letoeorsi

Worked closely with a fast-moving startup team to rapidly prototype and deploy multilingual speech systems in production, reducing speaker confusion rate by 30% versus open-source baselines such as PyAnnote.

Advanced TTS quality for 8 new Alexa voices using Direct Preference Optimization (DPO), improving conversational naturalness and expressiveness. Reduced word error rate (WER) by 50% and improved voice stability in LLM-based dubbing models, driving measurable gains in user experience. Designed and deployed a scalable ML pipeline for automatic tracking of speaker identity, accent, and prosody drift, enabling fast quality monitoring. Built a tone-sensitive neural TTS front-end for Japanese Alexa, reducing relative SER by 41% and earning strong listener preference in MUSHRA evaluations. Trained large-scale speech and multimodal models using distributed GPU infrastructure (32+ GPUs) on AWS.

Created and released speech recognition models for different languages (English, Spanish, Russian, Arabic). Increased OOV recognition rate by 38% using BPE-dropout augmentation for low-resource tasks, improving model robustness. Built an NLP-based punctuation restoration model that reaches 0.95 F1, integrated into the production ASR pipeline.