JP5609182B2 - Speech recognition apparatus, speech recognition method, and speech recognition program - Google Patents

Description

  The present invention relates to a speech recognition device that recognizes speech from an input signal, a speech recognition method, and a speech recognition program.

  The performance of a speech recognition system is significantly degraded due to the effects of noise. For this reason, in actual operation, a noise proofing method is required to enable a desired voice to be recognized even if noise is mixed. The cause of performance degradation due to noise is due to a mismatch between audio data used during acoustic model learning and an input signal used in actual operation. In order to suppress such mismatches, there are roughly two methods for noise resistance for speech recognition.

  One is a method (hereinafter referred to as a noise suppression method) in which noise components are suppressed / removed from an input signal, and the input signal side is brought close to voice data used during acoustic model learning. The other is a method of adapting the acoustic model side to the same noise environment as the input signal (hereinafter referred to as acoustic model adaptation method).

  As a noise suppression method, a spectral subtraction method (hereinafter referred to as SS method) is widely used (see, for example, Patent Document 1 and Non-Patent Document 1). The SS method is one of noise suppression methods in the frequency domain, and the noise contained in the input signal is subtracted from the power spectrum of the separately estimated noise from the power spectrum of the speech signal (input signal) containing noise. It is a way to suppress. Patent Document 1 describes a method of changing a suppression coefficient, which is a parameter for controlling a noise suppression amount, according to an SNR (signal-noise ratio: SN ratio). Further, as described in Non-Patent Document 2, a technique for further improving the accuracy of noise estimation has been studied.

  On the other hand, as an acoustic model adaptation method, an HMM synthesis method (see, for example, Non-Patent Document 1), a Jacobian method, a Vector Tallyer Series (VTS) method (see, for example, Non-Patent Document 3), and the like are known. .

  The HMM synthesis method synthesizes an HMM generated using clean speech that does not contain noise in advance (hereinafter referred to as a clean HMM) and an HMM generated using estimated noise, and is an object. This is a method for generating an HMM suitable for speech uttered in a noisy environment. One specific method of such an HMM synthesis method is, for example, a Parallel Model Combination method (hereinafter referred to as PMC method). In the PMC method, feature quantities are inversely converted into spectral domain quantities, and two HMMs are synthesized in the spectral domain.

  Further, the Jacobi method and the VTS method are methods that approximate how each distribution constituting the clean HMM changes in a noise environment by estimated noise by a linear expression.

JP 2000-330597 A

Hiroshi Matsumoto, “Speech Recognition Techniques for Noise Environments”, 2nd Information Science and Technology Forum (FIT2003), September 2003, p. 1-4 Masanori Kato, et al. , “Noise Suppression with High Speed Quality Base on Weighted Noise Estimation and MMSE STSA”, Proc IWAENC 2001, September 2001, p. 183-186 A. Acero, et al. , “HMM adaptation using vector Taylor series for noise spec recognition”, int. Conf. Spoken Language Processing (2000), 2000

  However, in the acoustic model adaptation method such as the HMM synthesis method, the Jacobi method, and the VTS method described above, and the noise suppression method such as the SS method, speech recognition is performed with high accuracy by following the changing noise environment. There is a problem that can not be done.

  For example, in the acoustic model adaptation method, there is a possibility that the calculation time required for noise adaptation of the acoustic model may not catch up with the noise change time, and in this case, there is a problem that the speech recognition accuracy deteriorates. In general, as the size of an acoustic model increases, the amount of calculation for adapting the acoustic model to noise also increases. For this reason, in a situation where the calculation time is longer than the noise change time, the acoustic model cannot be adapted to the changing noise environment, and the speech recognition accuracy deteriorates.

  In addition, a noise suppression method that suppresses estimated noise using calculation in the frequency domain, such as the SS method, has a problem of flooring processing, for example. In general speech recognition, the input signal is converted into a feature value such as a cepstrum, and the distance between the feature value and the probability density function for each phoneme included in the acoustic model is compared, and the input signal is supported. Search for the word string you want. In a logarithmic calculation performed when converting to a feature amount such as a cepstrum, a noise-suppressed signal that is used as an input signal may have an adverse effect.

  The following equation (1) represents a signal in which noise is suppressed in the SS method (that is, an audio signal output as a suppression result). Equation (1) is defined for each frequency band or subband.

