Vikentii Pankov

Experience

• 

  Gazpromneft Digital Solutions

  Jun 2019 - Dec 2021

  Lead Developer

  Development of a predictive system based on ML, statistical, and optimization methodsTechnologies: C#, PythonKey Achievements:1) Implemented features for associated petroleum gas production forecasting using C#, DBSCAN, linear regression, and derivative-free optimization methods.2) Developed a feature for estimating required volumes of water injection into wells using C#, linear regression, and numerical methods for solving differential equations.3) Conducted code refactoring and test coverage, resulting in a significant reduction in the number of bugs in the production solution.The developed methods have successfully contributed to forming and improving the accuracy of forecasts for geologists. Show less

• 

  Huawei

  Dec 2021 - Apr 2024

  Deep Learning Engineer

  R&D in Speech Processing using Deep Learning Methods: TTS, voice conversion, speaker recognitionKey achievements:1) Developed and trained a bimodal (speech and face) verification MVP solution using PyTorch, reducing the Equal Error Rate (EER) from 1.01% (speech only) to 0.34% (speech+face).2) Collected and preprocessed large datasets using SQL, PySpark, and clustering methods for testing and fine-tuning the release model.3) Leveraged expertise in PyTorch, computer vision, knowledge from paper overviews, and experiments with different training methods/architectures to develop an alternative speaker recognition solution, achieving a significant (1.1--2x) improvement in False Rejection Rate (FRR) compared to the baseline and competitors' solutions.3) Created a zero-shot speech synthesis solution using PyTorch, HuBERT model, and Variational Autoencoder-based Text-to-Speech (TTS) model. Achieved a remarkable 70% relative improvement in speaker similarity and significantly improved robustness to noise in the reference speech.4) Developed a noisy-robust zero-shot speech synthesis method by modifying the VITS architecture. Introduced a self-supervised DINO loss for joint training of a speaker verification model and a unit-based VITS, resulting in substantial enhancements in robustness and speaker similarity. Currently under review at ICASSP 2024 Show less

• 

  Rask AI

  Apr 2024 - now

  Deep Learning Engineer