Open to opportunities

Ivan Podluzhny

@ivanpodluzhny

Message

Applied Scientist specializing in speech recognition and machine learning.

United Kingdom

Message

What I'm looking for

I am seeking a role that fosters innovation in AI and speech technologies, with opportunities for professional growth and collaboration.

I am an Applied Scientist at Amazon, focusing on enhancing the naturalness and expressiveness of text-to-speech (TTS) voices. My work involves utilizing Direct Preference Optimization (DPO) to improve TTS synthesis generations. I have successfully designed and deployed an automation measurement pipeline that evaluates speaker identity, accent, and prosody drift, which has significantly contributed to the development of a tone-sensitive neural TTS front-end system for Alexa in Japanese, achieving a remarkable 41% reduction in relative SER.

Prior to my role at Amazon, I worked as a Speech Researcher at STC Group, where I created and released speech recognition models for multiple languages, including English, Spanish, Russian, and Arabic. My innovative approach led to a 38% increase in out-of-vocabulary (OOV) recognition rates using the BPE-dropout technique for low-resource tasks. I also developed a punctuation model for automatic speech recognition (ASR) results, achieving a 0.9 F1 score, and contributed to various academic papers and conferences in the field.

Experience

Work history, roles, and key accomplishments

Current

Applied Scientist

Current

Amazon

Jan 2023 - Present (2 years 5 months)

Improved naturalness and expressiveness for new TTS voices using Direct Preference Optimization (DPO). Designed and deployed an automation measurement pipeline for speaker identity, accent, and prosody drift in LLM TTS synthesis generations. This work included developing a tone-sensitive neural TTS front-end system for Alexa in Japanese, which reduced relative SER by 41% and achieved strong prefer

Direct Preference Optimization (DPO)LLM TTS Neural Networks Automation Speaker Identity Alexa PyTorch

Speech Researcher

STC Group

Jan 2019 - Present (6 years 5 months)

Created and released speech recognition models for various languages, including English, Spanish, Russian, and Arabic. Increased the Out-Of-Vocabulary (OOV) recognition rate by 38% using the BPE-dropout technique for low-resource tasks. Additionally, developed and published an NLP post-processing punctuation model for ASR results, achieving a 0.9 F1 score.

Speech Recognition NLP BPE Dropout Machine Learning Python Low Resource Languages F1 Score

Education

Degrees, certifications, and relevant coursework

ITMO University

Ph.D. in CS, Computer Science

Engaged in research on semi-supervised techniques for end-to-end Speech Recognition. Developed expertise in cutting-edge machine learning and AI methodologies.

Saint Petersburg State University

M.S. in Mathematics, Mathematics

Grade: 4.96 GPA

Focused on advanced topics in Probability Theory, Random Processes, and Statistics, achieving a high academic standing. Gained a comprehensive understanding of mathematical principles applicable to various scientific fields.