JPWO2015093025A1 - Audio processing apparatus, audio processing method, and audio processing program - Google Patents

Abstract

Even when the effect obtained by selecting the distribution included in the speech model based on the time transition information related to speech features is utilized, and state transitions related to past speech features are continuously strongly influenced by noise High noise suppression accuracy can be obtained. The speech processing device 4 includes a storage unit 40 that stores a first speech model having time transition information indicating the results of state transitions related to speech feature values, and a second speech model having no time transition information; From the information indicated by the input signal, which is a signal in which a speech component and a noise component are mixed, one of the first and second speech models is selected according to a predetermined criterion, and the speech component is selected using the selected speech model and input signal. An expected value calculation unit 60 that calculates an expected value for the noise, and a noise suppression unit 81 that generates noise-suppressed speech in which a noise component included in the input signal is suppressed using the expected value.

Description

  The present invention relates to a speech processing apparatus that performs processing for suppressing a noise component included in a speech signal.

  In recent years, a model-based noise suppression / speech enhancement technique that uses a speech model obtained by modeling a speech pattern for noise suppression has been developed. Such a technique is intended to reduce distortion caused by excessive noise suppression and to perform conversion suitable for voice recognition and abnormal sound detection.

  As a technique related to such a technique, Patent Document 1 discloses a noise suppression that obtains a temporary estimated speech spectrum from an input signal spectrum and an estimated value of a noise spectrum, and corrects the temporary estimated speech spectrum using a speech model. A system is disclosed. Such a noise suppression system calculates a noise suppression filter from the corrected temporary estimated speech spectrum and the estimated value of the noise spectrum, and calculates an estimated speech spectrum from the noise suppression filter and the input signal spectrum.

  Further, Patent Document 2 discloses a noise suppression system using a model to which time transition information of speech is added as a speech model when calculating speech feature values of estimated speech. Such a noise suppression system realizes feature amount correction by distribution selection along the transition of past speech feature amounts.

  Furthermore, Patent Document 3 holds a speech model having a plurality of details representing the characteristic properties of speech, and selects the detail closest to the feature properties of the input signal related to speech based on the speech model. A speech recognition apparatus is disclosed. The speech recognition apparatus controls parameters related to speech recognition according to the selected level of detail.

Japanese Patent No. 4754661 JP 2005-84653 A International Publication No. 2008/108232

Pedro J. Moreno, Bhisha Raj and Richard M. Stern, “A Vector Taylor Series Approach for Independent Inventive Recognition.

  The technique disclosed in Patent Document 2 can improve the noise suppression system more than the technique disclosed in Patent Document 1 in a normal case. However, particularly in a noisy environment, the use of the speech model added with the time transition information of speech for the speech feature amount calculation of the estimated speech disclosed in Patent Document 2 may reduce the noise suppression accuracy.

  For example, when past voice feature values are continuously strongly influenced by noise, the distribution selection accuracy decreases. At this time, in the noise suppression system disclosed in Patent Literature 2, there is a possibility that the noise suppression accuracy may be lower than that of the technology according to Patent Literature 1 that does not use the transition of past speech feature values.

  The technology disclosed in Patent Document 3 improves the speech recognition accuracy by using a plurality of speech models, but does not mention the technology for improving the noise suppression accuracy.

  Therefore, while taking advantage of the effect obtained by selecting the distribution included in the speech model based on time transition information, by preventing the deterioration of accuracy even when the transition of past speech feature is continuously strongly influenced by noise, It becomes a problem to perform signal processing with high noise suppression accuracy.

  The main object of the present invention is to provide an audio processing apparatus and the like that can solve the above-described problems.

  A speech processing apparatus according to an aspect of the present invention includes: a first speech model having time transition information indicating a result of state transition relating to speech feature amounts; and a second speech model not having the time transition information. From the information indicated by the storage means for storing and the input signal that is a signal in which a speech component and a noise component are mixed, the first and second speech models are selected according to a predetermined criterion, and the selected speech model and Expected value calculating means for calculating an expected value related to the speech component using the input signal, and generating a first noise-suppressed speech in which the noise component included in the input signal is suppressed using the expected value. Noise suppression means.

  In another aspect of achieving the above object, a speech processing method according to an aspect of the present invention includes: a first speech model having time transition information indicating a record of state transitions related to speech feature amounts by an information processing device; By referring to the storage means in which the second speech model having no time transition information is stored, the first and second information can be obtained from the information indicated by the input signal, which is a signal in which speech components and noise components are mixed. Is selected according to a predetermined criterion, and an expected value for the speech component is calculated using the selected speech model and the input signal, and is included in the input signal using the expected value. A first noise-suppressed speech in which the noise component is suppressed is generated.

  Further, in a further aspect of achieving the above object, a computer-readable recording medium according to one aspect of the present invention includes a first speech model having time transition information indicating a record of state transitions related to speech features, By referring to the storage means in which the second speech model having no time transition information is stored, the first and second information can be obtained from the information indicated by the input signal, which is a signal in which speech components and noise components are mixed. Selected according to a predetermined criterion, using the selected speech model and the input signal, an expected value calculation process for calculating an expected value for the speech component, and using the expected value, A sound processing program for causing a computer to execute noise suppression processing for generating first noise-suppressed speech in which the noise component included in the input signal is suppressed. .

  Furthermore, the present invention can also be realized by a computer program stored in such a storage medium.

  The present invention takes advantage of the effect obtained by selecting the distribution included in the speech model based on the time transition information indicating the state transition results regarding the speech feature amount, and the state transition regarding the past speech feature amount is caused by noise. High noise suppression accuracy can be obtained even when a strong influence is continuously received.

