George Roussos de Kock

My role at Vidispine was in a SaaS setting and my responsibilities regarded front-end objectives, but were occasionally backend-focused (backend built on .NET supported by a PostgreSQL database). I was involved in the development of a feature-rich, web-based (React/TypeScript) media ingestion tool used by broadcasting companies to schedule, record and transcode content. Additionally, I was responsible for building reusable React components for a commercial component library (using Google's MUI as a base).

As a software developer at Storytel, I developed an end-to-end neural network pipeline, which was responsible for synthesising speech from text, for generation of long-form audiobook content. I was responsible for and led the entire process, working independently (within a Data Science team) for the best part of my tenure at the company. My responsibilities included:

• Delivering a fully functional MVP for stakeholders (the venture was new to the company) that could synthesise speech in English and Swedish

• Developing a custom pipeline (for the Swedish language), in order to train style-controllable models and produce content, in multiple voices

• Designing methods for generation of long-form audiobook content

• Reviewing and implementing methodologies to maintain technological advancement

• Training and evaluation of the network and its models, by conducting various tests and collecting mean opinion scores

• Analysing, understanding and engineering the data the network used for the training sessions

• Leading the recording sessions, directing voice actors when recording content to use in the data for the pipeline

The pipeline I developed was eventually successfully used to train the company’s first AI voice and produce the company's first fully neural network-based audiobook content.

During my time at ReadSpeaker I gradually moved from speech synthesis related objectives to software development areas. It was a broad role in which I was tasked with:

• Supervision of QA documents in different languages

• Developing Python and C++ pipelines for text normalisation and disambiguation

• Development of voice engine features used by third-party products

• Training of various models (both supervised and unsupervised), such as parsers and taggers (spaCy pipelines) and grapheme- to- phoneme automata

Completion of Master's Education

Thesis Project: Transfer Learning in Speech Synthesis - Exploring Pretrained Weights Adaptation and Usage of Speaker Embeddings in Neural End-to-End Speech Synthesis

The project explored the feasibility of predict speech from text in pre-trained neural networks, where:

• available data was 80% smaller than required and unprocessed, so that pre-training the network on other data was an option

• obtaining data was not feasible, so zero-shot prediction, based on speaker embeddings, using clustering and multi-voice recordings, might be an option

Results showed that both options are viable for when training data is not available, but that the network is not able to pick on all of the new voice's vocal charactertics (such as prosody). The training pipeline was a mix of pyTorch and TensorFlow and training was done on a GCP instance.

Concentration in Natural Language Processing (NLP)

Concentration in Classics and Literature