Voice recognition method for recognizing a word in speech

A voice recognizing method in which a word included in an input voice is recognized, comprising the steps of:

• preparing a plurality of standard patterned phonemes;
• extracting a group of standard feature parameters from pieces of voice data obtained from voices of a small number of specified persons for each analyzing time, a standard differential voice power defined by differentiating a standard voice power with respect to time being included in the group of standard feature parameters;
• calculating a standard phoneme similarity between the group of standard feature parameters and one of the standard patterned phonemes for each analyzing time and for each standard patterned phoneme;
• preparing a standard similarity vector having the standard phoneme similarities at the same analyzing time for each analyzing time;
• preparing a standard regression coefficient vector from the standard phoneme similarities for each analyzing time;
• producing a plurality of standard patterned words to be collated with the input voice according to the standard phoneme similarities, each of the standard patterned words being formed of a time series of standard patterned voice pieces, and each of the standard patterned voice pieces being composed of the standard phoneme similarities for the standard patterned phonemes;
• extracting a group of feature parameters from an input voice for each analyzing time, a differential voice power defined by differentiating a voice power with respect to time being included in the group of feature parameters;
• storing a time series of differential voice powers relating to the input voice;
• calculating a phoneme similarity between the group of feature parameters and one of the standard patterned phonemes for each analyzing time and for each standard patterned phoneme;
• producing a similarity vector having the phoneme similarities at the same analyzing time as elements for each analyzing time, a time series of similarity vectors relating to the input voice being formed;
• producing a regression coefficient vector from the phoneme similarities for each analyzing time, a time series of regression coefficient vectors relating to the input voice being formed;
• producing a time series of input voice parameters composed of the time series of differential voice powers, the time series of similarity vectors and the time series of regression coefficient vectors;
• collating the time series of input voice parameters with the time series of standard patterned voice pieces indicating a standard patterned word for each standard patterned word;
• calculating an inter-frame similarity between an input voice parameter and a standard patterned voice piece at the same analyzing time for each analyzing time and for each standard patterned word, each of the inter-frame similarities being composed of a differential voice power similarity between one differential voice power and one standard differential voice power at the same analyzing time, a vector similarity between one similarity vector and one standard similarity vector at the same analyzing time and a regression similarity between one regression coefficient vector and one standard regression coefficient vector at the same analyzing time;
• accumulating the inter-frame similarities according to a continuous dynamic programming matching operation to obtain a word similarity for each standard patterned word;
• selecting a maximum word similarity having a maximum value from among the word similarities defined for the standard patterned words, the maximum word similarity being obtained by accumulating the inter-frame similarities for a correct standard patterned word;
• and
• outputting the correct standard patterned word as a recognized word of the input voice.

A voice recognizing method in which a word included in an input voice is recognized, comprising the steps of:

• preparing a plurality of standard patterned words to be collated with the input voice, each of the standard patterned words being formed of a time series of standard patterned voice pieces;
• preparing a plurality of standard patterned phonemes;
• extracting a group of feature parameters from the input voice for each analyzing time;
• calculating a phoneme similarity between the group of feature parameters and one of the standard patterned phonemes for each analyzing time and for each standard patterned phoneme, a value of each phoneme similarity for an analyzing time relating to a non-voice period placed between voice periods being low, and a value of each phoneme similarity for an analyzing time relating to a voice period being high;
• selecting a maximum phoneme similarity from among the phoneme similarities for all of the standard patterned phonemes at the same analyzing time, the maximum phoneme similarity being obtained for each analyzing time;
• producing a weighting factor according to each of the maximum phoneme similarities, a weighting factor produced from a first maximum phoneme similarity being equal to or larger than another weighting factor produced from a second maximum phoneme similarity having a value which is lower than that of the first maximum phoneme similarity, each weighting factor corresponding to one maximum phoneme similarity for an analyzing time relating to a non-voice period placed between voice periods being low, and each weighting factor corresponding to one maximum phoneme similarity for an analyzing time relating to a voice period being high;
• producing a similarity vector having the phoneme similarities at the same analyzing time as elements for each analyzing time, a time series of similarity vectors relating to the input voice being formed;
• producing a regression coefficient vector from the phoneme similarities for each analyzing time, a time series of regression coefficient vectors relating to the input voice being formed;
• producing a time series of input voice parameters composed of the time series of similarity vectors and the time series of regression coefficient vectors;
• collating the time series of input voice parameters with the time series of standard patterned voice pieces indicating a standard patterned word for each standard patterned word to obtain an inter-frame similarity between an input voice parameter and a standard patterned voice piece at the same analyzing time for each analyzing time and for each standard patterned word;
• subtracting a constant value from each of the inter-frame similarities respectively obtained for each analyzing time and for each standard patterned word to obtain a posterior probability similarity for each analyzing time and for each standard patterned word, the posterior probability similarities for a correct standard patterned word having positive values, and the posterior probability similarities for standard patterned words other than the correct standard patterned word having negative values;
• weighting each of the posterior probability similarities at an analyzing time with the weighting factor at the same analyzing time for each standard patterned word to produce a weighted posterior probability similarity for each analyzing time and for each standard patterned word, each posterior probability similarity for an analyzing time relating to a non-voice period placed between voice periods being weighted with one weighting factor having a low value, and each posterior probability similarity for an analyzing time relating to a voice period being weighted with another weighting factor having a high value;
• accumulating the weighted posterior probability similarities according to a continuous dynamic programming matching operation to obtain a word similarity for each standard patterned word;
• selecting a maximum word similarity having a maximum value from among the word similarities defined for the standard patterned words, the maximum word similarity being obtained by accumulating the weighted posterior probability similarities for the correct standard patterned word;
• and
• outputting the correct standard patterned word as a recognized word of the input voice.

