Worked on enhancing the Azure-Samples/cognitive-services-speech-sdk repository by delivering a new Custom TN Training capability focused on improving model customization for domain-specific vocabularies. The approach involved expanding multi-letter unit (mN) support and refining single-letter unit handling within the custom TN training workflow, which increased training data precision and improved end-user model performance. Emphasis was placed on code quality and alignment with project roadmap goals rather than bug fixes. The work was implemented in C++ using the Speech SDK, demonstrating a targeted application of C++ development skills to address data noise reduction and more accurate model behavior for customers.