Nobody should be surprised by this. To actually test your voice recognition software in a definitive manner you need to compare you kind of have to compare you

Nobody should be surprised by this. To actually test your voice recognition software in a definitive manner you need to compare you kind of have to compare your software's analysis to human analysis and that requires humans to listen to the audio or at least snippets of it. The software is competing with a human after all, you need to test its performance compared to the humans.