Articulatory copy synthesis (ACS), a subarea of speech inversion, refers to the reproduction of natural utterances and involves both the physiological articulatory processes and their corresponding acoustic results. This thesis proposes two novel methods for the ACS of human speech using the articulatory speech synthesizer VocalTractLab (VTL) to address or mitigate the existing problems of speech inversion, such as non-unique mapping, acoustic variation among different speakers, and the time-consuming nature of the process.
The first method involved finding appropriate VTL gestural scores for given natural utterances using a genetic algorithm. It consisted of two steps: gestural score initialization and optimization. In the first step, gestural scores were initialized using the given acoustic signals with speech recognition, grapheme-to-phoneme (G2P), and a VTL rule-based method for converting phoneme sequences to gestural scores. In the second step, the initial gestural scores were optimized by a genetic algorithm via an analysis-by-synthesis (ABS) procedure that sought to minimize the cosine distance between the acoustic features of the synthetic and natural utterances. The articulatory parameters were also regularized during the optimization process to restrict them to reasonable values.
The second method was based on long short-term memory (LSTM) and convolutional neural networks, which were responsible for capturing the temporal dependence and the spatial structure of the acoustic features, respectively. The neural network regression models were trained, which used acoustic features as inputs and produced articulatory trajectories as outputs. In addition, to cover as much of the articulatory and acoustic space as possible, the training samples were augmented by manipulating the phonation type, speaking effort, and the vocal tract length of the synthetic utterances. Furthermore, two regularization methods were proposed: one based on the smoothness loss of articulatory trajectories and another based on the acoustic loss between original and predicted acoustic features.
The best-performing genetic algorithms and convolutional LSTM systems (evaluated in terms of the difference between the estimated and reference VTL articulatory parameters) obtained average correlation coefficients of 0.985 and 0.983 for speaker-dependent utterances, respectively, and their reproduced speech achieved recognition accuracies of 86.25% and 64.69% for speaker-independent utterances of German words, respectively. When applied to German sentence utterances, as well as English and Mandarin Chinese word utterances, the neural network based ACS systems achieved recognition accuracies of 73.88%, 52.92%, and 52.41%, respectively. The results showed that both of these methods not only reproduced the articulatory processes but also reproduced the acoustic signals of reference utterances. Moreover, the regularization methods led to more physiologically plausible articulatory processes and made the estimated articulatory trajectories be more articulatorily preferred by VTL, thus reproducing more natural and intelligible speech. This study also found that the convolutional layers, when used in conjunction with batch normalization layers, automatically learned more distinctive features from log power spectrograms. Furthermore, the neural network based ACS systems trained using German data could be generalized to the utterances of other languages.
| Identifer | oai:union.ndltd.org:DRESDEN/oai:qucosa:de:qucosa:80324 |
| Date | 04 August 2022 |
| Creators | Gao, Yingming |
| Contributors | Birkholz, Peter, Möbius, Bernd, Laprie, Yves, Technische Universität Dresden |
| Source Sets | Hochschulschriftenserver (HSSS) der SLUB Dresden |
| Language | English |
| Detected Language | English |
| Type | info:eu-repo/semantics/publishedVersion, doc-type:doctoralThesis, info:eu-repo/semantics/doctoralThesis, doc-type:Text |
| Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.002 seconds