Description: I am really interested in AI research, and I love music. Therefore since september I decided to pursue my own independent research on deep learning topics applied to audio, speech and music. It involves thorough literature review on these topics, implementation of existing models and new paradigms. I mostly use Python for the implementation and libraries such as PyTorch, transformers, accelerate, etc... For heavy computing, I have access to Compute Canada cluster through my informal collaboration with the Mila-Québec institute of AI. My current work investigates neural audio editing (zero-shot mostly) based on diffusion models and inversion methods. I'm also trying to use a consistency model as the generative model and perform a novel "inversion" method that fits the consistency sampling. My previous project aimed to perform neural instrument modification on melodies using a triplet dataset (base melody, target text, target melody). I created the triplet dataset and trained a Latent Diffusion-Transformer model inspired by TangoFlux model. I worked mainly on the latent space of Music2Latent consistency autoencoder.

Keywords: Generative AI, Audio and Music generation, Deep Learning, Diffusion Models, Inversion, Consistency Models, AR Models, DSP, MIR

Description: As part of my end-of-studies internship, I was a research intern at Laboratoire Quosséça at Polytechnique Montréal under the supervisions of Gilles Pesant. My subject was "Conception and Implementation of a neuro-symbolic AI model for music melodies generation". The objective was to combine constraint programming and deep-learning models to improve the quality, the structure and the control over the generated melodies. For that I trained a transformer based model that generates a musical melody from a chord sequence. I added a constraint programming model to it with several constraints that are musically relevant. I also added to the inference, a beam search method to optimize the model and influence the generated result. . I have implemented an interface linking the whole architecture, which promotes collaboration with humans, and helps to increase testing and examples. I also set up quantitative (based on existing tools) and qualitative evaluation methods for the model (Turing Test), which have enabled me to judge the quality of the proposed model more accurately. Whether quantitatively, where our measurements show a performant model, or qualitatively when looking at the results of the Turing test I obtained satisfactory conclusions regarding the overall quality of the model. The result melodies are more diversified and more musical. For all that, I used mostly Python for the deep learning model, Java for the constraint programming model. I used a UNIX environment for the whole project and I used a Compute Canada (national Canadian cluster dedicated to research) to train the model. Apart from that I also did a lot of paper reading.

Keywords: Artificial Intelligence, Machine Learning, Deep Learning, Constraint Programming, Music Generation

Description: I carried out this internship between August 2021 to January 2022. It took place in the startup Otoqi, a company which aims to serve the car-sharing, car rental, new mobility and broad automotive industry. There I was a software engineer / developer and a data engineer intern. I worked in the technical team (a small young team - 4 people). As it was a growing start-up at the moment, I did have a lot of responsibilities. The tech team mostly used the low-code software Decisions, and I was in charge of the software issues resolution with the help of Decisions team. I was also in charge of the creation of a development environment for the team and of database management. Beyond these core tasks, I had lots of different classic tech mission (preparing features for clients, optimization of code, documentation, bug solving...). Apart from Decisions, I used mainly : Kibana (ElasticSearch), Python, Azure and MongoDB. Most of my internship was in English as we were working in an international environment.

Keywords: DB management, Software Engineer, Programming, Fast-paced start-up company