Mustafa Ebrar Aktaş

• Optimized Conv1D and quantization in "ctranslate2". • Created several internal web applications with Streamlit, including an LLM prompt revision manager, tools for content creation with LLMs and AI image generators, and observation tools connected to databases. • Optimized and integrated Whisper ASR into the existing pronunciation feedback library for providing robust pronunciation feedback to accented speakers. • Implemented a fast similarity based word group aligner in Rust with Python interface using PyO3.

• Developed and trained advanced acoustic models, including TDNN and Conformer/ContextNet, in TensorFlow to enhance streaming speech recognition for accented speech on mobile devices. • Developed a Rust library for providing real-time feedback on pronunciation, fluency, and intonation with very low latency on mobile devices. • Implemented a real-time, on-device streaming hybrid ASR library using lookahead FST composition with TensorFlow models in Rust, ensuring robust and efficient performance. • Developed a language-agnostic, similarity-based word group aligner using transformers. • Created and continue to maintain "tflitec", a cross-platform TensorFlow Lite Rust wrapper library.

• Trained several TDNN and Chain ASR models in Kaldi to enhance accented speech recognition, optimized for server implementation. • Developed and deployed a pronunciation feedback gRPC service. • Designed and deployed mobile-compatible DNN acoustic models using TensorFlow.

• Intern at KİTAL(Wireless Communication Technologies Laboratory ) • wrote a Cepstrum Analysis program on MATLAB • conducted research and developed a MATLAB program for Direction of Arrival Estimation.

Exchange Student at UIUC for Spring 2018