X = max [Y−N, α] (1)

  In Expression (1), X represents a power spectrum of speech, Y represents a power spectrum of an input signal, and N represents a power spectrum of noise estimated for the input signal. Further, max [A, B] indicates an operation that takes a larger value of A or B. Α represents a flooring coefficient.

  For example, in a low SNR environment or a section where no voice is uttered, the power spectrum Y of the input signal may be substantially equal to or smaller than the estimated power spectrum N of noise. Under these conditions, if the result of the subtraction operation is simply the power spectrum of the noise-suppressed signal, the value obtained by the subsequent logarithmic operation becomes very unstable or the value itself cannot be obtained. End up. In the SS method, in order to prevent such a situation, a value after subtraction performed in the frequency domain in order to suppress (remove) noise is negative by introducing a flooring coefficient α as in Equation (1). Value and values near 0 are prevented.

  However, in a changing noise environment, it is difficult to set an optimum value for the flooring coefficient α depending on the noise type and SNR. If the flooring coefficient α is too large, the recognition accuracy deteriorates. On the other hand, if the flooring coefficient α is too small, the flooring effect is lost and speech recognition is performed with an unstable value or the speech recognition process results in an error. Problem arises. Even when a technique for improving the accuracy of estimated noise as described in Non-Patent Document 2 is used, the power spectrum Y of the input signal is almost equal to the estimated power spectrum N of noise. In situations where they are equal or smaller, the flooring process will be performed with a small error, so the above problem still exists.

  Accordingly, an object of the present invention is to provide a speech recognition apparatus, a speech recognition method, and a speech recognition program that can perform speech recognition with high accuracy even in a changing noise environment.

  The speech recognition apparatus according to the present invention includes a noise statistic calculating unit that calculates a noise statistic from noise data estimated for a plurality of frames of an input signal, and a noise statistic calculated by the noise statistic calculating unit. An acoustic model using a short-time fluctuation noise component calculating means for calculating a short-time fluctuation component of noise included in each frame of the input signal based on the amount and a noise statistic calculated by the noise statistic calculation means; Acoustic model adaptation means for adapting to noise, noise suppression means for suppressing the short-time fluctuation component of noise included in the frame calculated by the short-time fluctuation noise component calculation means for each frame of the input signal, Speech recognition means for performing speech recognition on the input signal suppressed by the noise suppression means using an acoustic model noise-adapted by the acoustic model adaptation means And wherein the door.

  The speech recognition method according to the present invention calculates a noise statistic from noise data estimated for a plurality of frames of an input signal, and adapts an acoustic model to the noise using the calculated noise statistic. For each frame of the input signal, the short-term fluctuation component of the noise calculated based on the noise statistics used in the noise-adapted acoustic model is suppressed, and the short-term fluctuation component of the noise is suppressed. Speech recognition is performed on the signal using a noise-adapted acoustic model.

Speech recognition program according to the present invention, the computer, from the noise of the data estimated for a plurality of frames of the input signal, and a noise statistic calculation processing for calculating the noise statistics have been calculated by the noise statistic calculation processing Based on the noise statistic, the short-time fluctuation noise component calculation process that calculates the short-time fluctuation component of the noise included in each frame of the input signal, and the noise statistic calculated by the noise statistic calculation process , An acoustic model adaptation process for adapting the acoustic model to noise, and noise suppression for suppressing the short-time fluctuation component of the noise included in the frame calculated by the short-time fluctuation noise component calculation process for each frame of the input signal and processing the input signal is suppressed by the noise suppressing process, speech recognition is conducted using noise adapted acoustic models by the acoustic model adaptation process Characterized in that to execute the voice recognition process.

  According to the present invention, voice recognition can be performed with high accuracy even in a changing noise environment.