A voice recognizing method in which a word included in an input voice is recognized, comprising the steps of:

• preparing a plurality of standard patterned words to be collated with the input voice, each of the standard patterned words being formed of a time series of standard patterned voice pieces;
• preparing N standard patterned phonemes;
• preparing a standard patterned noise similar to a pattern of a noise;
• extracting a group of feature parameters from the input voice for each analyzing time;
• calculating a phoneme similarity d.sub.P (P=1 to N) between the group of feature parameters and one of the standard patterned phonemes for each analyzing time and for each standard patterned phoneme, a value of each phoneme similarity for an analyzing time relating to a noise period placed between voice periods being low, and a value of each phoneme similarity for an analyzing time relating to a voice period being high;
• calculating a noise similarity Xn between the group of feature parameters and the standard patterned noise for each analyzing time, a value of each noise similarity for an analyzing time relating to a noise period placed between voice periods being high, and a value of each noise similarity for an analyzing time relating to a voice period being low;
• producing an N-dimensional normalized similarity vector a having N normalized phoneme similarities a.sub.p at the same analyzing time as elements for each analyzing time, each of the normalized phoneme similarities a.sub.p being determined according to an equation
• a.sub.P =d.sub.P /(d.sub.1.sup.2 +d.sub.2.sup.2 + . . . +d.sub.N.sup.2 +Xn.sup.2).sup.1/2, and the N-dimensional normalized similarity vector a.sub.p being expressed by a=(a.sub.1, a.sub.2, . . . , a.sub.N);
• producing a time series of normalized similarity vectors relating to the input voice;
• producing a regression coefficient vector from the phoneme similarities for each analyzing time;
• producing a time series of regression coefficient vectors relating to the input voice;
• producing a time series of input voice parameters composed of the time series of normalized similarity vectors and the time series of regression coefficient vectors;
• collating the time series of input voice parameters with the time series of standard patterned voice pieces indicating a standard patterned word for each standard patterned word to obtain an inter-frame similarity between an input voice parameter and a standard patterned voice piece at the same analyzing time for each analyzing time and for each standard patterned word, a value of each inter-frame similarity for an analyzing time relating to a noise period placed between voice periods being low, and a value of each inter-frame similarity for an analyzing time relating to a voice period being high;
• subtracting a constant value from each of the inter-frame similarities respectively obtained for each analyzing time and for each standard patterned word to obtain a posterior probability similarity for each analyzing time and for each standard patterned word, each posterior probability similarity for a correct standard patterned word and an analyzing time relating to a voice period having a high positive value, each posterior probability similarity for one of standard patterned words other than the correct standard patterned word having a negative value, and each posterior probability similarity for an analyzing time relating to a noise period placed between voice periods having a negative value;
• accumulating the posterior probability similarities according to a continuous dynamic programming matching operation to obtain a word similarity for each standard patterned word;
• selecting a maximum word similarity having a maximum value from among the word similarities defined for the standard patterned words, the maximum word similarity being obtained by accumulating the posterior probability similarities for the correct standard patterned word;
• and
• outputting the correct standard patterned word as a recognized word of the input voice.

