A fun exchange from this week, I'm still unpacking: "Would it be OK if we remove everything with the labels "noise music" and "death-metal" from the training set? It make a perfectly working generator feel broken"
tag > ML
-
-
Spotify, a sleeping giant?
Puzzled by the lack of focus on AI at Spotify's StreamOn event today. Given the enormous potential of AI in music and its rapid advancements, it's surprising that they had zero related news, besides a brief mention of their new "AI DJ", which feels like an afterthought product.
-
I'm in love with Audio Superresolution ♥️
All those music records made from 1900 > 1960 that are so wonderful but have terrible quality? Audio Super-Resolution let's us reimagine them in pristine modern quality. Wild times.
-
Generative Music Tech Bro Communities
One of the prominent generative AI music communities distinctly feels like a ultra geeky tech bro festival, akin to hardcore gaming communities. Would Charlie Parker and Miles Davis (or any of the leading musicians of the past 100y) have intermingled with it?
-
Generative Bedtime Stories
A simple yet powerful app using Generative Voices + GPT + Apple Watch: "Bedtimes stories" where the style of the voice & story contents are continuously adapted based on biofeedback, to optimally put you to sleep. Such "Closed-Loop Feedback" tools will become very popular soon.
Illustration from research on "Closed-loop auditory stimulation"
-
CLAP: Contrastive Language-Audio Pretraining: Paper - Code & Model - Dataset
#Comment: Sometimes it's the simple things that bring joy to the heart: Training an ML model on a few million songs plus meta-data, produces a formidable "neural music search & recommendation engine". A very useful component for downstream applications.
-
Companies in the Generative AI space that do not have a well-planed strategy for creating a "clean dataset" free of copyright issues, are short-sighted. Investing in a best-effort approach now can unlock significant opportunities down the line.
-
Prediction: Neural Rendering will be THE topic across the games industry in 2024: generative AI pipelines that convert low quality (low FPS/poly/style) game graphics into near photorealistic ultra smooth graphics on the fly.
-
Illustration from a research project i published back in 2017, that explored generating music with ML guided by Musical Novelty Search. Fun time to revisit, as the tech has gotten dramatically better since.
-
Over the last hundred years, statisticians have uncovered a whole universe of statistical distributions and relationships between them.
-
Sounding The Secrets Of AudioLM - "Dive into the inner workings of Google AI’s language modeling approach to audio generation"
-
Prediction: "Dance-To-Music Generation" will become a big deal on social platforms and beyond.
(Image from: https://github.com/L-YeZhu/CDCD) - #Music #ML #Augmentation #Generative
