Project: Comprehensive ASR training program in Kazakh and Russian languages
Tasks:
- Created models competitive with other companies' solutions.
- Created and implemented the entire TTS model training pipeline, from data acquisition to production. Deployment, resulting in highly controlled synthesis with 4+ MOS for each speaker.
- Deployed an inference service to accelerate the TTS model, achieving high throughput (300+ RTFX) and low latency between audio fragments for more simultaneous streams.
- Optimized the training process by adding various methods, resulting in a much faster process. This reduced the waiting time for model training and improved overall efficiency.
- Deployed an inference service to optimize ASR models for Kazakh and Russian languages, achieving a high 450+ RTFX throughput and low latency for over 500 threads.
- Implemented MLOps techniques including data version control (DVC) and weighting and bias (W&B). Tracked experiments, which optimized the machine learning development cycle, improved the reproducibility of the experiments, and increased collaboration within the team.
- Developed and deployed an efficient API service using REST and gRPC to leverage proprietary ASR and TTS solutions.

Result: Achieved exceptional performance with low latency and high throughput.
Built a versatile pipeline for processing and preparing audio data to train ASR models. Accelerated performance
with low-resource languages from 1 week to as little as 1 day

Duration - 3 year

Project: Led the development and implementation of an advanced speech recognition service serving over 100 languages in both offline and streaming modes
Role: Lead MLE
Tasks:
- Led the successful implementation of a speech synthesis platform in production, with 6 Kazakh speakers and over 240 Russian speakers, achieving exceptional quality metrics.
- Implemented end-to-end development pipeline using ClearML, optimizing resource allocation, data version control, recording experiments and documenting hypotheses and solutions, and managing a cross-functional team of 2 ML.
Engineers, an MLOps expert, and a backend developer.
- Implemented a single language identification service integrated into the ASR pipeline, improving the overall system.

Opportunity Outcome: Implemented an advanced speech recognition service, improving the overall system and its performance

Duration - 1 year