Roman Kail

ML Engineer, Text to Speech team

Yandex

Yandex is a major Russian technology company, known for search, maps, cloud services, and consumer products used across Russia and other markets.

Improved TTS quality through training and data pipelines, and cut diffusion decoder inference cost 4× without quality loss.

• Implemented CTC loss for speech tokenizer training, improving intonation by 17% in side-by-side (SBS) evaluation • Cut inference cost of the diffusion decoder 4× with no quality loss: replaced diffusion with flow matching, halved the model size and distilled it • Built a synthetic emotional speech dataset, improving emotion quality by 11% in SBS • Designed an acceptance pipeline for studio-recorded dialogue data, improving naturalness by 5% in SBS

ML Engineer, Text to Speech team

SberDevices

SberDevices develops smart devices and voice technologies within Sber, including products built around its Salute virtual assistant.

Shipped emotional TTS and a C++ lexical stress module to production, improving speech for rare words and names.

• Introduced emotional speech synthesis into a neutral-only TTS model, shipping it to production • Developed a lexical stress prediction model and deployed it as a C++ inference module, enabling correct pronunciation of street names, surnames and other rare words in production

ML Lecturer

Yandex School of Data Analysis

Yandex School of Data Analysis is a selective graduate-level program run by Yandex, offering advanced training in machine learning and data science.

Created and teaches the speech synthesis course section for about 100 graduate students each year, including its lectures and assessment.

• Author and teach the speech synthesis part of the “Speech Technologies” course, ∼100 graduate students per year • Built lectures, assignments and a grading system from scratch; course materials are open source: course repository on GitHub

ML Engineer

Skoltech

Skoltech is an international graduate research university in Moscow, focused on science, technology, engineering, and innovation.

Researched uncertainty and ensembling, publishing three peer-reviewed papers with about 55 citations and authoring ScaleFace.

• 3 peer-reviewed papers on uncertainty and ensembling, two at CORE A venues (AISTATS 2022, IEEE DSAA 2023), ∼55 citations • Authored ScaleFace (DSAA 2023) — input-dependent uncertainty for face recognition at near-zero cost

ML program, Machine Learning

Yandex School of Data Analysis

MSc, Machine Learning

Higher School of Economics

BSc, Applied Mathematics and Physics

Moscow Institute of Physics and Technology