@edfonseca_

Research Scientist @GoogleDeepMind. Sound Understanding. Previously @GoogleAI and @mtg_upf. He/him.

 NYC
eduardofonseca.net/
Joined October 2017
  • Tweets184
  • Following566
  • Followers1.1K
23Media
🔊New paper! Recomposer allows editing sound events within complex scenes based on textual descriptions and event roll representations. And we discuss the details that matter! Work led by Dan Ellis w/ a bunch of Sound Understanding folks @GoogleDeepMind arxiv.org/abs/2509.05256
4
1
42
5,871
Eduardo Fonseca retweeted
Vivek Kumar@vivek_kumar10 Sep 2025
Excited to share our work from Sound Understanding team at @GoogleDeepMind! Ever wanted to remove a single cough from a recording or make a faint doorbell louder? Recomposer makes editing complex audio scenes possible! Paper: arxiv.org/abs/2509.05256 #AudioEditing #GenerativeAI

Recomposer: Event-roll-guided generative audio editing

Editing complex real-world sound scenes is difficult because individual sound sources overlap in time. Generative models can fill-in missing or corrupted details based on their strong prior...

arxiv.org
Eduardo Fonseca@edfonseca_10 Sep 2025
🔊New paper! Recomposer allows editing sound events within complex scenes based on textual descriptions and event roll representations. And we discuss the details that matter! Work led by Dan Ellis w/ a bunch of Sound Understanding folks @GoogleDeepMind arxiv.org/abs/2509.05256
1
5
25
3,771
Eduardo Fonseca retweeted
arXiv Sound@ArxivSound8 Sep 2025
Daniel P. W. Ellis, Eduardo Fonseca, Ron J. Weiss, Kevin Wilson, Scott Wisdom, Hakan Erdogan, John R. Hershey, Aren Jansen, R. Channing Moore, Manoj Plakal, "Recomposer: Event-roll-guided generative audio editing," arxiv.org/abs/2509.05256

Recomposer: Event-roll-guided generative audio editing

Editing complex real-world sound scenes is difficult because individual sound sources overlap in time. Generative models can fill-in missing or corrupted details based on their strong prior...

arxiv.org
7
1
23
2,423
Eduardo Fonseca retweeted
Yuma Koizumi@yuma_koizumi8 May 2025
New multilingual speech restoration paper out Miipher-2 🚀! The RTF on a TPU is 0.0078: 1 million hours of data can be cleaned in 3 days using just 100 TPUs! Paper: arxiv.org/abs/2505.04457 Demo: google.github.io/df-conforme…

Miipher-2: A Universal Speech Restoration Model for Million-Hour...

Training data cleaning is a new application for generative model-based speech restoration (SR). This paper introduces Miipher-2, an SR model designed for million-hour scale data, for training data...

arxiv.org
3
31
2
83
19,378
I'm looking for a PhD student to work on continual learning for audio. Funding available for 2 years to start with, to be extended to 4 later. Contact me through email if interested! If you participated to @DCASE_Challenge or are coming to @DCASE_Workshop even better!
3
32
2
112
15,285
🔊 We've released pre-trained models & code for our ICCV23 paper, Audiovisual Masked Autoencoders!! GitHub: github.com/google-research/s… Paper: arxiv.org/abs/2212.05922 Work led by Lili Georgescu and @anuragarnab , with Radu Ionescu, @MarioLucic_ and @CordeliaSchmid
5
2
46
3,540
Eduardo Fonseca retweeted
Vivek Kumar@vivek_kumar4 Oct 2023
It's so awesome to see the impact of the computational audio capabilities we developed featured in @madebygoogle 🎉 🎉 🎉 Congrats to John Hershey, @ScottTWisdom, @PGetreuer & everyone who contributed for pioneering new computational audio capabilities in Pixel8 #MadeByGoogle
Google Photos
@googlephotos4 Oct 2023
Check out the 4 new Google Photos features coming first to Pixel 8 and 8 Pro ↓ Whether it’s noise from wind, traffic, or barking dogs, Audio Magic Eraser in Google Photos reduces distracting sounds in your video in just a few taps! 🪄
4
16
3
59
20,324
🔊New paper out: Do you use data balancing in your AudioSet experiments? It gets you a little mAP boost? It might work differently than you think...😅 You might want to check our last paper, led by @ChannningMoore
Channing Moore@ChannningMoore5 Jul 2023
Our 2023 ICASSP paper is now up on arXiv: Dataset balancing can hurt model performance arxiv.org/abs/2307.00079 Dataset balancing works differently than you might assume: - can cause overfitting; - doesn’t improve performance on rare classes; - speeds up training convergence.

