Output details

'Fundamental Frequency Modulation in Singing Voice Synthesis'

This article presents the culmination of a series of studies, which have resulted in the development of a new model for the analysis and synthesis of the singing voice. The model offers an improvement in the perceived naturalness of synthesised singing by taking account of vocal drift, a form of unintentional low frequency modulation in the fundamental frequency.

In the study, fundamental frequency (f0) measurements are taken from vocalists producing a selected range of utterances without vibrato and trends in the data are observed. A probabilistic function is derived from these observations, with the aim of providing natural-sounding low frequency f0 modulation to synthesised singing. The perceptual relevance of this function is evaluated with subjective listening tests.

As an expert in audio analysis and author of the widely cited LibXtract feature extraction library (https://github.com/jamiebullock/libxtract), Bullock's role was to oversee the analysis component of the study. This included: specifying the design of tests (vocal techniques, recording techniques and equipment) to ensure an appropriate dataset was gathered; implementing the f0 tracking algorithm, audio segmentation and audio feature selection used in the measurement of f0 drift; deriving a novel probabilistic model from these measurements; specifying design of a method for evaluation of the proposed model against existing synthesis techniques.

Supporting research:

Stables, R., Athwal, C. & Bullock, J., 2011: ‘Towards a model for the humanisation of pitch drift in singing voice synthesis’, Proceedings of the 2011 International Computer Music Conference; ICMA, UK.

Stables, R., Athwal, C. & Bullock, J., 2011: ‘The Humanisation of Stochastic Processes for the Modelling of F0 in Singing’, Joint International Symposium: Frontiers of Research on Speech and Music & Computer Music Modeling and Retrieval; Bhubaneswar, India.

Bullock, J., 2007: ‘libxtract: a lightweight library for audio feature extraction’, Proceedings of the 2007 International Computer Music Conference; ICMA, Sweden.