Stability AI announces Stable Audio

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each weekday, we scour the web to aggregate the many stories related to artificial intelligence; we pass along the news, useful resources, tools or services, technical analysis and exciting developments in open source. Even if you aren’t an engineer, we’ll keep you in touch with what’s going on under the hood in AI.

Stability AI announces Stable Audio

The Story: Stability AI, primarily known for text-to-image generator Stable Diffusion, has launched Stable Audio - its first AI product for generating music. Stable Audio allows users to generate music tracks up to 90 seconds long by providing text prompts describing the desired style, instruments, mood etc. The model was trained on a large dataset of music and metadata provided by stock music company AudioSparx.

More Details:

• Stable Audio generates music tracks in response to text prompts - their example "Post-Rock, Guitars, Drum Kit, Bass, Strings, Euphoric, Up-Lifting, Moody, Flowing, Raw, Epic, Sentimental, 125 BPM" gets this output. Users can also specify the desired length.

• It uses a latent diffusion architecture that enables high-quality 44.1 kHz outputs, unlike most previous text-to-audio models with lower quality sound.

• There is a free version that generates 20 second clips, and a paid "Pro" version for 90 second tracks downloadable for commercial projects.

• At an unspecified date, Stability will open source models based on Stable Audio and training code for training your own audio generation models.

Takeaways: While the samples might not seem terribly impressive, this is currently state of the art in text-to-music along with Meta’s AudioCraft. These tools have already improved quite a bit from the earliest models and will continue to get better. While image generation has received much attention, AI-generated music will enable many new creative possibilities for musicians and creators.

There are, naturally, ethical and other concerns including copyright status and compensation for musicians using AI. As with other generative AI, striking the right balance between enabling creation and protecting rights will be an ongoing challenge. But overall, Stable Audio points to the rapid evolution of creative AI and the new frontiers it is opening up beyond language models and image generation.

The event sparked skepticism from some in Congress who argued the public should not be shut out of debates on critical technology issues as tech leaders lobby behind closed doors.

Additionally, some Senators refused to attend on principle. "I think the idea that it is some great breakthrough to hear from the biggest monopolists in the world - and that they are going to share with us their great wisdom - I just think the whole framework is wrong," said Sen. Josh Hawley, R-Mo.

"You got to take it with a grain of salt. You got to realize that they're interested parties, right? They stand to make a lot of money on this, which is fine," he continued, "but you got to know that I just think the whole framing that 'Oh, aren't we so graced by their presence?' - I mean, give me a break. These people are - they've done bad things for our country."

But they are asking to be regulated. "This is sort of an important and urgent and in some ways unprecedented moment," Altman told reporters. "And I think we really need the government to lead."

That’s what most Americans want too. And in related news, Tuesday Sen. Pete Ricketts, R-Neb. introduced legislation to water-mark all AI-generated content, including enforcement rules. The bill tasks all 4 of the following bodies to lay the guidelines: Department of Homeland Security, Department of Justice, Federal Communications Commission and Federal Trade Commission.

“We’re committed to delivering the best creative tools, services and value to our members so you can keep pushing the bounds of creativity. To get started with all of the new value in Creative Cloud, you can update your apps to get access to all the new AI-powered features in Photoshop, Illustrator, Premiere Pro, and After Effects, or visit the new Adobe Firefly web app to start creating content with generative AI or Adobe Express to create content and use the new Text to Image and Text Effects features.”

Adobe is also increasing Creative Cloud subscription prices starting November 1st, citing expanded features and AI costs. Prices for students will not change.

• China exploring using AI to manipulate audiences

• (Previously linked above) Adobe, IBM, Nvidia join US President Biden’s efforts to prevent AI misuse

• (Previously linked above) GOP senator demands federal standard for AI content identification

• Microsoft president and Nvidia chief scientist testify in Senate AI hearings

• AI quietly reshapes Apple iPhones, Watches

• UK space sector has sights set on artificial intelligence and machine learning professionals

• Artificial intelligence technology behind ChatGPT was built in Iowa - with a lot of water

• Salesforce introduces Einstein Copilot Studio to help customers customize their AI

• Coca-Cola adds limited-edition drink made by artificial intelligence

Here is an example of how RAG provides relevant contextual information to a large language model for a more accurate response:

Prompt: "What is the population of Paris?"

Without RAG: The language model may try to guess or fabricate an answer, since it does not have up-to-date factual information encoded. It may say something inaccurate like "The population of Paris is 5 million people."

With RAG: The information retrieval component searches through an external knowledge source and retrieves the fact "The population of Paris is 2,140,526 as of January 2022."

• More on RAG

• Tutorial: Fine-tune your own Llama 2

• How an archeological approach can help leverage biased data in AI to improve medicine

• Meet Phi-1.5, the new language model that could make training AI radically cheaper and faster