Dataset balancing can hurt model performance

Machine learning from training data with a skewed distribution of examples per class can lead to models that favor performance on common classes at the expense of performance on rare ones....

arxiv.org
1
16
2,357
Not sure my reviews were that "outstanding"... 😅, but the recognition is nice...  Thanks to the @ieeeICASSP committee.  #ICASSP2023
27
1,860
📢 The countdown to DCASE Challenge 2023 deadline is on! 🗓️ Deadlines: * System submission: 15/5 - 23.59 AoE * Technical reports: 22/5 - 23.59 AoE
2
1
6
1,002
Eduardo Fonseca retweeted
Romain Serizel@RSerizel6 Jul 2022
📣 The results for @DCASE_Challenge task 4 are finally out! 🥳 dcase.community/challenge202… @Nicoturpo @FraRonchini @edfonseca_ @SamueleCornell
1
4
17
🔊 A bit late, but happy to announce that I recently joined Google Research! I’m working in the Sound Understanding Group based out of NYC! research.google/people/Eduar…
8
7
154
Aaand that was after defending my PhD thesis some months ago "Training Sound Event Classifiers Using Different Types of Supervision" & taking some time off :) Thesis/video/slides & a quick summary available here: eduardofonseca.net/thesis/20… SUPER thankful to all @mtg_upf folks!!🙌
13
Eduardo Fonseca retweeted
HEAR Benchmark@hearbenchmark29 May 2022
HEAR PMLR journal submissions are open until 2022-06-30. neuralaudio.ai/hear2021-pmlr… Besides that, people have asked if they can run HEAR benchmarks, get on the leaderboard, cite us in the future. Yes! HEAR is here to stay. See our updated website: neuralaudio.ai/
5
2
17
Eduardo Fonseca retweeted
Romain Serizel@RSerizel25 May 2022
📣 We have a (super cool) PhD position in speech enhancement for patients with auditory neuropathy spectrum disorders. If you're interested in Audio Signal Processing/MAchine Learning/Audiology, contact us! More info ⤵️ adum.fr/as/ed/voirpropositio…
1
10
14
Eduardo Fonseca retweeted
Romain Serizel@RSerizel13 May 2022
Looking forward to seeing you all here! 🥳
DCASE Workshop@DCASE_Workshop13 May 2022
📢The #DCASE2022 workshop call for papers is out 🥳🎉 dcase.community/workshop2022… The abstract submission deadline is on 7th of July and the workshop will be held in presence from 3rd to 4th of November in Nancy. Looking forward to see you there! 😉
1
2
Eduardo Fonseca retweeted
daisukelab@nizumical27 Apr 2022
Our new paper is out! We explored a simple masked patch modeling w/o augmentation to learn a latent that describes the input spectrogram as it is. “Masked Spectrogram Modeling using Masked Autoencoders for Learning General-purpose Audio Representation” arxiv.org/abs/2204.12260

Masked Spectrogram Modeling using Masked Autoencoders for Learning...

Recent general-purpose audio representations show state-of-the-art performance on various audio tasks. These representations are pre-trained by self-supervised learning methods that create...

arxiv.org
2
2
29
Eduardo Fonseca retweeted
daisukelab@nizumical18 Apr 2022
arxiv.org/abs/2204.07402 For BYOL for Audio, an updated paper is out (submitted last year, still under review). It extends the initial BYOL-A for network architecture and data augmentation. We compare with 8 models (11 representations) using a benchmark with 10 tasks.

BYOL for Audio: Exploring Pre-trained General-purpose Audio Representations

Pre-trained models are essential as feature extractors in modern machine learning systems in various domains. In this study, we hypothesize that representations effective for general audio tasks...

arxiv.org
arXiv Sound@ArxivSound18 Apr 2022
``BYOL for Audio: Exploring Pre-trained General-purpose Audio Representations. (arXiv:2204.07402v1 [eess.AS]),'' Daisuke Niizumi, Daiki Takeuchi, Yasunori Ohishi, Noboru Harada, Kunio Kashino, ift.tt/b1pr6st
1
1
8
#DCASE2022 Challenge is officially open! You can now check the task descriptions and development data. Some tasks have delays with the baseline system, but those will be ready soon too. dcase.community/challenge202… @DCASE_Challenge #machinelistening #DCASE
11
21
📢 DCASE challenge 2022 task descriptions are out!! Enjoy ➡️ dcase.community/challenge202…
12
1
22
Load more