A fixed dimension and perceptually based dynamic sinusoidal model of speech

Abstract

This paper presents a fixed-and low-dimensional, perceptually based dynamic sinusoidal model of speech referred to as PDM (Perceptual Dynamic Model). To decrease and fix the number of sinusoidal components typically used in the standard sinusoidal model, we propose to use only one dynamic sinusoidal component per critical band. For each band, the sinusoid with the maximum spectral amplitude is selected and associated with the centre frequency of that critical band. The model is expanded at low frequencies by incorporating sinusoids at the boundaries of the corresponding bands while at the higher frequencies a modulated noise component is used. A listening test is conducted to compare speech reconstructed with PDM and state-of-the-art models of speech, where all models are constrained to use an equal number of parameters. The results show that PDM is clearly preferred in terms of quality over the other systems.

Fusion of multiple parameterisations for DNN-based sinusoidal speech synthesis with multi-task learning

“…For the DIR method, we model log|A| and log|B| explicitly. For this, we proposed in [11] synthesis, HDM is used for generating speech, where amplitudes at each harmonic (|A HDM | , |B HDM |) are assigned the amplitude of the centre frequency of the critical band in which they lie. Figure 1 gives an overview of both methods for integrating the DSM into DNN-based speech synthesis (see [12] for more detail).…”

An investigation of the application of dynamic sinusoidal models to statistical parametric speech synthesis

“…number of sinusoids) is higher than typical source-filter ones and varies from frame to frame. To address this problem, a perceptual dynamic sinusoidal model (PDM) [17] has been proposed to generate high quality speech with a fixed and low number of parameters.…”

“…In addition, [17] has shown that incorporating the dynamic slope of sinusoids can greatly improve quality in copy synthesis. It is natural, therefore, to consider including this dynamic feature for statistical modelling too.…”

Initial investigation of speech synthesis based on complex-valued neural networks

“…For spectral features, either i) 50 regularized discrete cepstra (RDC) extracted from the amplitudes of the harmonic dynamic model (HDM) [24] or ii) 50 highly correlated log amplitudes from the perceptual dynamic sinusoidal model (PDM) [25] are used as real-valued spectral output. 50 complex amplitudes with minimum phase extracted from PDM [19] are applied as complex-valued spectral output. Continuous logF 0 and a voiced/unvoiced (vuv) binary value together with either type of these spectral features are used to represent output features (total dimensions: 52).…”

“…This is motivated by the fact that for real-valued classification tasks, a CVNN has the same performance as a real-valued NN with a larger number of neurons [18]. Note that speech synthesis is a regression task, which is different from tasks reported in the literature; iii) Complex amplitudes extracted from [19] can be used as complex-valued outputs where phase is composed of linear phase, minimum phase and disperse phase. Here, linear phase should be omitted in the calculation of the amplitudephase objective function since analysis window position is unrelated to linguistic input.…”