It is a block diagram which shows the structure of the audio processing apparatus which concerns on 1st Embodiment of this invention. It is a flowchart which shows operation | movement of the audio processing apparatus which concerns on 1st Embodiment of this invention. It is a block diagram which shows the structure of the audio processing apparatus which concerns on 2nd Embodiment of this invention. It is a block diagram which shows the structure of the audio processing apparatus which concerns on 3rd Embodiment of this invention. It is a block diagram which shows the structure of the audio processing apparatus which concerns on 4th Embodiment of this invention. It is a block diagram which shows the structure of the information processing apparatus which can perform the audio | voice processing apparatus which concerns on each embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<First Embodiment>
FIG. 1 is a block diagram conceptually showing the structure of the speech processing apparatus 1 of the first embodiment.

The speech processing apparatus 1 according to the present embodiment uses “model adaptability” indicating how much the input signal is adapted to the speech model as a numerical value indicating the degree of noise burial in the input signal. Then, when performing noise suppression processing, the speech processing apparatus 1 uses a speech model that does not have time transition information when the model suitability is low, and speech that has time transition information when the model suitability is high. Use the model. Hereinafter, a speech model that does not have time transition information may be referred to as “model A”, and a speech model that has time transition information may be referred to as “model B”.
As shown in FIG. 1, the speech processing apparatus 1 includes an input signal acquisition unit 10, a noise estimation unit 20, a temporary estimated speech calculation unit 30, a storage unit 40, a speech model suitability calculation unit 50, an expected value calculation unit 60, and a filter calculation. Unit 70 and filtering unit 80.

  The input signal acquisition unit 10, the noise estimation unit 20, the provisional estimated speech calculation unit 30, the speech model fitness calculation unit 50, the expected value calculation unit 60, the filter calculation unit 70, and the filtering unit 80 may be electronic circuits. For example, it may be a computer program and a processor that operates according to the computer program. The storage unit 40 is an electronic circuit or an electronic device such as a magnetic disk or an electronic memory that is access-controlled by a computer program and a processor that operates according to the computer program.

  The input signal acquisition unit 10 has a function of acquiring an input signal related to sound. The input signal acquired by the input signal acquisition unit 10 is an input signal acquired from a microphone or the like through an A (Analog) / D (Digital) converter, for example. Examples of the input signal acquired by the input signal acquisition unit 10 include an input signal read from a hard disk and an input signal obtained from a communication packet. However, the input signal in the present invention is not limited to these.

  The input signal acquisition unit 10 classifies a digital signal corresponding to the acquired input signal for each unit time. Hereinafter, in the present application, the extracted digital signal for one section is referred to as a frame. Further, a frame corresponding to the input signal at time t is expressed as x (t−τ) (τ = 0,..., T−1: T is the number of samples included in the frame).

  For example, a digital signal generated by linear PCM (Pulse Code Modulation) having a sampling frequency of 8000 Hz (Hertz) and a quantization bit number of 16 bits includes a value of 8000 points per second. At this time, if the length of one frame is 25 milliseconds, one frame includes values for 200 points. That is, in this case, T = 200.

  Note that the input signal acquisition unit 10 may convert a frame corresponding to the input signal into an absolute value of a spectrum or a power spectrum using short-time discrete Fourier transform. In the following description, the input signal acquisition unit 10 outputs the absolute value of the spectrum of each frame, but the present invention is not limited to this. Hereinafter, the absolute value of the spectrum is also simply referred to as a spectrum.

  The spectrum of the input signal at time t is assumed to be X (t, k) (k = 0,..., K−1). Here, k is a frequency bin, and K is the number of frequency bins.

  The input signal acquisition unit 10 supplies the spectrum X (t, k) of the input signal to the noise estimation unit 20, the temporary estimated speech calculation unit 30, and the filtering unit 80, respectively.

  The noise estimation unit 20 has a function of estimating the spectrum of the noise component included in the spectrum of the input signal. Specifically, the noise estimation unit 20 acquires the spectrum X (t, k) of the input signal from the input signal acquisition unit 10. And the noise estimation part 20 estimates the spectrum of the noise component contained in the spectrum X (t, k) of the acquired input signal. The noise estimation unit 20 supplies the estimated spectrum of the noise component to the temporary estimated speech calculation unit 30 and the filter calculation unit 70, respectively.

  Hereinafter, the noise component estimated by the noise estimation unit 20 is referred to as estimated noise. Further, the spectrum of the estimated noise is N ^ (t, k) (k = 0,..., K−1) using a hat symbol “^” indicating an estimated value. In the present application, the hat symbol is written to the right of the immediately preceding character due to restrictions on document notation, but in general, the hat symbol “^” is written above the immediately preceding character. In the present embodiment, the noise estimation unit 20 calculates the estimated noise by using a weighted noise estimation method (WiNE) that is a known technique, but the method used by the present invention is the same. It is not limited. The noise estimation unit 20 may calculate the estimated noise using a predetermined method other than WiNE.

  The temporary estimated speech calculation unit 30 has a function of removing the estimated noise from the spectrum of the input signal and calculating the spectrum and speech feature amount of the temporary estimated speech. Specifically, the temporary estimated speech calculation unit 30 acquires the spectrum X (t, k) of the input signal from the input signal acquisition unit 10. In addition, the temporary estimated speech calculation unit 30 acquires the estimated noise spectrum N ^ (t, k) from the noise estimation unit 20. Then, the temporary estimated speech calculation unit 30 calculates the spectrum of the temporary estimated speech based on the acquired spectrum X (t, k) of the input signal and the estimated noise spectrum N ^ (t, k). Then, the temporary estimated speech calculation unit 30 calculates speech feature quantities such as a mel spectrum and a mel cepstrum based on the spectrum of the temporary estimated speech. Note that these audio feature quantities may include a primary dynamic component and a secondary dynamic component. However, the audio feature amount of the present invention is not limited to these. The temporary estimated speech calculation unit 30 supplies the calculated speech feature amount of the temporary estimated speech to the speech model suitability calculation unit 50 and the expected value calculation unit 60.

