Noise eliminating apparatus and speech recognition apparatus using the same

A speech recognition apparatus comprising:

• a) a noise eliminating apparatus for eliminating a noise contained in an input speech signal therefrom and outputting a noise-eliminated input speech signal;
• b) input pattern creating means, coupled to said noise eliminating apparatus, for creating an input pattern of said input speech signal from said noise-eliminated input speech signal;
• c) reference pattern storage means for storing reference patterns;
• and
• d) speech recognition means, coupled to said input pattern creating means and said reference pattern storage means, for comparing said input pattern created by said input pattern creating means with said reference patterns and for outputting one of said reference patterns which has the highest degree of similarity to said input pattern;
• wherein said noise eliminating apparatus includes:
• 1) first amplifier means for amplifying a first acoustic signal and for generating a first amplified acoustic signal, said first acoustic signal including the input speech signal to be recognized;
• 2) first feature extracting means, coupled to said first amplifier means, for generating a feature quantity of said first amplified acoustic signal for each of n channels, said feature quantity including a power spectrum;
• 3) second amplifier means for amplifying a second acoustic signal and for generating a second amplified acoustic signal;
• 4) second feature extracting means, coupled to said second amplifier means, for generating a feature quantity of said second amplified acoustic signal for each of said no channels, said feature quantity including a power spectrum;
• 5) speech section detecting means, coupled to said first and second feature extracting means, for detecting a speech section of said first acoustic signal for each of said n channels by using the feature quantity of said first amplified acoustic signal and the feature quantity of said second amplifier acoustic signal;
• and
• 6) a noise eliminating unit, coupled to said first and second feature extracting means and said speech section detecting means, said noise eliminating unit including:
• A) first means for generating a first noise eliminating variable k(f) and a second noise eliminating variable d(f) on the basis of a feature quantity Xn(f) extracted by said first feature extracting means and a feature quantity Nn(f) extracted by said second feature extracting means when said first acoustic signal does not include the input speech signal and thus said speech section detecting means does not detect the speech section;
• and
• B) second means for generating a feature quantity of said input speech signal which corresponds to said noise-eliminated input speech signal on the basis of a feature quantity Xs(f) extracted by said first feature extracting means and a feature quantity Ns(f) extracted by said second feature quantity extracting means when said first acoustic signal includes the input speech signal and thus said speech section detecting means detects the speech section;
• wherein said first noise eliminating variable k(f) and said second noise eliminating variable d(f) are defined as a function satisfying the following:
• (i) the first noise eliminating variable k(f) approaches zero and the second noise eliminating variable d(f) approaches the feature quantity Xn(f) as the feature quantity Xn(f) becomes smaller than a predetermined value;
• and
• (ii) the first noise eliminating variable k(f) approaches Xn(F)/Nn(f) and the second noise eliminating variable d(f) approaches zero as the feature quantity Xn(f) becomes greater than the predetermined value.

A method of eliminating noise especially to facilitate speech recognition, the method comprising:

• a) amplifying a first acoustic signal to generate a first amplified acoustic signal, the first acoustic signal including an input speech signal to be recognized;
• b) generating a feature quantity of the first amplified acoustic signal for each of n channels, the feature quantity including a power spectrum;
• c) amplifying a second acoustic signal to generate a second amplified acoustic signal;
• d) generating a feature quantity of the second amplified acoustic signal for each of the n channels, the second feature quantity including a power spectrum;
• e) detecting a speech section of the first acoustic signal for each of the n channels by using the feature quantity of the first amplified acoustic signal and the feature quantity of the second amplified acoustic signal;
• and
• f) a step including either;
• 1) generating a first noise eliminating variable k(f) and a second noise eliminating variable d(f) on the basis of a feature quantity Xn(f) extracted in the first generating step and a feature quantity Nn(f) extracted in the second generating step, when the first acoustic signal does not include the input speech signal;
• or
• 2) generating a feature quantity of the input speech signal which is to be supplied to a recognition unit on the basis of a feature quantity Xs(f) extracted in the first generating step and a feature quantity Ns(f) extracted in the second generating step, when the first acoustic signal includes the input speech signal;
• wherein the first noise eliminating variable k(f) and the second noise eliminating variable d(f) are defined as a junction satisfying the following:
• (i) the first noise eliminating variable k(f) approaches zero and the second noise eliminating variable d(f) approaches the feature quantity Xn(f) as the feature quantity Xn(f) becomes smaller than a predetermined value;
• and
• (ii) the first noise eliminating variable k(f) approaches Xn(f)/Nn(f) and the second noise eliminating variable d(f) approaches zero as the feature quantity Xn(f) becomes greater than the predetermined value.

A speech recognition method, comprising:

• 1) eliminating noise contained in an input speech signal and outputting a noise-eliminated input speech signal by performing the steps of:
• a) amplifying a first acoustic signal to generate a first amplified acoustic signal, the first acoustic signal including an input speech signal to be recognized;
• b) generating a feature quantity of the first amplified acoustic signal for each of n channels, the feature quantity including a power spectrum;
• c) amplifying a second acoustic signal to generate a second amplified acoustic signal;
• d) generating a feature quantity of the second amplified acoustic signal of reach of the n channels, the second feature quantity including a power spectrum;
• (e) detecting a speech section of the first acoustic signal for each of the n channels by using the feature quantity of the first amplified acoustic signal and the feature quantity of the second amplified acoustic signal;
• and
• f) a step including either;
• (1) generating a first noise eliminating variable k(f) and a second noise eliminating variable d(f) on the basis of a feature quantity Xn(f) extracted in the first generating step and a feature quantity Nn(f) extracted in the second generating step, when the first acoustic signal does not include the input speech signal;
• or
• (2) generating a feature quantity of the input speech signal which is to be supplied to a recognition unit on the basis of a feature quantity Xs(f) extracted in the first generating step and a feature quantity Ns(f) extracted in the second generating step, when the first acoustic signal includes the input speech signal;
• wherein the first noise eliminating variable k(f) and the second noise eliminating variable d(f) are defined as a function satisfying the following:
• (i) the first noise eliminate variable k(f) approaches zero and the second noise eliminating variable d(f) approaches the feature quantity Xn(f) as the feature quantity Xn(f) becomes smaller than a predetermined value;
• and
• (ii) the first noise eliminating variable k(f) approaches Xn(f)/Nn(f) and the second noise eliminating variable d(f) approaches zero as the feature quantity Xn(f) becomes greater than the predetermined value;
• 2) creating an input pattern of the input speech signal from the noise-eliminated input speech signal;
• 3) comparing the input pattern with stored reference patterns;
• and
• 4) outputting one of the reference patterns which has the highest degree of similarity to the input pattern.