Enhancing Singing Performances: Novel Method for Automatic Vocal Pitch Correction

Our research presents a novel method for automatically correcting the pitch of human vocals using a data-driven approach. Unlike existing commercial systems, which often shift vocal notes to predefined pitches or the closest pitch in a twelve-tone scale, here pitch is treated as a continuous value. This allows for greater flexibility in singing performances, including improvisation and harmonization. The system is developed and a neural network is trained by using a dataset of 4,702 karaoke from Sumel Inc. which had a great intonation. This model is trained in such a way that it learns to make corrections using the incorrect intonations also for the different pitch variations. The flaws in the previous work was monotonous pitch consideration within a controlled environment. To overcome this our system was proposed, the network architecture also integrates the convolutional layers which is then processed by gated recurrent units, which have shown promising performance in real-world score-free singing pitch correction, also known as auto tuning. The anticipated convolutional neural network accurateness is 93 % and error rate 7%.