A phonetic Hidden Markov Model speech synthesizer

Claim 1

A method for generating synthesized speech wherein an acoustic ergodic hidden Markov model (AEHMM), reflecting phonotactical constrains is correlated to a phonetic ergodic hidden Markov model (PhEHMM), the method comprising the steps of a) Building an AEHMM, considering as observations sequence the speech features vectors extracted from the frames in which the speech uttered during the training of said AEHMM is divided and as hidden sequence the sequence of the sources that most probably have emitted the speech utterance frames, b) Initializing said AEHMM by a Vector Quantization clustering scheme, having the same size of said AEHMM, c) Training said AEHMM by the Forward-Backward algorithm and Baum-Welch re-estimation formulas, d) Associating to each frame the label of the most probable source, e) Building a PhEHMM of the same size of said AEHMM, considering as observations sequence a synthetic observations sequence obtained from a written text and as hidden sequence the sequence of the labels, f) Initializing said PhEHMM transition probability matrix by assigning to the state transition probabilities the same values of the transition probabilities of the corresponding states of said AEHMM, g) Initializing said PhEHMM observation probability functions by: i) using a phonetically aligned speech corpus ii) generating for said speech corpus the sequence of the most probable labels, using the acoustic labelling feature of said AEHMM, iii) computing the observations probability function for each phoneme, counting the number of occurrences of the phoneme in a state divided by the total number of phonemes emitted by said state, h) Training said PhEHMM by the Baum-Welch algorithm on a proper synthetic observations corpus, i) Determining for each word to be synthesized a synthetic observation and through said PhEHMM the corresponding sequence of labels by means of a proper optimality criterion, j) Determining from the input text a set of additional parameters, as energy, prosody contours and voicing, by a prosodic processor, k) Determining for said sequence of labels the corresponding set of speech features vectors through said AEHMM, I) Transforming said speech features vectors into a set of filter coefficients preserving the spectral information and m) Using said set of filter coefficients and said additional parameters in a synthesis filter to produce a synthetic speech output.

Claim 2

A method for generating synthesized speech wherein an acoustic ergodic hidden Markov model (AEHMM), reflecting phonotactical constrains is correlated to a phonetic ergodic hidden Markov model (PhEHMM), the method comprising the steps of a) Building an AEHMM, considering as observations sequence the speech features vectors extracted from the frames in which the speech uttered during the training of said AEHMM is divided and as hidden sequence the sequence of the sources that most probably have emitted the speech utterance frames, b) Initializing said AEHMM by a Vector Quantization clustering scheme, having the same size of said AEHMM, c) Training said AEHMM by the Forward-Backward algorithm and Baum-Welch re-estimation formulas, d) Associating to each frame the label of the most probable source, e) Building a PhEHMM of the same size of said AEHMM, considering as observations sequence a synthetic observations sequence obtained from a written text and as hidden sequence the sequence of the labels, f) Initializing said PhEHMM transition probability matrix by assigning to the state transition probabilities the same values of the transition probabilities of the corresponding states of said AEHMM, g) Initializing said PhEHMM observation probability functions by: i) using a phonetically aligned speech corpus ii) generating for said speech corpus the sequence of the most probable labels, using the acoustic labelling feature of said AEHMM, iii) computing the observations probability function for each phoneme, counting the number of occurrences of the phoneme in a state divided by the total number of phonemes emitted by said state, h) Training said PhEHMM by the Baum-Welch algorithm on a proper synthetic observations corpus, i) Determining for each word to be synthesized a synthetic observation and through said PhEHMM the corresponding sequence of labels by means of a proper optimality criterion, j) Determining from the input text a set of additional parameters, as energy, prosody contours and voicing, by a prosodic processor, k) Determining for said sequence of labels the corresponding set of speech features vectors through said AEHMM, I) Transforming said speech features vectors into a set of filter coefficients preserving the spectral information and m) Using said set of filter coefficients and said additional parameters in a synthesis filter to produce a synthetic speech output.