Hereinafter, the spectrum of the temporary estimated speech calculated by the temporary estimated speech calculation unit 30 is denoted as S ^ (t, k) (k = 0,..., K−1). Further, the speech feature amount of the temporarily estimated speech is expressed as s ^ (t) (εR ^M , R ^M represents an M-dimensional real vector space). In the present embodiment, the temporary estimated speech calculation unit 30 uses a known technique (for example, spectral subtraction (SS), Wiener Filter (WF), etc.) for the spectrum of the temporary estimated speech. However, the method used by the present invention is not limited to these. The temporary estimated speech calculation unit 30 may calculate the spectrum of the temporary estimated speech using a predetermined method.

  The storage unit 40 stores a voice model obtained by modeling a voice feature amount. Specifically, as illustrated in FIG. 1, the storage unit 40 stores a mixed Gaussian distribution model GMM (Gaussian Mixture Model) 401 and a hidden Markov model HMM (Hidden Markov Model) 402.

  In this embodiment, the GMM and the HMM are different models, but a Gaussian distribution set related to the HMM can be handled as a GMM. Therefore, the storage unit 40 may store only the HMM.

  The GMM 401 is a speech model that does not have time transition information. This time transition information is created by using speech feature values extracted from previously collected speech data as learning data. Specifically, the GMM is composed of a plurality of Gaussian distributions. Each Gaussian distribution has a weight, a mean vector, and a variance matrix. Here, the time transition information refers to a constraint related to the temporal state transition of the model.

Hereinafter, the number of mixtures related to the GMM (the number of Gaussian distributions constituting the GMM) is N _GMM , the weight of the i-th Gaussian distribution is w _{i, GMM} , the mean vector is μ _{i, GMM} (∈R ^M ), and the variance matrix is Σ _{i, GMM} (εR ^{M × M} ) (i = 0,..., N _GMM− 1, εR ^{M × M} ).

  The speech data used for GMM learning is preferably subjected to a temporary estimated speech feature value calculation process similar to that of the temporary estimated speech calculation unit 30.

  The HMM 402 is a voice model having time transition information. Specifically, the HMM 402 has a state transition probability indicating a plurality of states and transition probabilities between the states. Each state has a GMM whose number of mixtures is a predetermined value. This state transition probability corresponds to the time transition information of the model.

Hereinafter, the number of states relating to the HMM 402 is denoted by S (state indices are denoted as s = 0,..., S-1), the number of mixtures relating to GMM in each state s is denoted by N _{s, HMM} , and the weight of the i-th Gaussian distribution. w _{i, s, HMM} , mean vector μ _{i, s, HMM} (∈R ^M ), variance matrix Σ _{i, s, HMM} (∈R ^{M × M} ) (i = 0,..., N _{s, HMM-} 1). Further, the state transition probability that the model transitions from the state p to the state q is denoted as a _qp .

・・・・・・（式１）The speech model suitability calculation unit 50 has a function of calculating the suitability of the temporarily estimated speech with respect to the speech model. Specifically, the speech model suitability calculating unit 50 acquires the feature amount s ^ (t) of the temporary estimated speech from the temporary estimated speech calculating unit 30. In addition, the speech model suitability calculation unit 50 acquires the GMM 401 stored in the storage unit 40. Then, the speech model suitability calculation unit 50 calculates the speech model suitability based on the acquired temporary estimated speech feature s ^ (t) and the GMM 401. The speech model suitability calculation unit 50 supplies the calculated likelihood to the expected value calculation unit 60.
The speech model fitness K (t) calculated by the speech model fitness calculator 50 can be expressed by Equation 1.

・ ・ ・ ・ ・ ・ (Formula 1)

・・・・・・（式２）Here, L (i | s ^ (t)) indicates the likelihood regarding the Gaussian distribution i when the feature quantity s ^ (t) of the temporary estimated speech is given, and can be expressed by Expression 2.

・ ・ ・ ・ ・ ・ (Formula 2)

・・・・・・（式３）
N (x; μ, Σ) can be expressed by Equation 3.

・ ・ ・ ・ ・ ・ (Formula 3)

・・・・・・（式４）In Equation 1, the maximum likelihood of the Gaussian distribution related to the GMM 401 is used as the speech model fitness, but the speech model fitness is not limited to this. For example, as shown in Equation 4, the speech model fitness is a sum of the likelihoods of the Gaussian distribution i up to the top N with respect to the GMM 401, and is a sum from the current time t to the time (t−T _K ). There may be.

・ ・ ・ ・ ・ ・ (Formula 4)

Here, U _N is a set of Gaussian distribution numbers corresponding to the top N when ranking with respect to i by the likelihood L (i | s ^ (t)) of the Gaussian distribution i.

  Further, in Formula 1, the speech model fitness is calculated from the likelihood based on the GMM 401, but may be calculated from the likelihood based on the HMM 402.

  The expected value calculation unit 60 has a function of calculating the expected value of the speech spectrum using the feature amount of the temporarily estimated speech, the speech model fitness, and the GMM 401 and HMM 402 stored in the storage unit 40. Specifically, the expected value calculation unit 60 acquires the feature amount s ^ (t) of the temporary estimated sound from the temporary estimated sound calculation unit 30. Further, the expected value calculation unit 60 acquires the speech model fitness K (t) from the speech model fitness calculation unit 50. Further, the expected value calculation unit 60 acquires the GMM 401 and the HMM 402 from the storage unit 40.

