively, these models are openly accessible on GitHub, marking a significant contribution to linguistic diversity in AI technology.

DragGAN

This past Friday, a group of researchers published a paper showing how they built DragGAN, an AI model that allows them to edit photos by simply dragging a point with the mouse.

Right now DragGAN is a demo and the is scheduled for release in June. We really can’t wait to see what the open-source community will build once they get their hands on it.

Resources

AI Insider 📰

1. We saw it, we liked it and now we’re sharing it with you. It’s a Coca-Cola ad that was released in early March and the production combines VFX & AI.
2. Microsoft has added new features to Bing, including chat history, data visualization, and the ability to export answers to PDF or Word.
3. Intel announced the  Aurora AI model with 1 trillion parameters, focused on scientific fields like Systems Biology and Climate Science.
4. Apple is in a hiring spree for machine learning specialists.
5. OpenAI released its ChatGPT app for iOS devices and had to fight with thousands of its clones to reach the top results. The app is only available in the US for now. Slowly it will be rolled out to the rest of the world and we are waiting for the Android app.
   

Learning Bytes 🧐

1. The Bulimia Project, wanting to explore what exactly are the “ideal bodies” being promoted on social media, gave the following prompts to Dall-E 2, Stable Diffusion and Midjourney:

   • “The ‘perfect’ female body according to social media in 2023”
   • “The ‘perfect’ male body according to social media in 2023”

   Their findings revealed that a significant portion of AI-generated images depicted unrealistic bodies, highlighting prevalent biases towards certain physical features such as blonde hair and dark skin for women, and brown hair and chartreuse skin for men, along with frequent inclusion of beards for men.

2. Research has shown that AI detectors often “tag” as AI-generated content from non-native English speakers.

Cool Finds 🤯

Human or Not? , it’s a social Turing Game, where you talk to someone for 2 minutes and try to figure out if they are human or not.

See you next week! 💚

#TogetherWeAI