About me
- Alexandre Défossez, defossez at fb.com.
- Research scientist at FAIR Paris.
-
Formerly
CIFRE PhD student at FAIR Paris and
Sierra at INRIA Paris,
under the supervision of Léon Bottou (FAIR), Nicolas Usunier (FAIR) and Francis Bach (INRIA).
-
[scholar]
[github]
[twitter]
Interests
Convex and non convex optimization, stochastic gradient methods, source separation, audio processing and synthesis.
Amateur DJ and composer [soundcloud].
Publications
-
A. Défossez, G. Synnaeve, Y. Adi.
Real Time Speech Enhancement in the Waveform Domain.
Interspeech 2020.
[preprint]
[audio samples]
[code]
-
A. Défossez, L. Bottou, F. Bach, N. Usunier.
A Simple Convergence Proof of Adam and Adagrad. Preprint 2020.
[preprint]
-
A. Défossez, N. Usunier, L. Bottou, F. Bach.
Music Source Separation in the Waveform Domain. Preprint 2019.
[preprint]
[github]
[audio samples]
-
A. Défossez, N. Usunier, L. Bottou, F. Bach.
Demucs: Deep Extractor for Music Sources with extra unlabeled data remixed. Preprint 2019.
[preprint]
-
L. Perotin, A. Défossez, E. Vincent, R. Serizel, A. Guérin
Regression versus classification for neural network based audio source localization.
WASPAA 2019.
[paper]
-
A. Défossez, N. Zeghidour, N. Usunier, L. Bottou, F. Bach.
SING:
Symbol-to-Instrument Neural Generator. NIPS 2018.
[paper]
[github]
[poster] [audio
samples]
[slides].
-
A. Défossez, F. Bach.
AdaBatch: Efficient Gradient Aggregation Rules for Sequential and Parallel
Stochastic Gradient Methods. Preprint 2017.
[preprint].
-
A. Défossez, F. Bach.
Constant step size least-mean-square: Bias-variance trade-offs and optimal
sampling distributions.
AI Stats 2015.
[AI Stats version],
[arXiv version].
Software
- Julius:
Efficient implementations of classical Digital Signal Processing algorithms in PyTorch,
fully differentiable and with CUDA support. Resampling, FFT based convolutions,
FIR low pass filters and decomposition of a signal over multiple frequency bands in the
waveform domain are implemented.
- Denoiser:
Real time speech denoising in the waveform domain. Can be used with Zoom or other
VC software with a virtual soundcard (e.g. Soundflower on a Mac). Live demo :)
- Demucs:
Music source separation in the waveform domain. Can separate drums, bass, and vocals
from the rest of the accompaniment. Jaime Altozano loves it!
Teaching
2018
Teaching assistant for the
Deep Learning: Do-It-Yourself!
class at Ecole Normale Superieure:
Misc.
I wrote my PhD manuscript on the
Optimization of Fast Deep Learning Network for Audio Analysis and Synthesis.
Half of it is on audio synthesis and source separation, and the other half is on adaptive and
stochastic optimization.