We present a continuing data collection effort towards building a rich database of English and Chinese pop songs for efficient speech-to-singing (STS) voice conversion. Parallel recordings of lyrics of songs, sung and read by professional singers, are recorded in a professional studio environment using high quality recording equipments under the supervision of a trained sound engineer. Word-level and sentence-level labeling of the read and sung audio files are performed manually. Then temporal alignment between words in the read lyrics and singing is performed automatically using dynamic time warping (DTW) with carefully crafted features. The accuracy of temporal alignment of frames of speech and singing voices is crucial for STS conversion, as it decides the effectiveness of mapping of parameters from speech signals to those of singing. The temporally aligned frames of speech and singing voices are used to map characteristics for STS conversion. The presented database of parallel recordings of speaking and singing voices of same linguistic content assist in facilitating efficient STS conversion, in addition to providing valuable resorts to singing voice analysis, understanding differences in production-perception of speech and singing voices and, evaluation of singing quality.
@InProceedings{VIJAYAN18.35, author = {Karthika Vijayan and Haizhou Li}, title = {Parallel Speak-Sing Corpus of English and Chinese Songs for Speech-to-Singing Voice Conversion}, booktitle = {Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)}, year = {2018}, month = {may}, date = {7-12}, location = {Miyazaki, Japan}, editor = {Kiyoaki Shirai}, publisher = {European Language Resources Association (ELRA)}, address = {Paris, France}, isbn = {979-10-95546-24-5}, language = {english} }
Summary of the paper