◦ Enhanced acoustic model precision and accelerated vocoder performance, elevating sound quality and efficiency. ◦ Spearheaded innovative research initiatives to mitigate speaker accents in diverse languages without the need for audio recordings. ◦ Executed a successful migration of the speech synthesis system to a modern architecture, ensuring scalability and performance enhancements.

◦ Spearheaded the development and rollout of a cutting-edge speech recognition service, catering to over 100 languages in both offline and streaming modes. ◦ Led the successful deployment of a speech synthesis platform into production, featuring 6 Kazakh speakers and over 240 Russian speakers, achieving exceptional quality metrics. ◦ Implemented an end-to-end development pipeline using ClearML, optimizing resource allocation, versioning data, logging experiments, and documenting hypotheses and solutions, while managing a cross-functional team of 2 ML engineers, MLOps expert, and Backend developer. ◦ Implemented a seamless language identification service integrated into the ASR pipeline, enhancing overall system capabilities.

◦ Developed and optimized an end-to-end ASR training pipeline for Kazakh and Russian languages, producing models that are competitive with other companies’ solutions. ◦ Built and implemented the entire TTS model training pipeline, from data collection through to production deployment, resulting in highly controlled synthesis with 4+ MOS for each speaker. ◦ Deployed an inference service to accelerate the TTS model, achieving a high bandwidth of 300+ RTFX and low latency between audio chunks for more concurrent streams. ◦ Optimized the training pipeline by adding various techniques, resulting in a significant speed-up of the process. This reduced the waiting time for model training and improved overall efficiency. ◦ Deployed an inference service to optimize the ASR models for Kazakh and Russian languages, achieving a high bandwidth of 450+ RTFX and an low latency for 500+ streams. ◦ Implemented MLOps practices including Data Version Control (DVC) and Weights & Biases (W&B) experiment tracking, which optimized the machine learning development cycle, improved reproducibility of experiments, and increased collaboration within the team. ◦ Designed and deployed an efficient API service using REST and gRPC to leverage proprietary ASR and TTS solutions. Achieved exceptional low-latency and high-throughput performance. ◦ Constructed a universal pipeline for audio data processing and preparation to train ASR models. Accelerated work with low-resource languages from 1 week to just 1 day.

◦ Built program for automated payment transfers via Kaspi app, processing 2000+ daily transactions for Kazakh speech data platform. Streamlined payments, drove growth, and replaced 10+ manual positions for platform success. ◦ Designed and developed a sophisticated Kazakh speech data collection platform, attracting over 44,000 users. Achieved successful collection of 10,000+ hours of speech in 3 months, on a limited budget.