Voice source characterization for prosodic and spectral manipulation
- Pérez Mayos, Javier
- Antonio Bonafonte Cávez Director
Defence university: Universitat Politècnica de Catalunya (UPC)
Fecha de defensa: 03 July 2012
- Climent Nadeu Camprubí Chair
- Daniel Erro Eslava Secretary
- Eduardo Rodríguez Banga Committee member
Type: Thesis
Abstract
The objective of this dissertation is to study and develop techniques to decompose the speech signal into its two main components: voice source and vocal tract. Our main efforts are on the glottal pulse analysis and characterization. We want to explore the utility of this model in different areas of speech processing: speech synthesis, voice conversion or emotion detection among others. Thus, we will study different techniques for prosodic and spectral manipulation. One of our requirements is that the methods should be robust enough to work with the large databases typical of speech synthesis. We use a speech production model in which the glottal flow produced by the vibrating vocal folds goes through the vocal (and nasal) tract cavities and its radiated by the lips. Removing the effect of the vocal tract from the speech signal to obtain the glottal pulse is known as inverse filtering. We use a parametric model fo the glottal pulse directly in the source-filter decomposition phase. In order to validate the accuracy of the parametrization algorithm, we designed a synthetic corpus using LF glottal parameters reported in the literature, complemented with our own results from the vowel database. The results show that our method gives satisfactory results in a wide range of glottal configurations and at different levels of SNR. Our method using the whitened residual compared favorably to this reference, achieving high quality ratings (Good-Excellent). Our full parametrized system scored lower than the other two ranking in third place, but still higher than the acceptance threshold (Fair-Good). Next we proposed two methods for prosody modification, one for each of the residual representations explained above. The first method used our full parametrization system and frame interpolation to perform the desired changes in pitch and duration. The second method used resampling on the residual waveform and a frame selection technique to generate a new sequence of frames to be synthesized. The results showed that both methods are rated similarly (Fair-Good) and that more work is needed in order to achieve quality levels similar to the reference methods. As part of this dissertation, we have studied the application of our models in three different areas: voice conversion, voice quality analysis and emotion recognition. We have included our speech production model in a reference voice conversion system, to evaluate the impact of our parametrization in this task. The results showed that the evaluators preferred our method over the original one, rating it with a higher score in the MOS scale. To study the voice quality, we recorded a small database consisting of isolated, sustained Spanish vowels in four different phonations (modal, rough, creaky and falsetto) and were later also used in our study of voice quality. Comparing the results with those reported in the literature, we found them to generally agree with previous findings. Some differences existed, but they could be attributed to the difficulties in comparing voice qualities produced by different speakers. At the same time we conducted experiments in the field of voice quality identification, with very good results. We have also evaluated the performance of an automatic emotion classifier based on GMM using glottal measures. For each emotion, we have trained an specific model using different features, comparing our parametrization to a baseline system using spectral and prosodic characteristics. The results of the test were very satisfactory, showing a relative error reduction of more than 20% with respect to the baseline system. The accuracy of the different emotions detection was also high, improving the results of previously reported works using the same database. Overall, we can conclude that the glottal source parameters extracted using our algorithm have a positive impact in the field of automatic emotion classification.