초록 ▼

A system and method for determining from continuous speech, the instantaneous values of a set of articulatory parameters. The continuous speech data is a sequence of spectral profiles obtained by spectrally sampling continuous speech. The spectral samples are presented in sequence to a plurality of

A system and method for determining from continuous speech, the instantaneous values of a set of articulatory parameters. The continuous speech data is a sequence of spectral profiles obtained by spectrally sampling continuous speech. The spectral samples are presented in sequence to a plurality of class transforms, each establishing a respective speech phoneme class which includes plurality of speech phoneme having similar spectral and articulatory characteristics. Each class transform converts a speech segment included in its class and contained in a spectral sample into a predetermined set of articulatory parameter values. A class-discriminating transform operates in parallel with the class transforms to produce a set of probability values, each indicating the probability that the spectral sample being transformed represents a phoneme in a respective speech phoneme class. An array of multipliers adjusts the predetermined values of the sets produced by the class transforms by multiplying the values of each set by the probability value produced for that set by the class-discriminating transform. The adjusted articulatory parameter value sets are combined by adding corresponding elements to produce a set of adjusted articulatory parameter values indicative of an articulatory tract configuration appropriate for producing the sampled speech.

대표청구항 ▼

A method of determining the values of a series of N articulatory parameters from speech data, comprising the steps of: creating a plurality of speech phoneme classes, each of said speech phoneme classes including a plurality of speech phonemes sharing similar spectral and articulatory characteristic

A method of determining the values of a series of N articulatory parameters from speech data, comprising the steps of: creating a plurality of speech phoneme classes, each of said speech phoneme classes including a plurality of speech phonemes sharing similar spectral and articulatory characteristics; providing a digital speech data signal representative of speech; selecting data segments of said speech data signal at predetermined sampling intervals according to predefined changes in energy levels in said speech data signal; transforming said selected data segments into spectral data segments; converting each of said spectral data segments into said speech phoneme classes so as to generate a weight for the probability that said segment corresponds to phonemes within each of said classes; converting each of said spectral data segments into a plurality of articulatory parameters for each of said speech phoneme classes so as to generate a series of N parameter values representative of articulatory characteristics in each speech phoneme class; and combining the weight for the probability that spectral data segments correspond to a given speech phoneme class with the output parameter values from each speech phoneme class so as to form a single series of N parameter values for selected data segments.