Anna Přibilová

Non-linear frequency scale mapping for voice conversion in text-to-speech system with cepstral description

Voice conversion, i.e. modification of a speech signal to sound as if spoken by a different speaker, finds its use in speech synthesis with a new voice without necessity of a new database. This paper introduces two new simple non-linear methods of frequency scale mapping for transformation of voice characteristics between male and female or childish. The frequency scale mapping methods were developed primarily for use in the Czech and Slovak text-to-speech (TTS) system designed for the blind and based on the Pocket PC device platform. It uses cepstral description of the diphone speech inventory of the male speaker using the source-filter speech model or the harmonic speech model. Three new diphone speech inventories corresponding to female, childish and young male voices are created from the original male speech inventory. Listening tests are used for evaluation of voice transformation and quality of synthetic speech.

Evaluation of influence of spectral and prosodic features on GMM classification of Czech and Slovak emotional speech

This article analyzes and compares influence of different types of spectral and prosodic features for Czech and Slovak emotional speech classification based on Gaussian mixture models (GMM). Influence of initial setting of parameters (number of mixture components and used number of iterations) for GMM training process was analyzed, too. Subsequently, analysis was performed to find how correctness of emotion classification depends on the number and the order of the parameters in the input feature vector and on the computation complexity. Another test was carried out to verify the functionality of the proposed two-level architecture comprising the gender recognizer and of the emotional speech classifier. Next tests were realized to find dependence of some negative aspect (processing of the input speech signal with too short time duration, the gender of a speaker incorrectly determined, etc.) on the stability of the results generated during the GMM classification process. Evaluations and tests were realized with the speech material in the form of sentences of male and female speakers expressing four emotional states (joy, sadness, anger, and a neutral state) in Czech and Slovak languages. In addition, a comparative experiment using the speech data corpus in other language (German) was performed. The mean classification error rate of the whole classifier structure achieves about 21% for all four emotions and both genders, and the best obtained error rate was 3.5% for the sadness style of the female gender. These values are acceptable in this first stage of development of the GMM classifier. On the other hand, the test showed the principal importance of correct classification of the speaker gender in the first level, which has heavy influence on the resulting recognition score of the emotion classification. This GMM classifier should be used for evaluation of the synthetic speech quality after applied voice conversion and emotional speech style transformation.

Spectrum Modification for Emotional Speech Synthesis

Emotional state of a speaker is accompanied by physiological changes affecting respiration, phonation, and articulation. These changes are manifested mainly in prosodic patterns of F0, energy, and duration, but also in segmental parameters of speech spectrum. Therefore, our new emotional speech synthesis method is supplemented with spectrum modification. It comprises non-linear frequency scale transformation of speech spectral envelope, filtering for emphasizing low or high frequency range, and controlling of spectral noise by spectral flatness measure according to knowledge of psychological and phonetic research. The proposed spectral modification is combined with linear modification of F0 mean, F0 range, energy, and duration. Speech resynthesis with applied modification that should represent joy, anger and sadness is evaluated by a listening test.

Emotional style conversion in the TTS system with cepstral description

This contribution describes experiments with emotional style conversion performed on the utterances produced by the Czech and Slovak textto-speech (TTS) system with cepstral description and basic prosody generated by rules. Emotional style conversion was realized as post-processing of the TTS output speech signal, and as a real-time implementation into the system. Emotional style prototypes representing three emotional states (sad, angry, and joyous) were obtained from the sentences with the same information content. The problem with the different frame length between the prototype and the target utterance was solved by linear time scale mapping (LTSM). The results were evaluated by a listening test of the resynthetized utterances.

Harmonic model for female voice emotional synthesis

Spectral and prosodic modifications for emotional speech synthesis using harmonic modelling are described. Autoregressive parameterization of inverse Fourier transformed log spectral envelope is used. Spectral flatness determines the voicing transition frequency dividing spectrum of synthesized speech into minimum phases and random phases of the harmonic model. Female emotional voice conversion is evaluated by a listening test.

