Pierre-Amaury Grumiaux

- Built end-to-end deep learning pipeline from scratch: dataset creation with custom annotations and audio augmentation strategies, model architecture, training, and evaluation with custom metrics - Researched and implemented state-of-the-art audio processing techniques for audio event detection - Optimized models for embedded systems through compression, quantization, latency reduction, and memory optimization while balancing accuracy vs. computational constraints - Prepared and deployed models for production environments - Led technical strategy: defined AI roadmap, prioritized features, and integrated customer feedback into model improvements - Collaborated cross-functionally with product, hardware, and backend teams while maintaining technical documentation

Bandwidth extension of music signals with deep learning models.

Deep learning for speaker counting and localization with Ambisonics signals • State-of-the-art review on sound source counting and localization with signal processing and deep-learning based methods. Survey article on DL-based sound source localization • Multi-source speech signals dataset generation with reverberation and noise, based on real and simulated spatial impulse responses (Ambisonics format); • Implementation and evaluation of neural networks for speaker counting on multi-channel signals: CNN, RNN, CRNN • Implementation and evaluation of neural networks for speaker localization on multi-channel signals: CNN, RNN, CRNN, dilated convolutions, residual connections, self-attention mechanism • Investigations on a novel input feature representation of a sound field (Time Domain Velocity Vector) for single-speaker localization • Implementation and evaluation of DL-based hybrid methods for joint speaker counting and localization: number of sources injection, multi-task networks • Co-supervision of several research interns

Automatic drum transcription using deep neural networks • State-of-the-art review on automatic music transcription, especially automatic drums transcription • Implementation and evaluation of neural networks for automatic drums transcription : RNN, CNN, CRN • Implementation and evaluation of the addition of beats and downbeats information to improve automatic drums transcription • Implementation and evaluation of the student-teacher paradigm for automatic drums transcription (leading to the scrapping of YouTube music videos for a big training dataset)

Audio-to-lyrics alignment for polyphonic music • State-of-the-art review of lyrics-to-audio alignment • Creation of a training dataset based on TIMIT signals • Implementation of a state-of-the-art methods based on dynamic time warping and hidden Markov models jointly with a proprietary system of voice extraction.

Physical modeling based sound synthesis tools in Faust language • Learning about waveguide physical modeling and Faust language • Implementation of several physical modeling modules in Faust (excitation elements, strings, terminations) • Creation of two Python scripts for fast creation of Faust physical models o Modal model from an impulse response o Modal model from a set of geometrical and material properties

As a technician for one month, my job was to meticulously edit videos. These videos are to recognize face movements to control audio effects. The face becomes an audio controller.