@dadabots not sure how well it will work at a 4096 waveform patch size but this is one of the more fun things to do with "just" a codec
English
Patrick O'Reilly
67 posts
Patrick O'Reilly
@ReillyOPatrick
PhD student in @NorthwesternU's Interactive Audio Lab, researching audio & ML
Katılım Ekim 2021
173 Takip Edilen180 Takipçiler
everyone you know is vibecoding generative loopers with a b&w color scheme
Sonic Field@sonic_field
ORAM v0.3 is out ✿ New additions: - loop zone variations powered by @stableaudio inpainting - proper master recorder - fade in / fade out - undo / redo - per-layer reverse playback - nuke now actually nukes everything - quieter, quicker, prettier FX palette - smoother GUI
English
of course the cool new causal latent GAN paper is from the Musika author (+friends)
Marco Pasini@marco_ppasini
Meet LiveBand: a real-time AI jamming companion! 🎸 It generates live music accompaniments with zero perceived latency ⚡️ It runs locally on Macbooks, can generate any instrument (more than one at a time), is wildly robust, and is trained from scratch on a single GPU! 🧵👇
English
from a mixing standpoint I've always been blown away by how well treats wrestles a bunch of extremely harsh and peaky sounds into a listenable whole, wouldn't have the same venom if they just fed it into a "resonance suppressor"
English
feel like at least 50% of this is just letting guitars & vocals annihilate the upper midrange in ways that would terrify today's soothe-brained producers
kenjataimu jones 🌹@originaldaveyj
How did they make that one Sleigh Bells album so loud
English
for future discourse farming I would recommend producing from scratch and then passing through an old VAE
HuangF@Iam_HuangF
I don’t usually join challenges like this, but this sample really intrigued me :0
English
and also that you can engreeble audio by adding a lil noise
English
I do love that CoDiCodec's ambient latent space is a greeble choir
English
Slides and code for our Interspeech tutorial are up now!
Patrick O'Reilly@ReillyOPatrick
Had a blast presenting a tutorial on speech watermarking at #Interspeech2025 in Rotterdam! If you’re in town and want to talk watermarking, definitely reach out! And while you’re here, I recommend the pinball museum for the big lebowski sensory overload chamber
English
stemgen/diff-a-riff I will never forget you
Geraldo Ramos@geraldoramos
We just released a new instrument-based AI music model at @moises_ai. Check it out!
English
Had a blast presenting a tutorial on speech watermarking at #Interspeech2025 in Rotterdam!
If you’re in town and want to talk watermarking, definitely reach out! And while you’re here, I recommend the pinball museum for the big lebowski sensory overload chamber
English
train on strongly metered rhyming lyrics —> model learns to associate phrases with similar metric & phonetic structure, even if semantically nonsensical
very Max Martin coded tbh
Zachary Novack@zacknovack
Suno + Veo 3 generate highly similar versions of popular songs purely based on *phonetically* similar gibberish lyrics?!?! Presenting Bob’s Confetti: Phonetic Memorization Attacks in Music and Video Generation 🔊: jrohsc.github.io/music_attack/ 📖: arxiv.org/abs/2507.17937 🧵1/n
English
@serrjoa Wouldn't that just be @marco_ppasini's Music2Latent2? With the register-like summary embeddings etc.
English
Creating first the audio version of this TiTok tokenizer. And then, instead of inpainting, temporal continuation.
English
Let's do this for audio, no?
Kwang Moo Yi@kwangmoo_yi
Preprint of today: Beyer et al., "Highly Compressed Tokenizer Can Generate Without Training" -- github.com/lukaslaobeyer/… The latent space of tokenizers already provides a good enough abstraction to work with -- you don't have to use a diffusion model on top to inpaint, etc!
English
@TeeJayBaker @latentspaces @SonyCSLMusic @SonyCSLParis Very cool work! And nice to see the comparison to a Sketch2Sound-like variant (LiLAC*) w/o fine-tuning
English
Hey! Our paper, 🌸 “LiLAC: A Lightweight Latent ControlNet for Musical Audio Generation” got accepted as ISMIR 2025!
It presents a novel control paradigm for audio diffusion models with greatly reduced parameters, encouraging users to train individual, modular models.
English
@ZDi____ If you’re trying to tune a spectrogram codec, might be worth checking out: arxiv.org/abs/2406.17672
English
Audio language modeling has always involved people training codecs to VQ audio directly. But what if we tokenized mel spectrograms, then trained a vocoder like iSTFTNet, and our AR prior on mel spectrogram indices? We can easily language model 44.1KHz audio with a single codebook
English
@serrjoa Presumably this excludes watermarking? But yeah, definitely feels like we’re entering the “adversarial attacks on medical imaging” era of deepfake detection papers
English
Hot take: Papers on deepfake detection should be desk-rejected*. Especially if they are based on classifiers.
* Based on the reasoning that if generative models mimic the training data distribution sufficiently well, all deepfake detectors can pick is just spurious correlation.
English