About Education Experience Research Contact
AI Scientist — Thales Research & Technology

Damien Martins Gomes

I build speech and audio models that hold up on real hardware, under real constraints — turning degraded signals into intelligence, one waveform at a time.

Speech Enhancement
Denoising & dereverberation, real-time on-device
Neural Audio Codecs
Ultra-low-bitrate compression for edge devices
Audio Generation
Flow matching & foundation models
Multimodal Fusion
Cross-modal signal representations
Text-to-Speech
Natural, low-latency voice synthesis
Audio Recognition
Robust ASR under real-world noise
Scroll
About

Who I am

Damien Martins Gomes

AI Scientist at Thales Research & Technology, I work at the intersection of deep learning and signal processing, designing intelligent audio systems under real-world constraints. My research spans speech enhancement, embedded and real-time models, as well as generative audio systems where efficiency, robustness, and reliability are critical.

I recently completed a research master's degree in Computer Science at Concordia University, alongside a dual degree in Aerospace Engineering specialized in Telecommunications at IPSA Toulouse. This interdisciplinary background allows me to bridge physical systems, signal processing, and modern machine learning to tackle complex, high-impact problems.

Previously at Mila – Quebec AI Institute, I developed AdaFisher, a second-order optimization algorithm accepted at ICLR 2025. Driven by curiosity and a strong research mindset, I enjoy turning raw signals into intelligent systems and exploring how AI can robustly perceive and understand the world.

Machine Learning Deep Learning Dynamics Signal Processing Physics
Education

Education

Concordia UniversityMontreal, CA

Concordia University

MCompSc: Master of Research in Optimization and Machine Learning

SEP 2023 — APR 2025

CGPA 4.1 / 4.3

Thesis awarded with "Outstanding" distinction

Relevant Coursework

  • Representation Learning (MILA Institute)
  • Geometric Data Analysis (MILA Institute)
  • Algorithm Design Techniques
  • Parallel Programming
IPSA ToulouseToulouse, FR

IPSA Toulouse, France

MSc in Aerospace Engineering: Embedded Systems and Signal Processing

SEP 2019 — APR 2025

CGPA 3.8 / 4.0

Obtained with "Outstanding" distinction

Relevant Coursework

  • Non-Linear Optimization, Numerical Linear Algebra
  • Distributed Intelligent Systems, Guided Propagation
  • Advanced FPGA Circuits, On-board Networks
  • Real-Time Information Systems
Experience

Experience

Thales

AI Scientist

May 2025 — Now

  • Designed end-to-end audio AI systems, from research and model development to integration support on edge and embedded devices
  • Developed lightweight, real-time speech enhancement models under strict latency, memory, and compute constraints for embedded deployment
  • Conducted research on generative audio models, including flow matching and audio foundation models, and adapted them to multiple downstream tasks
  • Investigated hallucination phenomena in generative audio models, focusing on their quantification and mitigation
  • Worked with diverse model architectures such as Transformers and Mamba, contributing to three industrial patents

MILA — Quebec AI Institute

AI/ML Research Student

September 2023 — April 2025

  • Developed AdaFisher, a novel second-order optimizer that significantly improves performance over Adam in image classification and language modeling
  • Designed and ran large-scale training experiments on vision and LLMs (GPT-1), optimizing hyperparameters, fine-tuning architectures, and scaling distributed training
  • Led theoretical research on neural network optimization, curvature-aware updates, MVPs, and second-order methods
  • Served as a reviewer for top-tier conferences: ECCV 2024, NeurIPS 2024, CVPR 2025, ICML 2025
  • Supervised by Dr. Eugene Belilovsky, Dr. Guy Wolf, and Dr. Mahdi S. Hosseini

Yale University × MILA

AI/ML Research Collaborator

September 2024 — January 2025

  • Led research on developing novel neural network architectures based on an innovative mathematical framework
  • Implemented and optimized architectures within LLMs to enhance performance and computational efficiency
  • Developed methods to capture complex oscillatory patterns in NLP and speech processing tasks
  • Collaborated with Dr. Yanlei Zhang and Yale University research team