・・・・・・（式５）Then, the expected value calculation unit 60 calculates the expected value S _{E E} (t, k) (k) of the speech spectrum from the feature amount of the acquired temporary estimated speech, the acquired speech model fitness, and the acquired GMM 401 and HMM 402. = 0, ..., K-1). The expected value S ^ _E (t, k) of the speech spectrum is calculated by Equation 5.

・ ・ ・ ・ ・ ・ (Formula 5)

Here, Expression 5 calculates the speech spectrum expected value S ^ _E (t, k) using the GMM 401 or uses the HMM 402 according to the result of comparing the speech model fitness K (t) and the threshold th1. Indicates that it is to be calculated. In Equation 5, an expected speech spectrum value S ^ _GMM (t, k) is an expected speech spectrum value calculated using _GMM 401 described later. The expected speech spectrum value S ^ _HMM (t, k) is an expected speech spectrum value calculated using the _HMM 402 described later. An experimentally calculated value can be used for the threshold th1.

  In Equation 5, a model is selected for each frame t, but a model may be selected for each voice section.

・・・・・・（式６）Next, the expected speech spectrum value ＾ _GMM (t, k) calculated using _GMM 401 and the expected speech spectrum value ＾ _HMM (t, k) calculated using _HMM 402 in Equation 5 will be described.
The expected speech spectrum value S ^ _GMM (t, k) calculated using the _GMM 401 is calculated by Equation 6.

・ ・ ・ ・ ・ ・ (Formula 6)

Here, S _{μ, i, GMM} (k) (k = 0,..., K−1) is a value obtained by converting an average vector μ _{i, GMM} of a Gaussian distribution i related to _GMM into a spectrum.

・・・・・・（式７）Next, the expected speech spectrum value S ^ _HMM (t, k) calculated using the _HMM 402 is calculated by Equation 7. Based on the calculation method used in Patent Document 2, the calculation method of the expected speech spectrum value S ^ _HMM (t, k) will be described.

・ ・ ・ ・ ・ ・ (Formula 7)

・・・・・・(式８)

・・・・・・（式９）Here, S indicates the number of states, and S _{μ, i, s, HMM} (k) (k = 0,..., K−1) is an average vector μ _i, Gaussian distribution i in the state s of the HMM _{. This} is a value obtained by converting _{s and HMM} into a spectrum. Further, the weighting coefficient α (s, t) is recursively calculated by Expressions 8 and 9 using the state transition probability a _qp that is time transition information. Note that the second term on the right side of Equation 9 is a natural logarithm.

(Equation 8)

・ ・ ・ ・ ・ ・ (Formula 9)

  As shown in Equation 5, the expected value calculation unit 60 selects either a GMM having time transition information or an HMM having no time transition information as a speech model used when calculating an expected value of the speech spectrum. The expected value calculation unit 60 performs this selection based on the model suitability K (t), that is, the suitability of the temporary estimated speech with respect to the speech model. The expected value calculation unit 60 calculates the expected value of the speech spectrum using a model having time transition information when the degree of fitness for the speech model is equal to or greater than a threshold value. The expected value calculation unit 60 calculates the expected value of the speech spectrum using a model that does not have time transition information when the degree of fitness for the speech model is less than the threshold. The expected value calculation unit 60 supplies the calculated expected value of the speech spectrum to the filter calculation unit 70.

・・・・・・（式１０）The filter calculation unit 70 has a function of calculating a suppression filter that suppresses noise based on the estimated noise spectrum and the expected value of the speech spectrum. Specifically, the filter calculation unit 70 acquires the estimated noise spectrum N ^ (t, k) from the noise estimation unit 20. Further, the filter calculation unit 70 acquires the expected value S ^ _E (t, k) of the speech spectrum from the expected value calculation unit 60. The filter calculation unit 70 calculates a suppression filter based on the acquired spectrum of estimated noise and the expected value of the speech spectrum. Here, the suppression filter is expressed as W (t, k) (k = 0,..., K−1). The filter calculation unit 70 calculates the suppression filter W (t, k) using Expression 10.

・ ・ ・ ・ ・ ・ (Formula 10)

  Thereafter, the filter calculation unit 70 supplies the calculated suppression filter W (t, k) to the filtering unit 80.

・・・・・・（式１１）The filtering unit 80 has a function of outputting an estimated voice (output signal) by performing a filtering process using a suppression filter on the input signal. Specifically, the filtering unit 80 acquires the spectrum X (t, k) of the input signal from the input signal acquisition unit 10. Further, the filtering unit 80 acquires the suppression filter W (t, k) from the filter calculation unit 70. Then, the filtering unit 80 calculates the spectrum of the estimated speech output by the speech processing device 1 using the spectrum of the received input signal and the suppression filter. Here, the spectrum of the estimated speech is denoted as S _OUT (t, k) (k = 0,..., K−1). The filtering unit 80 calculates the spectrum of the estimated speech using Expression 11.

... (Formula 11)

  The filtering unit 80 outputs the spectrum of the estimated speech calculated using Expression 11. When the output target is a speech recognition device, the filtering unit 80 converts the spectrum of the estimated speech into a feature amount vector and outputs the feature amount vector of the estimated speech. In addition, when the output target is a sound reproduction device such as a speaker, the filtering unit 80 performs inverse Fourier transform on the spectrum of the estimated sound, thereby converting the spectrum into a time domain signal and outputs the digital signal.

  Next, the operation (processing) of the sound processing apparatus 1 of the present embodiment will be described in detail with reference to the flowchart of FIG.

  The input signal acquisition unit 10 acquires an input signal (step S101). The input signal acquisition unit 10 divides a digital signal generated based on the acquired input signal into frames for each unit time. The input signal acquisition unit 10 converts the cut frame into an absolute value of the spectrum and outputs it as an input signal spectrum (step S102). The noise estimation unit 20 estimates the spectrum of the noise component based on the input signal spectrum output from the input signal acquisition unit 10, and outputs the estimated estimated noise spectrum (step S103).

