Speech recognition method and apparatus using frequency warping of linear prediction coefficients

A speech recognition apparatus comprising:

• mel-linear predictive analysis means for calculating mel-linear predictive coefficients corresponding to a spectrum envelope having different resolutions for different frequencies, from input speech, according to a frequency weighting function based on human auditory characteristics, said met-linear predictive analysis means including:
• extracting means for extracting an input signal portion, having a specified time length, from an input speech signal;
• multiple stages of all-pass filters operable to produce output signals, and a linear coupling unit operable to receive the output signals and to perform frequency warping of linear prediction coefficients to thereby obtain frequency-warped output signals, wherein the linear prediction coefficients correspond to a spectrum envelope having different resolutions for different frequencies according to a frequency weighting function based on human auditory characteristics;
• correlation function means for performing a correlation function on a mel-frequency axis, said correlation function being a product-sum operation between the input signal portion and the output signal from each filter performed within a range restricted to the time length of the input signal portion and represented by the following formula:
• φ ⁢ ⁢ ( i , j ) = ∑ n = 0 N - 1 ⁢ ⁢ x ⁡ [ n ] · y ( i - j ) ⁡ [ n ]
• wherein φ(i,j) is the correlation function, x[n] is the input signal, and y(i-j)[n] is the output signal from each filter;
• and
• obtaining means for obtaining the parameters of the spectrum envelope based on mel-linear predictive coefficients from the correlation function on the mel-frequency axis, and recognizing the input speech based on the mel-linear predictive coefficients or cepstrum coefficients obtained from the mel-linear predictive coefficients.

A speech recognition method using frequency warping of linear prediction coefficients employing the following steps as a method for obtaining parameters corresponding to a spectrum envelope based on human auditory characteristics from an input speech:

• cutting out an input signal of a specific time length from an input speech, and filtering the signal of the time length using multiple stages of all-pass filters to obtain output signals from the respective filters;
• obtaining a correlation function on a mel-frequency axis by-performing a product-sum operation between the input signal and the output signal from each filter, which product-sum operation is performed within a range restricted to the time length of the input signal as represented by the following formula, φ ⁢ ⁢ ( i , j ) = ∑ n = 0 N - 1 ⁢ ⁢ x ⁡ [ n ] · y ( i - j ) ⁡ [ n ]
• wherein φ(i,j) is the correlation function, x[n] is the input signal, and y(i-j)[n] is the output signal from each filter;
• and
• obtaining mel-linear predictive coefficients from the correlation function on the mel-frequency axis;
• wherein the input speech is recognized using the mel-linear predictive coefficients, or cepstrum coefficients obtained from the mel-linear predictive coefficients.

A speech recognition apparatus using frequency warping of linear prediction coefficients, comprising:

• a mel-linear predictive analysis means for obtaining parameters corresponding to a spectrum envelope based on human auditory characteristics, from an input speech, said mel-linear predictive analysis means comprising:
• a first means for cutting out an input signal of a specific time length from an input speech, and filtering a signal of a time length using multiple stages of all-pass filters to obtain output signals from the respective filters;
• a second means for obtaining a correlation function on a mel-frequency axis by performing a product-sum operation between the input signal and the output signal from each filter, which product-sum operation is performed within a range restricted to the time length of the input signal as represented by the following formula, φ ⁢ ⁢ ( i , j ) = ∑ n = 0 N - 1 ⁢ ⁢ x ⁡ [ n ] · y ( i - j ) ⁡ [ n ]
• wherein φ(i,j) is the correlation function, x[n] is the input signal, and y(i-j)[n] is the output signal from each filter;
• and
• a third means for obtaining mel-linear predictive coefficients from the correlation function on the mel-frequency axis.