Abstract

This talk will cover three different directions of my research journey in the speech domain: recognition, generation and privacy. We will start with self-training algorithms that have recently emerged as a powerful strategy for semi-supervised learning in large-scale speech recognition. We will discuss different aspects of self-training to make it simple and resource-efficient. Then we will delve into the language modeling for speech recognition where I show that at the proper scale error correction models outperform conventional language models. Next point of discussion will be the joint modeling of text-to-speech and speech-to-text with the help of a simple tokenization scheme. Finally I discuss how speech recognition systems can be trained with differential privacy and federated learning.

Biography

Tatiana is a research scientist at the machine learning research team, Apple. Prior to Apple, she was a postdoctoral research scientist in the speech recognition team, Facebook AI Research. Back in the day, Tatiana received a Ph.D. in mixed type partial differential equations from Moscow State University. For 4 years she worked on applications of machine learning to high energy physics as a researcher in the joint lab at Yandex and CERN. The main focus of her recent research is stability of transformers training, speech recognition (robustness and data-efficiency) and generation, private federated learning.