Vocal Effort Modification for Singing Synthesis

Vocal effort modification of natural speech is an asset to various applications, in particular, for adding flexibility to concatenative voice synthesis systems. Although decreasing vocal effort is not particularly difficult, increasing vocal effort is a challenging issue. It requires the generation of artificial harmonics in the voice spectrum, along with transformation of the spectral envelope. After a raw source-filter decomposition, harmonic enrichment is achieved by 1/ increasing the source signal impulsiveness using time distortion, 2/ mixing the distorted and natural signals’ spectra. Two types of spectral envelope transformations are used: spectral morphing and spectral modeling. Spectral morphing is the transplantation of natural spectral envelopes. Spectral modeling focuses on spectral tilt, formant amplitudes and first formant position modifications. The effectiveness of source enrichment, spectrum morphing, and spectrum modeling for vocal effort modification of sung vowels was evaluated with the help of a perceptive experiment. Results showed a significant positive influence of harmonic enrichment on vocal effort perception with both spectral envelope transformations. Spectral envelope morphing and harmonic enrichment applied on soft voices were perceptively close to natural loud voices. Automatic spectral envelope modeling did not match the results of spectral envelope morphing, but it significantly increased the perception of vocal effort.