Dimitrios Bralios
@DBralios
Followers
63
Following
775
Media
2
Statuses
12
AI + Audio PhD Student, UIUC
Joined October 2020
Great audio AEs/codecs exist, but when you need structured latents or a tweaked bottleneck for a downstream task (e.g. generation), retraining is expensive & brittle. We Re-Bottleneck👇
Dimitrios Bralios, Jonah Casebeer, Paris Smaragdis, "Re-Bottleneck: Latent Re-Structuring for Neural Audio Autoencoders,"
1
5
20
Paper (🏆 IEEE MLSP 2025 Best Paper): https://t.co/9wOrzTgxar Code: https://t.co/q83LcCgypx In collaboration with @CasebeerJonah, @Psmaragdis See also our WASPAA’25 paper on latent-domain upsampling/upmixing:
0
0
6
We measure latent “diffusability” by training text-to-audio diffusion models. Showing in practice how our framework enables rapid prototyping.
1
0
2
We demonstrate: - Ordered latents for channel ranking and rate control. - Equivariance between input and latent spaces under specified transforms (latent filtering). - Semantic latents by contrastive alignment to external audio or text embeddings (e.g. BEATs/T5).
1
0
2
Our method is simple: we add a small inner AE and train only in latent space (latent reconstruction + latent discriminator). No waveform objectives, base AE stays untouched.
1
0
2
Re-Bottleneck: Latent Re-Structuring for Neural Audio Autoencoders.
arxiv.org
Neural audio codecs and autoencoders have emerged as versatile models for audio compression, transmission, feature-extraction, and latent-space generation. However, a key limitation is that most...
0
2
6
I'm thrilled to announce that our paper, "Generation or Replication: Auscultating Audio Latent Diffusion Models" 🩺 with the Speech & Audio team at MERL, has been accepted for publication at #ICASSP2024!
1
1
14
``Generation or Replication: Auscultating Audio Latent Diffusion Models. (arXiv:2310.10604v1 [ https://t.co/3pcQCkeyAA]),'' Dimitrios Bralios, Gordon Wichern, François G. Germain, Zexu Pan, Sameer Khurana, Chiori Hori, Jonathan Le Roux,
0
4
9
Complete and separate: Conditional separation with missing target source attribute completion Dimitrios Bralios, Efthymios Tzinis, Paris Smaragdis https://t.co/rv8Q4j32Yf
arxiv.org
Recent approaches in source separation leverage semantic information about their input mixtures and constituent sources that when used in conditional separation models can achieve impressive...
0
1
3