  The temporary estimated speech calculation unit 30 uses the input signal spectrum output from the input signal acquisition unit 10 and the estimated noise spectrum output from the noise estimation unit 20 to calculate the spectrum and speech feature amount of the temporary estimated speech. . The temporary estimated speech calculation unit 30 outputs the calculated temporary estimated speech feature amount (step S104).

  The speech model suitability calculation unit 50 calculates the speech model suitability using the speech feature amount of the temporary estimated speech output from the temporary estimated speech calculation unit 30 and the GMM 401 or HMM 402 stored in the storage unit 40. . The speech model fitness calculation unit 50 outputs the calculated speech model fitness to the expected value calculation unit 60 (step S105).

  When the speech model suitability is equal to or higher than the threshold (Yes in step S106), the expected value calculation unit 60 stores the speech feature amount of the temporary estimated speech output from the temporary estimated speech calculation unit 30 and the storage unit 40. The expected value of the voice spectrum is calculated using the HMM 402. The expected value calculation unit 60 outputs the calculated expected value of the speech spectrum to the filter calculation unit 70 (step S107).

  When the speech model fitness is less than the threshold value (No in step S106), the expected value calculation unit 60 stores the speech feature amount of the temporary estimated speech output from the temporary estimated speech calculation unit 30 and the storage unit 40. The expected value of the speech spectrum is calculated using the GMM 401. The expected value calculation unit 60 outputs the calculated expected value of the speech spectrum to the filter calculation unit 70 (step S108).

  The filter calculation unit 70 calculates a suppression filter using the estimated noise spectrum output from the noise estimation unit 20 and the expected value of the speech spectrum output from the expected value calculation unit 60. The filter calculation unit 70 outputs the calculated suppression filter to the filtering unit 80 (step S109). The filtering unit 80 calculates the estimated speech using the input signal spectrum output from the input signal acquisition unit 10 and the suppression filter output from the filter calculation unit 70 (step S110).

  The input signal acquisition unit 10 checks whether or not there is an input signal. If there is an input signal (Yes in step S111), the process returns to S101, and if there is no input signal (No in step S111), the entire The process ends.

  The speech processing apparatus 1 according to the present embodiment takes advantage of the effect obtained by selecting the distribution included in the speech model based on the time transition information, and the state transition related to the speech feature is continuously influenced by noise. Even in the case of receiving noise, high noise suppression accuracy can be obtained. The reason is that when the expected value calculation unit 60 calculates the expected value of the speech spectrum, the HMM 402 is used when the speech model suitability is greater than or equal to the threshold, and the GMM 401 when the speech model fit is less than the threshold. It is because it uses.

  The speech processing apparatus 1 according to the present embodiment uses the property that the degree of model suitability for the speech signal decreases as the proportion of the noise component included in the speech signal increases. That is, when calculating the expected value of the speech spectrum, the speech processing apparatus 1 selects a distribution included in the speech model based on the time transition information of the HMM by using the HMM 402 when the model suitability is high. The effect obtained by doing can be utilized. Then, the speech processing apparatus 1 uses the GMM 401 while avoiding the HMM 402 when the model conformity is low, so that even when the state transition related to the speech feature is strongly influenced by noise, high noise suppression is achieved. Accuracy can be obtained.

  In addition, although this embodiment employs a method similar to the model-based Wiener filtering method (Patent Document 1) as shown in Equation 5, for example, as a noise suppression method, another method may be used. . The noise suppression processed by the present embodiment may be model-based noise suppression, and may be performed, for example, in the form of a vector Taylor series expansion described in Non-Patent Document 1.

<Second Embodiment>
FIG. 3 is a block diagram conceptually showing the configuration of the speech processing apparatus 2 of the second embodiment. For convenience of explanation, components having the same functions as those of the components included in FIG. 1 described for the first embodiment are denoted by the same reference numerals as those of the first embodiment, and description thereof is omitted. .

The speech processing apparatus 2 according to the present embodiment uses a speech noise ratio “SNR (Signal Noise Ratio)” as a numerical value indicating the degree of noise burying in the input signal. When the noise processing is performed, the speech processing apparatus 2 uses a speech model (model A) that does not have time transition information when the SNR is low, and speech that has time transition information when the SNR is high. A model (Model B) is used.
As shown in FIG. 3, the configuration of the speech processing device 2 is the same as that of the speech processing device 1 except that the speech model suitability calculation unit 50 in the speech processing device 1 is replaced with an SNR calculation unit 90.

  Note that, unlike the first embodiment, the noise estimation unit 20 according to the present embodiment uses not only the temporary estimated speech calculation unit 30 and the filter calculation unit 70 but also the SNR calculation unit as the estimated noise component spectrum. 90 is also supplied. Further, the temporary estimated speech calculation unit 30 according to the present embodiment supplies the calculated spectrum of the temporary estimated speech to the SNR calculation unit 90. Then, the expected value calculation unit 60 according to the present embodiment acquires the SNR calculated by the SNR calculation unit 90.

  The SNR calculator 90 calculates the SNR based on the temporary estimated speech spectrum and the estimated noise spectrum. Specifically, the SNR calculator 90 acquires the spectrum S ^ (t, k) of the temporary estimated speech from the temporary estimated speech calculator 30. In addition, the SNR calculation unit 90 acquires the estimated noise spectrum N ^ (t, k) from the noise estimation unit 20. Then, the SNR calculation unit 90 supplies the calculated SNR to the expected value calculation unit 60.

・・・・・・（式１２）Hereinafter, when the SNR calculated by the SNR calculation unit 90 is SNR (t), the SNR (t) can be expressed by Expression 12.

