Diet

A technique is described for producing smooth speech parametric representation evolutions by means of an application of traditional hidden Markov modeling techniques. It is based on the consideration that HMM training is able to locate the main acoustic events which occur during the speech process; thus the decomposition obtained can be used to reproduce the linguistic units utilized for training. This is accomplished by interpolating the state-related features values with some weighting functions, as it is done for temporal decomposition technique. Results will be given for isolated word synthesis.

A new duration intrinsic model for improved speech recognition by HMM techniques is presented. Assuming an exponentially decaying time dependency of the states loop probability, the duration density can be factorized and a path early pruning theorem demonstrated. As a consequence, computational complexity is greatly reduced with respect to explicit models, whereas recognition performances improve considerably.

A simple solution to the word spotting problem is experimented. Continuous Speech Recognition (CSR) techniques are adopted, with the background speech represented by a filler templates set defined through vector quantization algorithms. Quantitative results about the performed experiments are given.

A synthetic speech segmental quality assessment method based on nonsense pseudo-word list is proposed. The pseudo-words are automatically generated by a Stochastic Phonotax, whose transition diagram captures various aspects of the phonological structure of the language, such as the syllabic stress, and morphological constraints (Falaschi, 1987). The composition of the test material thus reflects the real oppositions of the language. Full coverage of the phonemic contexts and phonotactical roles is ensured by the use of a compacting selection procedure (Falaschi, 1990). The phonemic confusion and the recognition rate of listeners are automatically scored by an algorithm for recognizer performance evaluation. Analysis of the test results provides valuable diagnostic hints about the synthesizer's main weaknesses. Test results are given, involving five synthesizers for Italian plus one human voice as control data. The Stochastic Phonotax which generates the pseudo-words may be built for any language, so that the method is naturally suited for the definition of typical complexity test sets across different languages.

Un esperimento iniziato nel 2001, e divenuto nel tempo un riferimento culturale su scala nazionale. Ritengo che il fine del mio lavoro sia aiutare a comprendere, e che la cultura debba essere libera e senza ostacoli di accesso. Chi acquista il formato cartaceo, ebook, od effettua una donazione, entra a far parte dei sostenitori del progetto, e può scaricare la versione PDF dell'edizione più recente, perfettamente navigabile e stampabile. L'edizione più recente è di Giugno 2019, la trovi presso https://teoriadeisegnali.it/

Abstract The authors deal with the extension of the hidden control neural network (HCNN) architecture to the ergodic case, ie, if all the control state sequences are allowed. This scheme gives a deeper understanding of the modeling capabilities offered by the HCNN formalism. In fact, the control input binary digits status can be considered as the presence/absence of a posteriori defined binary phonetic features, forcing the network to produce a low prediction error on pairs of speech frames.

Abstract A low-bit-rate codec based on an ergodic hidden Markov model (EHMM) is presented. A 256-state autoregressive Gaussian EHMM is trained on speech uttered by eight different speakers by means of the Baum-Welch algorithm. Initial estimates are obtained from vector quantization. The resulting EHMM is then utilized for a Viterbi decoding of incoming speech data. The state sequence obtained is frame synchronously encoded.