It is a block diagram which shows the structural example of the speech recognition apparatus of 1st Embodiment. It is a flowchart which shows an example of the whole operation | movement of the speech recognition apparatus of 1st Embodiment. It is a flowchart which shows an example of the processing flow of a voice suppression / voice recognition process. It is a flowchart which shows an example of the processing flow of an acoustic model noise adaptation process. It is explanatory drawing which shows an example of the time change of the power spectrum N_S of noise, the average power spectrum N_L of the noise component calculated from the noise, and the power spectrum N_S of the short-time fluctuation component of the noise. It is explanatory drawing which shows the other example of the time change of the power spectrum N_N of the noise, the average power spectrum N_L of the noise component calculated from the noise, and the power spectrum N_S of the short-time fluctuation component of the noise. It is explanatory drawing which shows the example of the operation | movement timing of the acoustic model noise adaptation process and the noise suppression process with respect to this input signal. It is explanatory drawing which shows the example of the operation | movement timing of the acoustic model noise adaptation process and the noise suppression process with respect to this input signal. It is a block diagram which shows the structural example of the speech recognition apparatus of 2nd Embodiment. It is a flowchart which shows the operation example of the speech recognition apparatus of 2nd Embodiment. It is explanatory drawing which shows the example of the operation timing of the acoustic model noise adaptation process in 2nd Embodiment, and the noise suppression process with respect to this input signal. It is explanatory drawing which shows the example of the operation timing of the acoustic model noise adaptation process in 2nd Embodiment, and the noise suppression process with respect to this input signal. It is a block diagram which shows the structural example of the speech recognition apparatus of 3rd Embodiment. It is a flowchart which shows an example of the processing flow of the acoustic model noise adaptation process of 3rd Embodiment. It is a block diagram which shows the outline | summary of this invention.

Embodiment 1. FIG.
Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram illustrating a configuration example of the speech recognition apparatus according to the first embodiment of this invention. The speech recognition apparatus shown in FIG. 1 includes a speech recognition unit 100 and an acoustic model adaptation unit 200.

  The speech recognition unit 100 also includes an input signal acquisition unit 101, a noise estimation unit 102, a short-time fluctuation noise component calculation unit 103, a noise suppression unit 104, a search unit 105, and an acoustic model storage unit 106. . The acoustic model adaptation unit 200 includes an estimated noise holding unit 201, a noise statistic calculation unit 202, and an acoustic model adaptation unit 203.

  The input signal acquisition unit 101 inputs a signal collected using a microphone or the like, and extracts and acquires a time series of the input signal for each frame. The noise estimation unit 102 estimates noise from the time series of the input signal acquired by the input signal acquisition unit 101.

  The short-time variation noise component calculation unit 103 calculates a short-time variation component of noise based on the noise estimated by the noise estimation unit 102 and the noise statistic calculated by the noise statistic calculation unit 202 described later. To do. Hereinafter, the short-time fluctuation component of noise may be expressed as a short-time fluctuation noise component.

  Here, the noise statistic is, for example, the average or variance of the noise, and is an amount that does not change over a relatively long time from the time of adaptation of the acoustic model to the time of adaptation of the next acoustic model. Assuming that the calculation time necessary for noise adaptation of the acoustic model is T, it is set to an amount that does not change with respect to a time equal to or longer than T. Therefore, for example, an average or variance obtained from noise data in a predetermined time interval equal to or longer than T time may be used. Therefore, for example, an average or variance obtained from noise data in a predetermined time interval equal to or longer than T time may be used. Note that the mean and variance of noise are not limited to the mean and variance of noise components such as power spectrum, but the mean and variance of noise in the feature region, as well as first-order differentiation, second-order differentiation,. Including mean and variance of. An example of the feature amount is cepstrum. In addition to the cepstrum, various feature quantities for speech recognition such as the primary difference component, the secondary difference component, and the pitch frequency value can be used.

  On the other hand, the short-time fluctuation component of noise is a component that conceptually changes in a relatively short time of noise (in this case, a time interval shorter than the time interval for obtaining the statistical amount of noise). Specifically, it means the remaining noise component obtained by subtracting the average, which is one of the noise statistics, from the noise. Hereinafter, the expression “average of noise components” means the average of noise components as one of the statistical quantities of noise used for obtaining the short-time fluctuation component.

  In the present embodiment, the average of the noise components used when the short-time fluctuation noise component calculation unit 103 calculates the short-time fluctuation noise component is used during noise adaptation of the acoustic model applied to the input signal. Use noise statistics. For example, when the noise statistic calculation unit 202 stores an acoustic model adapted to noise in the acoustic model storage unit 106 described later, the noise statistic calculation unit 202 associates and stores information on the noise statistic adapted to the acoustic model, The statistics information may be output in response to a noise statistics acquisition request from the short-time fluctuation noise component calculation unit 103. In this way, the noise environment of noise to be suppressed by the input signal is synchronized with the noise environment of noise adapted to the acoustic model.

  The noise suppression unit 104 performs noise suppression processing on the time series of the input signal to suppress the short-time fluctuation noise component calculated by the short-time fluctuation noise component calculation unit 103.

