Data-driven voice soruce waveform modelling

In this work a glottal model loosely based on the Ishizaka and Flanagan model is proposed, where the number of parameters is drastically reduced. First, the glottal excitation waveform is estimated, together with the vocal tract filter parameters, using inverse filtering techniques. Then the estimated waveform is used in order to identify the nonlinear glottal model, represented by a closedloop configuration of two blocks: a second order resonant filter, tuned with respect to the signal pitch, and a regressor-based functional, whose coefficients are estimated via nonlinear identification techniques. The results show that an accurate identification of real data can be achieved with less than ½¼ regressors of the nonlinear functional, and that an intuitive control of fundamental features, such as pitch and intensity, is allowed by acting on the physically informed parameters of the model.

Jeevakumar Kanagaratnam Joint Estimation of Vocal Tract and Source Parameters of a Speech Production Model Phd Thesis Dublin City University, 1993

I hereby certify that this material, which I now submit for assessment on the programme o f study leading to the award of Ph.D. is entirely my own work and has not been taken from the work of others save and to the extent that such work has been cited and acknowledged within the text o f my work.

2007

Traditionally, voice production has been modeled using the source-filter approach. Now, the glottal airflow can be interpreted as the source and the oral and nasal cavities can be modeled as the filter. The filters are based either on representing the formants, i.e., the resonances of the vocal tract, or on modeling the articulation, i.e., the shape of the vocal tract. The source signal depends on the voice to be produced, and it is traditionally modeled as a sequence of geometric pulses or as random noise. However, more accurate source signals can be obtained by utilizing the inverse filtering technique or by modeling the vocal folds as oscillating masses.

There is a need for automatic methods for parametrization of the voice source signals. Representatives of the two types of methods that have been used most often for parametrization were tested and compared. For this purpose a novel evaluation procedure is proposed which makes it possible to perform the numerous tests needed for a detailed comparison of the methods. This evaluation procedure revealed that in order to reduce the average error in the estimated voice source parameters the estimation methods should be able to estimate noninteger values of these parameters. The proposed evaluation method was also used to study the influence of low-pass filtering on the estimated voice source parameters. The factor low-pass filtering was chosen because low-pass filtering is probably used in all methods in which voice source parameters are estimated. It turned out that low-pass filtering causes an error in all estimated voice source parameters. On average, the smallest errors were found for a parametrization method in which a voice source model is fitted to the voice source signals, and in which the voice source model is low-pass filtered with the same filter as the voice source signals.

Two methods are presented for automatic calculation of the voice source parameters from continuous speech. Both methods are used to calculate the voice source parameters for natural speech. However, for natural speech no objective test procedure seems available. Therefore, both methods were also tested on synthetic speech.

… International Conference on Speech Science and …, 2006

We describe a new method to estimate vocal cord dynamics using a parametric model of glottis movements in order to assess the health of the vocal cords and detect pathological conditions of the larynx. The underlying model is based on that proposed by the Fujisaki-Ljungqvist, modified by reducing the number of parameters. The choice of a parametric model enabled us to assess the impact of the various model characteristics as well as the number of parameters. In the first step of the process we estimated the transfer function of the vocal tract and used it to approximate the model-generated speech. The general approach to the vocal tract estimation was based on the hypothesis that the vocal tract filtering action and the glottal forcing functions affect different, non-overlapping, frequency bands, and can therefore, be separated by homomorphic filtering. In the second step we used filtering of the acoustic waveform, using the estimated vocal tract transfer function H(ω) to estimate the shape of the glottal pulses. The actual estimates of H(ω) were obtained by applying an inverse filtering approach, using the cepstral method in conjunction with liftering (filtering) in the cepstral domain. The model parameters were estimated by maximizing the match between the model-generated and the observed speech signal. The general approach consists of finding the parameters of the FL model that would maximize the correspondence between the observed and synthetic utterances filtered by H(ω). The optimization was performed using the Nelder-Mead simplex search method because of the strong nonlinearities, discontinuities and the complex interactions among the model parameters. Prior to estimating the transfer function, the observed speech was filtered to remove the effects of lip radiation. In order to evaluate this approach we used the Kay Elemetrics Disordered Voice database, which comprises over 1,400 voice samples of approximately 700 subjects and includes sustained phonation and running speech samples from patients with a wide variety of organic, neurological, traumatic, and psychogenic voice disorders, as well as from 53 normal speakers. Estimated parameters of the FL model and its combinations were analyzed to determine the potential of this method to assess the health state of the speaker. We will illustrate the applicability of this technique to the problem of discriminating between healthy and pathological speech samples. The classification is based on the resulting estimates of the model parameters. The results suggest that the parameter estimates may provide a useful clinical tool for rapid unobtrusive, triage and diagnosis.

2001

A physically-informed glottal model is proposed; some physical information is retained in a linear block that accounts for fold mechanics, while non-linear coupling with the airflow is modeled using a regressorbased mapping. The model is used in an identification/resynthesis scheme. Given a real signal, system parameters are estimated via non-linear identification techniques; then the model is used for resynthesizing the signal. With a proper choice of the regressor set the system accurately fits the target waveform and is stable during resynthesis. Physical parameters can be used to change voice quality and speaker identity.

2010 IEEE International Conference on Acoustics, Speech and Signal Processing, 2010

There are numerous models of varying complexities which seek to efficiently represent the voice source signal. These models are typically based on data and observations which can come from air-flow masks, electroglottographs, mechanical systems, and the inversefiltering of speech signals. The first part of this study examines observations from the high-speed imaging of the larynx and proposes a new source model, which is shown to provide a better fit for the observed data than existing models. The proposed source model is then used in an automatic source estimation application, based on methods introduced in an earlier study . Results, on average, show that the proposed model provides a more accurate estimation of the source signal compared with the Liljencrants-Fant model.

Interspeech 2021, 2021

Source-filter modelling is among the fundamental techniques in speech processing with a wide range of applications. In acoustic modelling, features such as MFCC and PLP which parametrise the filter component are widely employed. In this paper, we investigate the efficacy of building acoustic models from the raw filter and source components. The raw magnitude spectrum, as the primary information stream, is decomposed into the excitation and vocal tract information streams via cepstral liftering. Then, acoustic models are built via multi-head CNNs which, among others, allow for processing each individual stream via a sequence of bespoke transforms and fusing them at an optimal level of abstraction. We discuss the possible advantages of such information factorisation and recombination, investigate the dynamics of these models and explore the optimal fusion level. Furthermore, we illustrate the CNN's learned filters and provide some interpretation for the captured patterns. The proposed approach with optimal fusion scheme results in up to 14% and 7% relative WER reduction in WSJ and Aurora-4 tasks.

Estimation of the glottal source has applications in many areas of speech processing. Therefore, a noise-robust automatic source estimation algorithm is proposed in this paper. The source signal is estimated using a codebook search approach. The glottal area waveforms extracted from high-speed recordings of the glottis is converted to the glottal flow signals in order to evaluate the performance of the proposed source estimation algorithm. Results in clean and noisy conditions, on average, show that the proposed algorithm provides more accurate estimation than the software toolkit Aparat [1] as well as an earlier approach [2].