Noise estimation apparatus, noise estimation method, noise estimation program, and recording medium

A noise estimation method comprising:

• a step, by circuitry of a noise estimation apparatus, of receiving, as an input, complex spectra of inputted observed waveform signals, which are acoustic signals that include clean speech mixed with a noise signal, up to a current frame;
• obtaining a variance of the noise signal, where the noise signal follows a complex Gaussian distribution, such that a value of weighted addition of sums becomes large, wherein:
• each of the sums is obtained by adding a first product and a second product;
• the first product in each frame is a product of a log likelihood of a model of an observed signal expressed by a Gaussian distribution in a speech segment and a speech posterior probability;
• and the second product in each frame is a product of a log likelihood of a model of an observed signal expressed by a Gaussian distribution in a non-speech segment and a non-speech posterior probability;
• and
• the method includes estimating, by the circuitry, a variance σv,i2 of the noise signal in the current frame i by weighted addition of a complex spectrum Yi of an observed signal in the current frame i and a variance σv,i-τ2 of the noise signal estimated in a past frame where τ is an integer greater than 1, on the basis of a non-speech posterior probability estimated in the current frame, and
• outputting the variance σv,i2 of the noise signal for cancellation of the noise signal from the acoustic signals, wherein the cancellation of the noise signal includes subtracting a power spectrum of the noise signal, which is estimated based on the outputted variance σv,i2 from a power spectrum of the observed waveform signals.

A non-transitory computer-readable recording medium having recorded thereon a noise estimation program which when executed by a noise estimation apparatus, causes the noise estimation apparatus to perform a method comprising:

• a step, by circuitry of a noise estimation apparatus, of receiving, as an input, complex spectra of inputted observed waveform signals, which are acoustic signals that include clean speech mixed with a noise signal, up to a current frame;
• obtaining a variance of the noise signal, where the noise signal follows a complex Gaussian distribution, such that a value of weighted addition of sums becomes large, wherein:
• each of the sums is obtained by adding a first product and a second product, the first product in each frame is a product of a log likelihood of a model of an observed signal expressed by a Gaussian distribution in a speech segment and a speech posterior probability;
• and the second product in each frame is a product of a log likelihood of a model of an observed signal expressed by a Gaussian distribution in a non-speech segment and a non-speech posterior probability;
• and
• the method includes estimating, by the circuitry, a variance σv,i2 of the noise signal in the current frame i by weighted addition of a complex spectrum Yi of an observed signal in the current frame i and a variance σv,i-τ2 of the noise signal estimated in a past frame where τ is an integer greater than 1, on the basis of a non-speech posterior probability estimated in the current frame, and
• outputting the variance σv,i2 of the noise signal for cancellation of the noise signal from the acoustic signals, wherein the cancellation of the noise signal includes subtracting a power spectrum of the noise signal, which is estimated based on the outputted variance σv,i2, from a power spectrum of the observed waveform signals.