・ ・ ・ ・ ・ ・ (Formula 12)

Here, T _avg is the number of frames used for the average value calculation. Note that the frames used for the average value calculation may be those belonging to the speech section obtained by speech detection. The SNR calculation method is not limited to Equation 12.

  The expected value calculation unit 60 calculates the expected value of the speech spectrum using the speech feature amount of the temporary estimated speech, the SNR, and the GMM 401 and HMM 402 stored in the storage unit 40. Specifically, the expected value calculation unit 60 acquires the feature amount s ^ (t) of the temporary estimated speech from the temporary estimated speech calculation unit 30. In addition, the expected value calculation unit 60 acquires SNR (t) from the SNR calculation unit 90. Further, the expected value calculation unit 60 acquires the GMM 401 and the HMM 402 from the storage unit 40.

・・・・・・（式１３）Then, the expected value calculation unit 60 calculates the expected value S ^ _E (t, k) (k = 0) of the speech spectrum from the speech feature amount of the acquired temporary estimated speech, the acquired SNR, and the acquired GMM 401 and HMM 402. , ..., K-1). The expected value S ^ _E (t, k) of the speech spectrum is calculated by Equation 13.

... (Formula 13)

Here, the expression 13 compares the SNR (t) and the threshold th2 to set the expected speech spectrum value S ^ _E (t, k) as the expected speech spectrum value S _GMM ^ (t, k). It indicates that it is determined whether or not the speech spectrum expected value S _HMM ^ (t, k) is to be used. Here, the expected speech spectrum value S _GMM ^ (t, k) is an expected value related to the speech spectrum calculated using the _GMM 401. The expected speech spectrum value S _HMM ^ (t, k) is an expected value related to the speech spectrum calculated using the _HMM 402. An experimentally calculated value can be used as the threshold th2.

  The speech processing apparatus 2 according to the present embodiment utilizes the effect obtained by selecting the distribution included in the speech model based on the time transition information, and the state transition related to the speech feature is continuously influenced by noise. Even in the case of receiving noise, high noise suppression accuracy can be obtained. The reason is that when the expected value calculation unit 60 calculates the expected value of the speech spectrum, the HMM 402 is used when the SNR is greater than or equal to the threshold value, and the GMM 401 is used when the speech model suitability is less than the threshold value. It is.

  The audio processing device 2 according to the present embodiment uses the property that the SNR relating to the audio signal decreases as the ratio of the noise component included in the audio signal increases. That is, when calculating the expected value of the speech spectrum, the speech processing device 2 uses the HMM 402 when the SNR is high, thereby selecting a distribution included in the speech model based on the time transition information of the HMM. The effect obtained by can be utilized. Then, when the SNR is low, the speech processing apparatus 2 avoids the HMM 402 and uses the GMM 401, so that even when the state transition related to the speech feature amount is continuously strongly influenced by the noise, high noise suppression accuracy can be obtained. Can be obtained.

<Third Embodiment>
FIG. 4 is a block diagram conceptually showing the configuration of the speech processing apparatus 3 of the third embodiment. For convenience of explanation, components having the same functions as those of the components included in FIG. 1 described for the first embodiment are denoted by the same reference numerals as those of the first embodiment, and description thereof is omitted. .

  The speech processing apparatus 3 according to the present embodiment uses the property that when the input signal is buried in noise, it easily adapts to a specific distribution included in the speech model, and the input signal is adapted to the specific distribution. Judgment based on whether or not to do. The voice processing device 3 uses a voice model (model A) that does not have time transition information when the input signal matches a specific distribution when performing noise suppression processing, and the input signal matches the specific distribution. If not, a speech model (model B) having time transition information is used.

  As shown in FIG. 4, the configuration of the speech processing device 3 is the same as that of the speech processing device 1 except that the speech model suitability calculation unit 50 is replaced with a designated distribution matching unit 91 and a designated distribution storage unit 41 is added. This is the same as the sound processing device 1.

  The designated distribution storage unit 41 stores a designated distribution. Specifically, the designated distribution storage unit 41 stores a distribution set P designated in advance among the Gaussian distributions included in the GMM 401 stored in the storage unit 40. Then, the designated distribution storage unit 41 supplies this distribution set P to the designated distribution matching unit 91. Here, the designated distribution is a distribution that easily matches an input signal buried in noise, and can be experimentally obtained in advance, for example.

  The designated distribution matching unit 91 determines whether or not a distribution that gives the maximum likelihood to the speech feature amount related to the temporarily estimated speech among the Gaussian distributions included in the GMM 401 is included in the designated distribution set P. Specifically, the designated distribution matching unit 91 acquires the feature amount s ^ (t) of the temporary estimated speech from the temporary estimated speech calculation unit 30. Also, the designated distribution matching unit 91 acquires the designated distribution set P from the designated distribution storage unit 41. Then, the designated distribution matching unit 91 obtains a distribution that gives the maximum likelihood to the feature quantity s ^ (t) of the temporary estimated speech from among the Gaussian distributions included in the GMM 401 and performs matching with the distribution set P. . The designated distribution matching unit 91 supplies the matching result to the expected value calculation unit 60.

・・・・・・（式１４）If the matching result of the designated distribution matching unit 91 is Q (t), the matching result Q (t) can be expressed by Expression 14.

・ ・ ・ ・ ・ ・ (Formula 14)

  Here, p ′ indicates a distribution that gives the maximum likelihood to the feature quantity s ^ (t) of the temporarily estimated speech among the Gaussian distributions included in the GMM 401.

  The expected value calculation unit 60 calculates the expected value of the speech spectrum using the feature amount of the temporarily estimated speech, the matching result, and the GMM 401 and HMM 402 stored in the storage unit 40. Specifically, the expected value calculation unit 60 acquires the feature amount s ^ (t) of the temporary estimated sound from the temporary estimated sound calculation unit 30. In addition, the expected value calculation unit 60 acquires Q (t) from the designated distribution matching unit 91. Further, the expected value calculation unit 60 acquires the GMM 401 and the HMM 402 from the storage unit 40.

