The unique work that makes this speech recognition superior to other tools is in those 284 lines of code: https://github.com/DeutscheKI/tevr-asr-tool/blob/maste

The unique work that makes this speech recognition superior to other tools is in those 284 lines of code: https://github.com/DeutscheKI/tevr-asr-tool/blob/master/tevr_asr_tool.cc#L57-L160 https://github.com/DeutscheKI/tevr-asr-tool/blob/master/tevr... That's a custom-designed beam search decoder implemented in C++ and based on the research for my TEVR paper. It increases performance by a relative 16% reduction in word error rate.