Bernardo Torres

PhD Student @TelecomParis, @tp_adasp.
Previously intern at @researchdeezer @SonyCSLMusic.
AI/ML for audio and music signal processing and synthesis.

[Bernardo Torres @torres_be_]

@stark_reborn @AssafShocher Actually it's not really the same approach since the encoder/decoder don't follow the linearizer architecture they propose.
But I would be really interested in simply trying to replace the encoder/decoder with their linearizer and a similar compression ratio to see if it works!

[Bernardo Torres @torres_be_]

@stark_reborn @AssafShocher Hello! I might be a little late for the party but this is exactly was was done in the following paper (for pre-training):

https://t.co/TqcyoOOJ9V

Check it out!

[Alain Riou @howariou]

We just released the full training code, as well as our best pretrained model! 🎉

Feel free to use our SOTA checkpoint in your own project with 3 lines of code, or to retrain on your own data using our Lightning+Hydra+Dora codebase ⚡️🐍
🌐 https://t.co/gzloylhihP

[Bernardo Torres @torres_be_]

+ if you work on MIR and use the CQT for tasks which require frame-wise/time-varying estimation, consider using the Variable-Q Transform (VQT) instead.

The huge CQT kernels at low frequencies are often underestimated and they might be hurting your performances.

[Bernardo Torres @torres_be_]

The PESTO extension paper was published in TISMIR!

Here's the link to the publication:

https://t.co/0osdiF4Toz

It features several improvements to the model, larger cross dataset evaluation, and a real-time implementation.

[Alain Riou @howariou]

Estremamente felice di annunciare that our new PESTO recipe has been published in the famous TISMIR cookbook 👨‍🍳🤌

With chef @torres_be_, we revisited this traditional sauce invented at ISMIR 2023 in Milan with some Brazilian flavours 🇮🇹🇧🇷

https://t.co/M9JRzUbQxN

[Zineb Lahrichi @zineb_lahrichi]

🔉New paper out!

Recent audio codecs typically learn compression and quantization jointly, limiting the choice of quantization layers, non-differentiable by definition.

What if we used powerful neural quantizers like Qinco2 and trained them offline?

https://t.co/UJBOTJKwui

[Bernardo Torres @torres_be_]

@RomyBeaute Amazing! Great work :)

[Bernardo Torres @torres_be_]

We’ve added fresh ingredients and cooking tricks to make the best, lightest, and fastest neural pitch estimator even better! 🌿🔥

Shoutout to all collaborators and to the amazing chef & SSL titan, @howariou

[Alain Riou @howariou]

PESTO 2.0 è rilasciato! 🥳🥳🥳
With Brazilian chef @torres_be_ (and others), we revisit this traditional italian sauce, invented in Milan at @ISMIRConf 2023 🇮🇹

And you can taste it in REAL-TIME at home (~5 ms latency) ⏱️

1/6

[Alain Riou @howariou]

PESTO 2.0 è rilasciato! 🥳🥳🥳
With Brazilian chef @torres_be_ (and others), we revisit this traditional italian sauce, invented in Milan at @ISMIRConf 2023 🇮🇹

And you can taste it in REAL-TIME at home (~5 ms latency) ⏱️

1/6

[Stefan Lattner @deeplearnmusic]

🌟My keynote at the @c4dm workshop about "Models of Musical Signals: Representation, Learning & Generation" is now on YouTube, giving an overview on developments in self-supervised learning for audio since 2020, low-level representation learning, audio (stem) generation and much more 🧵👇

https://t.co/k3lTKNkzD3

@SonyCSLMusic @SonyCSLParis

[Bernardo Torres @torres_be_]

@howariou 👀

[Victor Letzelter @ ICML 2026 @VLetzelter]

Interested in ill-posed learning tasks, uncertainty prediction, conditional density estimation or multi-head deep neural networks ?

In our new paper, accepted at #ICML24, we tackle these challenges by exploring the Winner-Takes-All (WTA) training scheme.
[1/n]

[Bernardo Torres @torres_be_]

Signals with a high degree of autocorrelation (such as pitched signals) make the training of convnets on raw audio unstable. For Gaussian initialization, the greater the input’s autocorrelation, the greater the variance of the output.

Huge relief seeing these kinds of papers <3

[Vincent Lostanlen @lostanlen]

Training convnets on waveforms is hard—far harder than on magnitude spectrograms.

"Instabilities in Convnets for Raw Audio" approaches this phenomenon from the perspective of sensitivity to initialization.

IEEE Signal Processing Letters vol. 31
preprint: https://t.co/Cj6BAGffQR

[Bernardo Torres @torres_be_]

So many amazing people in this photo!

[Peeters Geoffroy @GeoffroyPeeters]

@adasp group from ⁦⁦@telecomparis⁩ @ieeeICASSP⁩ 2014, Seoul, Korea

[Bernardo Torres @torres_be_]

Great stuff!

[YCY @yoyolicoris]

https://t.co/KIPD0aU7ti
The paper preprint is out! We'll make an official announcement soon but feel free to have a look first! :)

[Bernardo Torres @torres_be_]

@92HsChoi Hello! We struggled to make it work out of the box for more realistic data on a DDSP-like setting. The loss provides the gradient to move in frequency, but having to estimate the n. of harmonics + F0 creates many local minima/instabilities.
Getting past that is work in progress!

[Bernardo Torres @torres_be_]

Happy to share some work accepted to ICASSP :)

We experiment with a loss function inspired by optimal transport to compare spectra.

We test it on a synthesis-based frequency localization (and F0 estimation) toy task using a harmonic synthesizer.

Paper: https://t.co/zTwBw1IJvD

[Peeters Geoffroy @GeoffroyPeeters]

10 papers from the Audio group @tp_adasp of Télécom-Paris will be presented at @icassp2014 @ieeeICASSP.
---
Paper #1: Bernardo Torres et al. "Unsupervised Harmonic Parameter Estimation Using Differentiable DSP and Spectral Optimal Transport"

[Bernardo Torres @torres_be_]

The SOT loss is based on the p-Wasserstein distance, which has a closed-form solution in 1D.

First we compute the cumulative function of both spectra. Then, we invert it to get the quantiles. We add up the differences (raised to power p) and that’s it!

[Bernardo Torres @torres_be_]

We call this loss Spectral Optimal Transport (SOT). Compared to L1 and L2 spectral losses on sinusoidal signals, SOT has a very nice convex curve leading to the right oscillator frequency.

Its gradient also does not vanish when the frequency difference is high.

[Stefan Lattner @deeplearnmusic]

🥳 We present our #ICASSP2024 paper:

A diffusion model that generates production-quality (bass) audio stems to any audio input. 😎🎸 According to our experience, that's more useful to artists than generating full mixes. 🙃

📜Paper: https://t.co/3MMhHTz9VF
🎶Demo: https://t.co/ivCtWolSPK

by @marco_ppasini 👈💪@SonyCSLMusic #MusicAI