  The search means 105 performs speech recognition on the input signal subjected to noise suppression processing by the noise suppression means 104 using an acoustic model stored in an acoustic model storage means 106 described later. The search means 105 extracts, for example, a feature value from the time series of the noise-suppressed input signal, compares the distance between the feature value of the input signal and the probability density function for each phoneme included in the acoustic model, A word string corresponding to the input signal may be searched, and the search result may be output as a speech recognition result of the input signal.

  The acoustic model storage unit 106 stores the acoustic model that has been subjected to noise adaptation in the acoustic model adaptation unit 200. In the present embodiment, the acoustic model storage unit 106 stores an acoustic model that has been subjected to noise adaptation based on noise statistics by an acoustic model adaptation unit 203 described later. The acoustic model storage unit 106 may store, for example, parameter values that define the acoustic model as acoustic model information. For example, a parameter value of an acoustic model such as an HMM in which a change in speech is represented by a probability density function of a feature amount for each phoneme is stored. The acoustic model storage means 106 may store information on the statistical amount of noise adapted to the acoustic model, along with the information on the acoustic model. For example, an average value of noise components of noise statistics or a value indicating noise statistics held separately may be stored in association with the acoustic model.

  The estimated noise holding means 201 holds noise data estimated by the noise estimating means 102. It is assumed that the estimated noise holding unit 201 can sequentially hold at least noise data necessary for the noise statistic calculating unit 202 to obtain a noise statistic.

  The noise statistic calculation unit 202 calculates a noise statistic using the noise data held in the estimated noise holding unit 201.

  The acoustic model adaptation unit 203 uses the noise statistic calculated by the noise statistic calculation unit 202 to perform noise adaptation on the acoustic model. The acoustic model adaptation means 205 generates, for example, the average and variance of noise components shown as noise statistics, and the average and variance values of noise calculated in the feature quantity region using speech that does not contain noise. It may be added to the value of the acoustic model parameter. In this way, the acoustic model is adapted to the noise environment indicated by the noise statistics. As already described, the acoustic model subjected to noise adaptation by the acoustic model adaptation unit 203 is stored in the acoustic model storage unit 106 and used by the search unit 105. When the acoustic model adaptation unit 203 stores the acoustic model information in the acoustic model storage unit 106, the acoustic model adaptation unit 203 may also store information on noise statistics adapted to the acoustic model.

  In the present embodiment, the input signal acquisition unit 101 is realized by a sound collection device such as a microphone and a CPU or the like that operates according to a program. The noise estimation unit 102, the short-time fluctuation noise component calculation unit 103, the noise suppression unit 104, the search unit 105, the noise statistic calculation unit 202, and the acoustic model adaptation unit 203 are realized by a CPU or the like that operates according to a program, for example. . The acoustic model storage unit 106 and the estimated noise holding unit 201 are realized by a storage device, for example.

  In this example, the speech recognition unit 100 includes the acoustic model storage unit 106, and the acoustic model adaptation unit 200 includes the estimated noise holding unit 201. However, the storage unit includes the speech recognition unit 100 and the acoustic model. It is also possible to provide the storage means shared with the adaptation unit 200 independently inside or outside the speech recognition apparatus. In addition, the CPU included in the speech recognition apparatus implements a multi-thread environment, for example, so that each unit configuring the speech recognition unit 100 and each unit configuring the acoustic model adaptation unit 200 operate as separate threads. It is more preferable. In the speech recognition unit 100, the input signal acquisition unit 101, the noise estimation unit 102, and other units may be operated as separate threads.

  Next, the operation of this embodiment will be described. FIG. 2 is a flowchart showing an example of the overall operation of the speech recognition apparatus of this embodiment. Note that the example shown in FIG. 2 operates by receiving an event notifying that a signal (sound data) has been input, for example. In addition, the processing for the input signal is performed for each frame.

  As shown in FIG. 2, first, in the speech recognition unit 100, the input signal acquisition unit 101 cuts out and acquires collected time-series input sound data for each frame of unit time (step S101). For example, when the input sound data is 16-bit liner-PCM data with a sampling frequency of 8000 Hz, waveform data for 8000 points per second is input. For example, the input signal acquisition unit 101 sequentially cuts out such waveform data along a time series with a frame width of 200 points (25 milliseconds) and a frame shift of 80 points (10 milliseconds), and a predetermined storage area. May hold data. At that time, short-time discrete Fourier transform (FFT) may be performed on the extracted waveform data to convert it into a power spectrum.