Experiment with GMM-Based Artefact Localization in Czech Synthetic Speech

The paper describes an experiment with using the statistical approach based on the Gaussian mixture models GMM for localization of artefacts in the synthetic speech produced by the Czech text-to-speech system employing the unit selection principle. In addition, the paper analyzes influence of different number of used GMM mixtures, and the influence of setting of the frame shift during the spectral feature analysis on the resulting artefact position accuracy. Obtained results of performed experiments confirm proper function of the chosen concept and the presented artefact position localizer can be used as an alternative to the standardly applied manual localization method.

EVALUATION OF SPECTRAL AND PROSODIC FEATURES OF SPEECH AFFECTED BY ORTHODONTIC APPLIANCES USING THE GMM CLASSIFIER

Abstract The paper describes our experiment with using the Gaussian mixture models (GMM) for classification of speech uttered by a person wearing orthodontic appliances. For the GMM classification, the input feature vectors comprise the basic and the complementary spectral properties as well as the supra-segmental parameters. Dependence of classification correctness on the number of the parameters in the input feature vector and on the computation complexity is also evaluated. In addition, an influence of the initial setting of the parameters for GMM training process was analyzed. Obtained recognition results are compared visually in the form of graphs as well as numerically in the form of tables and confusion matrices for tested sentences uttered using three configurations of orthodontic appliances.

An experiment with spectral analysis of emotional speech affected by orthodontic appliances

The contribution describes the effect of the fixed and removable orthodontic appliances on spectral properties of emotional speech. Spectral changes were analyzed and evaluated by spectrograms and mean Welch’s periodograms. This alternative approach to the standard listening test enables to obtain objective comparison based on statistical analysis by ANOVA and hypothesis tests. Obtained results of analysis performed on short sentences of a female speaker in four emotional states (joyous, sad, angry, and neutral) show that, first of all, the removable orthodontic appliance affects the spectrograms of produced speech.

Statistical analysis of complementary spectral features of emotional speech in Czech and Slovak

Several spectral features quantify speaker-dependent as well as emotion-dependent characteristics of a speech signal. It means these features provide information which complements the vocal tract characteristics. This paper analyzes and compares complementary spectral features distribution (spectral centroid, spectral flatness measure, Shannon entropy) of male and female acted emotional speech in Czech and Slovak languages.

GMM-Based Evaluation of Emotional Style Transformation in Czech and Slovak

Abstract In the development of the voice conversion and the emotional speech style transformation in the text-to-speech systems, it is very important to obtain feedback information about the users’ opinion on the resulting synthetic speech quality. For this reason, the evaluations of the quality of the produced synthetic speech must often be performed for comparison. The main aim of the experiments described in this paper was to find out whether the classifier based on Gaussian mixture models (GMMs) could be applied for evaluation of male and female resynthesized speech that had been transformed from neutral to four emotional states (joy, surprise, sadness, and anger) spoken in Czech and Slovak languages. We suppose that it is possible to combine this GMM-based statistical evaluation with the classical one in the form of listening tests or it can replace them. For verification of our working hypothesis, a simple GMM emotional speech classifier with a one-level structure was realized. The next task of the performed experiment was to investigate the influence of different types and values (mean, median, standard deviation, relative maximum, etc.) of the used speech features (spectral and/or supra-segmental) on the GMM classification accuracy. The obtained GMM evaluation scores are compared with the results of the conventional listening tests based on the mean opinion scores. In addition, correctness of the GMM classification is analyzed with respect to the influence of the setting of the parameters during the GMM training—the number of mixture components and the types of speech features. The paper also describes the comparison experiment with the reference speech corpus taken from the Berlin database of emotional speech in German language as the benchmark for the evaluation of the performance of our one-level GMM classifier. The obtained results confirm practical usability of the developed GMM classifier, so we will continue in this research with the aim to increase the classification accuracy and compare it with other approaches like the support vector machines.