・・・・・・（式１５）Then, the expected value calculation unit 60 calculates the expected value S ^ _E (t, k) (k = 0) of the speech spectrum from the feature amount of the acquired temporary estimated speech, the acquired matching result, and the acquired GMM 401 and HMM 402. , ..., K-1). The expected value S ^ _E (t, k) of the speech spectrum is calculated by the following equation 15.

・ ・ ・ ・ ・ ・ (Formula 15)

Here, in Expression 15, depending on whether or not the value indicated by Q (t) is 1, the expected speech spectrum value S ^ _E (t, k) is changed to the expected speech spectrum value S _GMM ^ (t, k). ) Or the expected speech spectrum value S _HMM ^ (t, k). Here, the expected speech spectrum value S _GMM ^ (t, k) is an expected value related to the speech spectrum calculated using the _GMM 401. The expected speech spectrum value S _HMM ^ (t, k) is an expected value related to the speech spectrum calculated using the _HMM 402.

  In the present embodiment, the example in which the speech processing device 3 matches the distribution set P with the maximum likelihood included in the GMM 401 calculated from the temporary estimated speech is shown. The speech processing device 3 may take a match between the state relating to the HMM 402 calculated from the temporary estimated speech and the specified state set. Alternatively, the voice processing device 3 may take a matching between the state series related to the HMM 402 calculated from the temporary estimated voice and the designated state series set.

  The speech processing apparatus 3 according to the present embodiment utilizes the effect obtained by selecting the distribution included in the speech model based on the time transition information, and the state transition related to the speech feature is continuously influenced by noise. Even in the case of receiving noise, high noise suppression accuracy can be obtained. The reason is that when the expected value calculation unit 60 calculates the expected value of the speech spectrum, the HMM 402 is used when the input signal does not match the specific distribution, and the GMM 401 when the input signal matches the specific distribution. It is because it uses.

  The sound processing device 3 according to the present embodiment uses the property that the sound signal is likely to be adapted to a specific distribution included in the sound model as the ratio of the noise component included in the sound signal increases. That is, when calculating the expected value of the speech spectrum, the speech processing device 3 uses the HMM 402 when the input signal does not match a specific distribution, thereby converting the speech model into a speech model based on the time transition information of the HMM. The effect obtained by selecting the included distribution can be utilized. When the input signal matches a specific distribution, the speech processing apparatus 3 avoids the HMM 402 and uses the GMM 401 so that the state transition related to the speech feature amount is strongly influenced by noise. However, high noise suppression accuracy can be obtained.

<Fourth Embodiment>
FIG. 5 is a block diagram conceptually showing the structure of the speech processing apparatus 4 of the fourth embodiment.

  The speech processing apparatus 4 according to the present embodiment includes a storage unit 40, an expected value calculation unit 60, and a noise suppression unit 81.

  The memory | storage part 40 memorize | stores the 1st audio | voice model which has the time transition information which shows the performance of the state transition regarding an audio | voice feature-value, and the 2nd audio | voice model which does not have the time transition information.

  The expected value calculation unit 60 selects one of the first and second speech models from information indicated by the input signal, which is a signal in which speech components and noise components are mixed, according to a predetermined criterion. The expected value calculation unit 60 calculates an expected value related to the speech component using the selected speech model and input signal.

  The noise suppression unit 81 uses the expected value calculated by the expected value calculation unit 60 to generate noise-suppressed speech in which the noise component included in the input signal is suppressed.

  The speech processing apparatus 4 according to the present embodiment makes use of the effect obtained by selecting the distribution included in the speech model based on the time transition information, and the state transition related to the speech feature is continuously affected by noise. Even in the case of receiving noise, high noise suppression accuracy can be obtained. The reason is that when the expected value calculation unit 60 calculates the expected value of the speech spectrum, based on the information indicated by the input signal, it selects one of the first and second speech models according to a predetermined criterion, This is because the selected speech model is used.

<Hardware configuration example>
1 and 3 to 5 in the above-described embodiment can be realized by a dedicated HW (electronic circuit). Each unit (excluding the storage unit 40) can be regarded as a function (processing) unit (software module) of the software program. However, the division of each part shown in these drawings is a configuration for convenience of explanation, and various configurations can be assumed for mounting. An example of the hardware environment in this case will be described with reference to FIG.

  FIG. 6 is a diagram for exemplarily explaining the configuration of an information processing apparatus 900 (computer) that can execute the speech processing apparatus according to each exemplary embodiment of the present invention. That is, FIG. 6 shows a configuration of a computer (information processing apparatus) capable of realizing the sound processing apparatus shown in FIGS. 1 and 3 to 5, and hardware capable of realizing each function in the above-described embodiment. Represents the environment.

The information processing apparatus 900 illustrated in FIG. 6 includes the following as constituent elements.
CPU 901 (Central_Processing_Unit),
ROM 902 (Read_Only_Memory),
RAM 903 (Random_Access_Memory),
-Hard disk 904 (storage device),
A communication interface 905 with an external device,
A reader / writer 908 capable of reading and writing data stored in a storage medium 907 such as a CD-ROM (Compact_Disc_Read_Only_Memory)
-I / O interface 909,
The information processing apparatus 900 is a general computer in which these configurations are connected via a bus 906 (communication line).

  The present invention described using the above-described embodiment as an example supplies a computer program capable of realizing the following functions to the information processing apparatus 900 illustrated in FIG. 6. The function is a function of a block configuration diagram (FIGS. 1 and 3 to 5) or a flowchart (FIG. 2) referred to in the description of the embodiment. The present invention is then achieved by reading the computer program into the hardware CPU 901 for interpretation and execution. The computer program supplied to the apparatus may be stored in a readable / writable volatile storage memory (RAM 903) or a nonvolatile storage device such as the hard disk 904.