  Next, the noise estimation unit 103 estimates the noise included in the input signal from the input signal acquired by the input signal acquisition unit 101, and holds the estimated noise data in the estimated noise holding unit 201 (step S102). The noise estimation method may be, for example, a method using an average value of input signals in a non-speech section before the target speech is uttered, or a method described in Non-Patent Document 2 (hereinafter referred to as weighted noise estimation). May be used). The weighted noise estimation method is a method mainly composed of three types of processes of SNR estimation, weight coefficient calculation, and averaging, and weights an input signal with a weight coefficient calculated based on the estimated SNR, In this method, estimated noise is estimated based on an average value of weighted input signals (see Non-Patent Document 2).

  Here, if it is the adaptation timing of the acoustic model (Yes in step S103), the process proceeds to step S106, where the acoustic model adaptation unit 200 performs noise adaptation processing for the acoustic model. In the following description, it is assumed that the acoustic model used for the input signal has already been noise-adapted and is not at the adaptation timing of the acoustic model at this time.

  When it is not the adaptation timing of the acoustic model (No in step S103) or when the noise adaptation processing of the acoustic model used for the input signal is completed, the speech recognition unit 100 performs noise suppression / speech recognition processing (step S104). FIG. 2 shows an example in which all input signals are subjected to speech recognition processing. However, for example, the subsequent noise suppression / speech recognition processing may be performed only when there is an instruction for speech recognition processing. Good. Further, for example, when calculating noise statistics using all noise data within a predetermined time interval including the input signal, noise statistics are calculated, and noise adaptation processing of the acoustic model is performed based on the noise statistics. Until the completion, the noise suppression / speech recognition process (step S104) is made to stand by, and noise estimation is performed in advance as much as necessary, so that after the acoustic model is noise-adapted, the subsequent noise suppression / speech is performed. Recognition processing may be performed.

  FIG. 3 is a flowchart showing an example of a processing flow of the voice suppression / voice recognition process (step S104 in FIG. 2). As shown in FIG. 3, in the speech suppression / recognition processing, first, the short-time fluctuation noise component calculation unit 103 acquires the noise statistic included in the input signal, and the acquired noise statistic (more specifically, Is the average of the noise components) and the noise estimated by the noise estimation means 103 is calculated for the short-time fluctuation noise component for the input signal (step S111 in FIG. 3). The noise statistic included in the input signal is a noise statistic adapted to the acoustic model used for speech recognition of the input signal, and is a value acquired from the noise statistic calculation unit 202. Use it. Here, the statistical quantity of the estimated noise of the input signal in the time interval including the input signal or the time interval before the input signal, which is calculated by the noise statistic calculation unit 202, is acquired.

  In step S111, for example, if the noise power spectrum estimated by the noise estimation unit 103 is N and the average power spectrum of the noise component included in the input signal calculated by the noise statistic calculation unit 202 is N_L, the input The power spectrum N_S of the short-time fluctuation noise component of the signal is calculated by the following equation (2). Note that the short-time fluctuation noise component calculation unit 103 performs, for example, the calculation of Expression (2) for each predetermined unit defined in the frequency domain such as each frequency band or each subband.

N_S = N−N_L (2)

  Next, the noise suppression unit 104 suppresses the short-time fluctuation noise component calculated by the short-time fluctuation noise component calculation unit 103 from the time series of the input signal to be processed (step S112). As a suppression method, for example, a method such as a spectral subtraction method (SS method) or a Wiener filter method may be used. For example, the following formula (3) is an example of suppression using the SS method.

X = max [Y−N_S, α] (3)

  It should be noted that YN_S (input signal in which the short-time fluctuation component of noise is suppressed) in Expression (3) does not remove all of the noise, but theoretically indicates the noise component indicated by the noise statistic. The component corresponding to the average remains.

  Also, the following equations (4) and (5) represent input signals that are noise-suppressed by the Wiener filter method. Noise suppression may be performed using such equations (5) and (6).

X = D_t / (D_t + N_S) * Y Expression (4)
D_t = γ * D_ (t−1) + (1−γ) * max [Y−N_S, α] (5)

  In equations (4) and (5), t represents a frame number. Α represents a flooring coefficient. Further, γ> 0.9.

  The noise suppression unit 104 performs, for example, the calculation of the above formula (3) or the above formula (4) and the formula (5) for each predetermined unit defined in the frequency domain such as each frequency band or each subband.

  Next, the search unit 105 extracts a feature amount used for speech recognition from the signal (noise suppression signal X) subjected to noise suppression by the noise suppression unit 104 (step S113). Next, the distance between the extracted feature quantity and the probability density function for each phoneme included in the acoustic model is compared, and a word string corresponding to the input signal is searched (step S114). An example of the acoustic model is HMM.