Murmuration SAS

Machine Learning Engineer

June 2023 — August 2023

  • Contributed to the European research project "DeepCube" developing a sophisticated price engine model
  • Implemented constrained optimization algorithms using SciPy and reinforcement learning models based on DQN
  • Integrated a Generative Adversarial Network (GAN) for synthetic data generation

EuroMoonMars

Commander / Analog Astronaut

June 2021 — September 2021

  • Commanded the EMMPOL7 mission simulating lunar base conditions, organized by EMM, ILEWG, and AATC
  • Conducted reinforcement learning experiments on rover navigation in challenging terrain
  • Designed payload for a lunar launcher in CATIA for 3D printing
Research

Publications

AdaFisher Research
ICLR 2025 · Main Conference

AdaFisher: Adaptive Second-Order Optimization via Fisher Information

We propose AdaFisher, a novel second-order optimizer that integrates Fisher information into the Adam family. AdaFisher achieves consistently faster convergence and improved generalization compared to Adam, AdamW, AdaHessian, and Shampoo across image classification and language modeling tasks, while remaining computationally efficient.

GitHub OpenReview
Preprint · 2025

BDN: Blaschke Decomposition Networks

A neural architecture that iteratively "unwinds" real- and complex-valued 1-D and 2-D signals into interpretable oscillatory components via the Blaschke decomposition, demonstrated on ECG and phase holographic microscopy data.

Paper →
arXiv · 2026

BASENet: Band-Adapted Speech Enhancement Network with Cross-Band Attention

A Bark-scale, band-adapted architecture with cross-band attention for real-time speech enhancement — PESQ 3.55 / STOI 96% on VoiceBank+DEMAND with only 0.83M parameters, the fewest among methods at that quality bar.

arXiv →
MobiSys 2026 · Paper

Design and Deployment of Lightweight Neural Speech Enhancement and AI Codecs on GAP9 for Ultra-Low-Bitrate Edge Communication

With C.B. Huynh, R. Lampe, M. De Closets, F. Sausset, S. Lorin. On-device neural speech enhancement and codecs running on the GAP9 embedded processor for ultra-low-bitrate edge communication.

ACM DL →
MobiSys 2026 · Demo

VoxClean — AI-Powered Speech Enhancement for Tactical and Mobile Communication Pipelines

With S. Lorin. A deployed speech enhancement pipeline for tactical and mobile communications, presented as a demo at the 24th ACM International Conference on Mobile Systems, Applications, and Services.

ACM DL →
Master's Thesis · 2025

Towards Practical Second-Order Optimizers in Deep Learning: Insights from Fisher Information

A unified analysis of zeroth-, first-, and second-order optimization methods, culminating in a detailed theoretical and empirical study of AdaFisher.

Manuscript →
Honors

Honors & Awards

Fonds de recherche du Québec (FRQNT) Masters (B1X) Scholarship

2024 — 2025

$26,667

Concordia University Split Merit Scholarship

2023 — 2024

$5,000

Ivan Velan Student Award

2023 · Outstanding athletic performance

$1,000
Archive

Earlier Builds

Numerical Methods

N-Bodies Problem

Analytical & numerical study of the N-body problem, applied to Voyager 2's trajectory.

 PDE Simulation

HeatWave

Numerical solution of the heat equation in one and two dimensions.

 Signal Processing

PixelPulse

Webcam-based heart rate detection via Fourier analysis of forehead pixel intensity.

 Sequence Modeling

LinguaNet

Encoder/decoder neural machine translation from English to German in TensorFlow.

 Systems / C++

ConquerZone

A Warzone-compatible implementation of Risk, built in C++.
Contact

Get in touch

Location

Thales Research & Technology
1 Avenue Augustin Fresnel
Palaiseau, 91120, France

Email

damien.martins.gomes@gmail.com
Connect