  In the above case, a general procedure can be adopted as a method for supplying the computer program into the hardware. The procedure includes, for example, a method of installing in the apparatus via various storage media 907 such as a CD-ROM, and a method of downloading from the outside via a communication line such as the Internet. In such a case, it can be understood that the present invention is configured by a code constituting the computer program or a storage medium 907 in which the code is stored.

  The present invention has been described above using the above-described embodiment as an exemplary example. However, the present invention is not limited to the above-described embodiment. That is, the present invention can apply various modes that can be understood by those skilled in the art within the scope of the present invention.

  This application claims the priority on the basis of Japanese application Japanese Patent Application No. 2013-259846 for which it applied on December 17, 2013, and takes in those the indications of all here.

DESCRIPTION OF SYMBOLS 1 Audio processing apparatus 2 Audio processing apparatus 3 Audio processing apparatus 4 Audio processing apparatus 10 Input signal acquisition part 20 Noise estimation part 30 Temporary estimation audio | voice calculation part 40 Memory | storage part 401 GMM
402 HMM
DESCRIPTION OF SYMBOLS 41 Designated distribution memory | storage part 50 Speech model adaptation calculation part 60 Expected value calculation part 70 Filter calculation part 80 Filtering part 81 Noise suppression part 90 SNR calculation part 91 Designated distribution matching part 900 Information processing apparatus 901 CPU
902 ROM
903 RAM
904 Hard disk 905 Communication interface 906 Bus 907 Storage medium 908 Reader / writer 909 Input / output interface

Claims (10)

Storage means for storing a first speech model having time transition information indicating a result of state transition relating to a speech feature, and a second speech model not having the time transition information;
From the information indicated by the input signal, which is a signal in which a speech component and a noise component are mixed, one of the first and second speech models is selected according to a predetermined criterion, and the selected speech model and the input signal are used. , An expected value calculating means for calculating an expected value for the audio component;
Noise suppression means for generating a first noise-suppressed speech in which the noise component included in the input signal is suppressed using the expected value;
A speech processing apparatus comprising: A temporary estimated speech calculation means for generating a signal related to a second noise-suppressed speech in which the noise component contained in the input signal is suppressed by a predetermined method;
The expected value calculation means selects one of the first and second speech models from information represented by a signal related to the second noise-suppressed speech.
The speech processing apparatus according to claim 1. Voice model suitability calculating means for calculating a model suitability indicating the degree to which the second noise-suppressed speech is adapted to the first or second speech model;
The expected value calculation means selects the first speech model when the value indicated by the speech model suitability is greater than or equal to a threshold, and when the value indicated by the speech model suitability is less than the threshold, Select the second voice model,
The speech processing apparatus according to claim 2. For the second noise-suppressed speech, further comprising speech noise ratio calculating means for calculating a speech noise ratio indicating a ratio of the speech component to the noise component,
The expected value calculation means selects the first speech model when the value indicated by the speech noise ratio is equal to or greater than a threshold value, and selects the second speech model when the value indicated by the speech noise ratio is less than the threshold value. Select the voice model for
The speech processing apparatus according to claim 2 or 3. Designated distribution storage means for storing a predetermined distribution set included in the first or second speech model;
Designated distribution matching that performs pattern matching between the third speech model related to the second noise-suppressed speech and the predetermined distribution set, and determines whether or not the third speech model is included in the predetermined distribution set Means,
Further comprising
The expected value calculation means selects the second speech model when the third speech model is included in the predetermined distribution set, and the third speech model is included in the predetermined distribution set. If not, select the first speech model,
The speech processing apparatus according to claim 2. The specified distribution matching means uses a mixed Gaussian distribution model as the third speech model, and with respect to the speech feature amount related to the second noise-suppressed speech among a plurality of Gaussian distributions included in the mixed Gaussian distribution model. Pattern matching between the Gaussian distribution giving the maximum likelihood and the predetermined distribution set,
The speech processing apparatus according to claim 5. Depending on the information processing device,
By referring to the storage means in which the first speech model having time transition information indicating the results of state transition relating to the speech feature amount and the second speech model not having the time transition information are stored, the speech From information indicated by an input signal that is a signal in which a component and a noise component are mixed, either the first or second speech model is selected according to a predetermined criterion, and the selected speech model and the input signal are used. Calculating an expected value for the audio component;
Using the expected value, a first noise-suppressed speech in which the noise component included in the input signal is suppressed is generated.
Audio processing method. Generating a signal relating to a second noise-suppressed speech in which the noise component contained in the input signal is suppressed by a predetermined method;
From the information represented by the signal related to the second noise-suppressed speech, select one of the first and second speech models.
The voice processing method according to claim 7. By referring to the storage means in which the first speech model having time transition information indicating the results of state transition relating to the speech feature amount and the second speech model not having the time transition information are stored, the speech From information indicated by an input signal that is a signal in which a component and a noise component are mixed, either the first or second speech model is selected according to a predetermined criterion, and the selected speech model and the input signal are used. An expected value calculation process for calculating an expected value related to the audio component;
Using the expected value, a noise suppression process for generating a first noise-suppressed speech in which the noise component included in the input signal is suppressed;
A computer-readable recording medium in which a sound processing program for causing a computer to execute is stored. A temporary estimated speech calculation process for generating a signal relating to a second noise-suppressed speech in which the noise component contained in the input signal is suppressed by a predetermined method;
From the information represented by the signal relating to the second noise-suppressed speech, the expected value calculation process for selecting one of the first and second speech models;
A computer-readable recording medium in which the sound processing program according to claim 9 is stored.

Images