  In the noise suppression / speech recognition process of this example, the processes in steps S111 to S114 are performed on all input signals that have been noise-estimated and that are the targets of the speech recognition process. When there is no unprocessed input signal, the process returns to the caller (No in step S115, proceeds to step S105 in FIG. 2).

  In step S105, it is determined whether or not the signal input has ended. If the signal input has ended (Yes in step S107), the series of processing ends. If the signal input has not ended, the same processing is performed again from the input signal acquisition processing (return to step S101).

  In addition, when the acoustic model adaptation timing comes while performing the speech recognition process on the input signal in this way (Yes in step S103), the acoustic model adaptation unit 200 performs the acoustic model noise adaptation process (step S106) may be operated. As the adaptation timing of the acoustic model, for example, it can be considered that it operates several seconds after sound collection by a microphone is started, and thereafter operates at regular intervals. Note that the acoustic model noise adaptation process may be performed at least once as an initial operation before performing the speech recognition process on the input signal, regardless of whether the adaptation timing of the acoustic model has come.

  FIG. 4 is a flowchart showing an example of the processing flow of the acoustic model noise adaptation processing (step S106 in FIG. 2). As shown in FIG. 4, in the acoustic model noise adaptation process, first, the noise statistic calculation unit 202 calculates noise statistic using the noise data held in the estimated noise holding unit 201 (step S121). ). For example, the noise statistic calculation unit 202 may calculate a noise statistic from the latest K pieces of noise data held in the estimated noise holding unit 201.

  The noise statistic calculation unit 202 may calculate, for example, an average value of K power spectra as the noise statistic. In addition to this, the variance may be further calculated based on the average value, or the average value and variance of the N-th derivative may be calculated. Alternatively, the noise component may be converted into a feature quantity such as a cepstrum, and an average value or variance in the feature quantity region, or an average value or variance of the Nth order derivative may be calculated. In addition, what kind of statistic should be calculated may be determined according to the acoustic model to which noise is applied. It should be noted that at least the average of noise components is included.

  Note that the noise statistic calculation unit 202 may calculate the noise statistic in association with the time interval of the input signal, for example, when the acoustic model is subjected to noise adaptation periodically. For example, it waits for the noise data for all input signals in the time interval starting from the current time when the acoustic model has been adapted to be held, and the noise that is valid in that time interval is used using those noise data. It may be calculated as a statistic. Also, for example, the noise data of the previous time interval already held in the estimated noise holding means 201 at the time when the acoustic model adaptation timing is reached (ie, estimated from the input signal included in the previous time interval) Calculated noise data), it may be calculated as a statistical quantity of noise that is effective in the time interval starting from the present time. When the noise statistic calculating unit 202 calculates the noise statistic in association with the time interval of the input signal, the noise statistic calculating unit 202 obtains information on the time interval of the input signal corresponding to the statistic together with the noise statistic. It may be output.

  Next, the acoustic model adaptation unit 206 extracts a noise feature amount based on the noise statistic calculated by the noise statistic calculation unit 202, and noise-adapts the acoustic model (steps S122 and S123). Note that if the statistical amount in the feature amount region has already been calculated as the statistical amount of noise, the process of step S122 is omitted.

  As an application method of the acoustic model, for example, a VTS method or a PMC method may be used. For example, when the VTS method, which is one of acoustic model adaptation methods, is used, the acoustic model adaptation unit 206 uses the noise statistic calculated by the noise statistic calculation unit 202 and the parameters of the acoustic model. Thus, the acoustic model may be noise-adapted by performing a linear operation. In the VTS method, an approximation used when an acoustic model is adapted to a noise environment is a linear approximation. The relationship between the parameters of the acoustic model and the noise statistic of the noise environment is non-linear, but by approximating this linearly, it is a method of approximately generating an acoustic model learned from speech in the noise environment with only linear computation. (Refer nonpatent literature 3.).

  Then, the acoustic model adaptation unit 206 stores the acoustic model subjected to noise adaptation in the acoustic model storage unit 106. For example, the acoustic model adaptation unit 206 may store the parameter value of the acoustic model in the acoustic model storage unit 106 as the acoustic model information. Also, for example, in association with the information of the acoustic model, the average value of the noise component in the statistic of noise adapted to the acoustic model, and the time interval information of the corresponding input signal if necessary are stored. Also good.

