Pitch Detection Algorithm Evaluation Framework

When implementing an existing or designing a brand new pitch detection algorithm (PDA) there is a need to evaluate it and maybe compare it to another one. We could make some comparisons or rough estimates based on graphical output, but this approach is ineffective and definitely inconvenient for larger file-bases. In this paper design and implementation of PDA speech-oriented evaluation framework is presented. It is based on existing ECESS PMA/PDA manually pitch marked reference database. The database consists of four-channel recordings in various environments with varying SNR and thus is suitable for testing robustness of pitch detection algorithms. Standard evaluation criteria widely used by many authors in the field are discussed and few improvements are suggested in this area. Finally results of comparing several PDA algorithms are presented and further idea of using the best one in a real-time punctuation detector is presented too.