Xinyi Zhao

Master's

Data Science

Pingan Technology

Machine Learning Engineer Intern

• Fine-tuned VITS, a Text-to-Speech deep learning model, on a 10-hour Mandarin-English multi-speaker dataset, improving 0.2 MOS (Mean Opinion Score) values through the manipulation of speaker embeddings, pitch, and tone. • Developed automated data quality filtering and cleaning Python scripts using librosa and numpy, enabling bulk examination of inconsistent pronunciations, noise, or fractured sounds, and trimming silences. • Replicated and experimented with over 6 existing Text-to-Speech models, including Fastspeech2 and YourTTS, on cloud servers, comparing performance across more than 4 metrics. • Implemented Fernet encryption to ensure the security and confidentiality of Python codes and models. Deployed the fine-tuned models to a production environment and collaborated on updates via Git.