  Note that the acoustic model and the time interval of the input signal are not necessarily associated with each other, and are at least the average noise component in the noise statistic adapted to the acoustic model and the input signal input to the acoustic model. Each information should just be matched so that the average of the noise component left to the noise suppression signal may correspond. When noise statistics are calculated in association with the time interval of the input signal, information on the time interval of the noise data that is the source of the calculation may be retained.

  FIG. 5 shows the time change of the power spectrum N_S of the noise (estimated noise), the average power spectrum N_L of the noise component calculated from the noise, and the power spectrum N_S of the short-time fluctuation component (short-time fluctuation noise component) of the noise. It is explanatory drawing which shows an example. FIG. 5A is an explanatory diagram showing an example of a temporal change of the noise power spectrum N in a certain frequency band or subband. FIG. 5B is an explanatory diagram illustrating an example of a temporal change in the average power spectrum N_L of the noise component calculated from the noise illustrated in FIG. FIG.5 (c) is explanatory drawing which shows an example of the time change of the power spectrum N_S of the short time fluctuation component of the noise shown to Fig.5 (a). Moreover, the downward white arrow with respect to the time axis shown in the upper part of FIG. 5A is an example of the timing for applying noise to the acoustic model.

  The example shown in FIG. 5 is an example in which the acoustic model is periodically subjected to noise adaptation, and the noise statistic included in each input signal in each time interval is calculated at the end of the time interval. It is an example calculated from noise data of an input signal. For example, in the example shown in FIG. 5, Ft0, Ft1, Ft2, Ft3, Ft4,... In the time series of the input signal are noise adaptation timings of the acoustic model. In this example, for example, the statistical amount of noise included in the input signal in the time interval Ft0-Ft1 is calculated using the noise data of the input signal in the time interval Ft0-Ft1. Therefore, the noise suppression process for a certain input signal frame in the time interval Ft0-Ft1 is performed after the noise estimation process for all input signals in the time interval Ft0-Ft1. Note that the speech recognition processing is performed after completion of the adaptation processing of the speech recognition model based on the statistics of the noise.

  As another example, when emphasizing the immediate responsiveness of voice recognition, at the start of each time interval, the noise data of the input signal in the immediately preceding time interval is used to convert the input signal in that time interval. It is also possible to calculate the statistics of the included noise. In this case, the time interval shown in FIG. 5A will be described as an example. The statistical amount of noise included in the input signal in the time interval from Ft1 to Ft2 is from Ft0 to Ft1 that is the previous time interval. It is calculated from the noise data of the time interval. In such a case, for example, only noise estimation is performed on an input signal for a certain time from the start (for example, from Ft0 to Ft1), and the acoustic model is subjected to noise adaptation based on the noise data. Using the acoustic model generated at the present time, it is possible to immediately perform noise suppression processing and speech recognition processing for each frame of the input signal.

  FIG. 6 shows, for the power spectrum N of the noise (estimated noise) of the input signal, the average power spectrum N_L of the noise component calculated from the noise and the power of the short-time fluctuation component (short-time fluctuation noise component) of the noise. It is explanatory drawing which shows the other example of the time change of spectrum N_S. 6A and 6B, FIG. 6A shows a noise power spectrum N, FIG. 6B shows an average power spectrum N_L of noise components calculated from the noise, and FIG. c) shows a power spectrum N_S of a short-time fluctuation component of noise. In the example shown in FIG. 6, Ft1, Ft2, Ft3, Ft4,... In the time series of the input signal are noise adaptation timings of the acoustic model. In this example, for example, the statistic of noise included in the input signal of the time interval Ft1-Ft2 is stored in the estimated noise holding unit 201 at the timing of noise adaptation of the acoustic model, for the time interval Ft1-Ft2. Calculation is performed using the noise data of the input signal in the previous time interval Ft0-Ft1. Therefore, noise suppression processing and speech recognition processing for a certain input signal frame in the time interval Ft1-Ft2 can be performed without waiting for noise estimation processing of all input signals in the time interval Ft0-Ft1.

  Note that, as shown in FIGS. 5 and 6, the statistical amount of noise included in the input signal does not change in the time interval. In other words, after the noise statistic is calculated at the timing of noise adaptation of the acoustic model, a constant value is held until the timing of noise adaptation of the next acoustic model.

  FIGS. 7 and 8 are explanatory diagrams showing examples of operation timings of the acoustic model noise adaptation process and the noise suppression process for the input signal, along with examples of noise estimated from the input signal.