A voice recognizing method in which a word included in an input voice is recognized, comprising the steps of:

• preparing a plurality of standard patterned phonemes;
• extracting a group of standard feature parameters from pieces of voice data obtained from voices of a small number of specified persons for each analyzing time, a standard differential voice power defined by differentiating a standard voice power with respect to time being included in the group of standard feature parameters;
• calculating a standard phoneme similarity between the group of standard feature parameters and one of the standard patterned phonemes for each analyzing time and for each standard patterned phoneme, a value of each phoneme similarity for an analyzing time relating to a non-voice period placed between voice periods being low, and a value of each phoneme similarity for an analyzing time relating to a voice period being high;
• preparing a standard similarity vector having the standard phoneme similarities at the same analyzing time for each analyzing time;
• preparing a standard regression coefficient vector from the standard phoneme similarities for each analyzing time;
• producing a plurality of standard patterned words to be collated with the input voice according to the standard phoneme similarities, each of the standard patterned words being formed of a time series of standard patterned voice pieces, and each of the standard patterned voice pieces being composed of the standard phoneme similarities for the standard patterned phonemes;
• extracting a group of feature parameters from an input voice for each analyzing time, a differential voice power defined by differentiating a voice power with respect to time being included in the group of feature parameters;
• storing a time series of differential voice powers relating to the input voice;
• calculating a phoneme similarity between the group of feature parameters and one of the standard patterned phonemes for each analyzing time and for each standard patterned phoneme;
• selecting a maximum phoneme similarity from among the phoneme similarities for all of the standard patterned phonemes at the same analyzing time, the maximum phoneme similarity being obtained for each analyzing time;
• producing a weighting factor according to each of the maximum phoneme similarities, a weighting factor produced from a first maximum phoneme similarity being equal to or larger than another weighting factor produced from a second maximum phoneme similarity having a value which is lower than that of the first maximum phoneme similarity, each weighting factor corresponding to one maximum phoneme similarity for an analyzing time relating to a non-voice period placed between voice periods being low, and each weighting factor corresponding to one maximum phoneme similarity for an analyzing time relating to a voice period being high;
• producing a similarity vector having the phoneme similarities at the same analyzing time as elements for each analyzing time, a time series of similarity vectors relating to the input voice being formed;
• producing a regression coefficient vector from the phoneme similarities for each analyzing time, a time series of regression coefficient vectors relating to the input voice being formed;
• producing a time series of input voice parameters composed of the time series of differential voice powers, the time series of similarity vectors and the time series of regression coefficient vectors;
• collating the time series of input voice parameters with the time series of standard patterned voice pieces indicating a standard patterned word for each standard patterned word;
• calculating an inter-frame similarity between an input voice parameter and a standard patterned voice piece at the same analyzing time for each analyzing time and for each standard patterned word, each of the inter-frame similarities being composed of a differential voice power similarity between one differential voice power and one standard differential voice power at the same analyzing time, a vector similarity between one similarity vector and one standard similarity vector at the same analyzing time and a regression similarity between one regression coefficient vector and one standard regression coefficient vector at the same analyzing time;
• subtracting a constant value from each of the inter-frame similarities respectively obtained for each analyzing time and for each standard patterned word to obtain a posterior probability similarity for each analyzing time and for each standard patterned word, the posterior probability similarities for a correct standard patterned word having positive values, and the posterior probability similarities for standard patterned words other than the correct standard patterned word having negative values, and
• weighting each of the posterior probability similarities at an analyzing time with the weighting factor at the same analyzing time for each standard patterned word to produce a weighted posterior probability similarity for each analyzing time and for each standard patterned word, each posterior probability similarity for an analyzing time relating to a non-voice period placed between voice periods being weighted with one weighting factor having a low value, and each posterior probability similarity for an analyzing time relating to a voice period being weighted with another weighting factor having a high value;
• accumulating the weighted posterior probability similarities according to a continuous dynamic programming matching operation to obtain a word similarity for each standard patterned word;
• selecting a maximum word similarity having a maximum value from among the word similarities defined for the standard patterned words, the maximum word similarity being obtained by accumulating the weighted posterior probability similarities for the correct standard patterned word;
• and
• outputting the correct standard patterned word as a recognized word of the input voice.