  FIG. 7A is an explanatory diagram illustrating an example of noise estimated from an input signal. In addition, the downward white arrow with respect to the time axis shown in the upper part of FIG. 7A is an example of timing at which the acoustic model is adapted to noise. In FIG. 7A, Ft0, Ft1, Ft2, Ft3, Ft4,... In the time series of the input signal are shown as noise adaptation timings of the acoustic model. The right arrow at the bottom of FIG. 7 (a) and its time t1 indicate that the noise data used to calculate the statistical quantity of noise included in the input signal in the time interval starting from the start position is input to the time interval. This is signal noise data, indicating that the noise data is for time t1.

  FIG. 7B is an explanatory diagram showing an example of noise statistics calculated for the noise shown in FIG. 7A and the operation timing of the noise adaptation processing of the acoustic model. In FIG. 7B, the downward white arrow with respect to the time axis shown at the top indicates the timing at which the speech model adaptation unit 200 (more specifically, the noise statistic calculation unit 202) starts the noise statistic calculation process. Is shown. In FIG. 7B, the timing at which the noise statistic calculating unit 202 operates is the time when t1 time has elapsed from the timing at which the acoustic model shown in FIG. In this example, it is assumed that it takes time t2 from the start of the noise statistic calculation process to the completion of the acoustic model noise adaptation process. In FIG. 7B, even if the noise statistic is calculated, the time until the noise statistic is applied to the acoustic model, that is, until the noise adaptation processing of the acoustic model is completed is short. Since it is not the average of the noise components output to the time fluctuation noise component calculation means 103, it is indicated by a broken line indicating that adaptation is in progress. A value indicated by a solid line is output to the short-time fluctuation noise component calculation means 103.

  FIG. 7C is an explanatory diagram illustrating an example of the operation timing of the voice recognition process. In FIG. 7C, the downward white arrow with respect to the time axis shown at the top indicates the timing at which the speech recognition unit 100 (more specifically, the short-time fluctuation noise component calculation unit 103) starts the noise suppression / speech recognition processing. Show. In this example, in response to the completion of the noise adaptation processing for the acoustic model, the short-time fluctuation noise component calculation means 103 is the noise data from which the noise statistic used for the noise adaptation of the acoustic model is calculated again. Noise suppression processing for the obtained input signal is started. That is, after estimating the noise for the input signal, wait for the acoustic model used for the input signal to be applied with noise, and obtain the statistics of the noise used for the acoustic model. Noise suppression is performed. Thereafter, the speech recognition process is performed using the currently stored acoustic model using the suppression signal obtained thereby as an input signal.

  8 shows an example of operation timing different from that in FIG. FIG. 8A is an explanatory diagram illustrating an example of noise estimated from an input signal. In addition, the downward white arrow with respect to the time axis shown in the upper part of FIG. 8A is an example of timing at which the acoustic model is adapted to noise. In FIG. 8A, Ft1, Ft2, Ft3, Ft4,... In the time series of the input signal are shown as noise adaptation timing of the acoustic model. Also, the left arrow at the bottom of FIG. 8A and its time t1 indicate that the noise data used to calculate the statistical amount of noise included in the input signal in the time interval starting from the start position is input in the previous time interval. This is signal noise data, indicating that the noise data is for time t1.

  FIG. 8B is an explanatory diagram showing an example of noise statistics calculated for the noise shown in FIG. 8A and the operation timing of the noise adaptation processing of the acoustic model. In FIG. 8B, the timing at which the noise statistic calculating unit 202 operates is the same as the timing at which the acoustic model shown in FIG. That is, the noise statistics are immediately calculated using the noise data currently held. Other points are the same as in FIG. 7B.

  FIG. 8C is an explanatory diagram illustrating an example of operation timing of the voice recognition process. In the example shown in FIG. 8C, the operation timing of the noise suppression / voice recognition processing by the short-time fluctuation noise component calculation unit 103 and the search unit 105 is not particularly specified. That is, if there is a noise-adapted acoustic model for the input signal currently being processed, the noise statistic used in the acoustic model is acquired, and noise suppression is performed based on that. In this example, an acoustic model that is noise-adapted based on noise statistics based on noise data in a previous time interval is used for an input signal in a certain time interval.

  As described above, according to the present embodiment, it is possible to prevent an adverse effect due to the flooring coefficient α by suppressing the short-time fluctuation noise component instead of the estimated noise from the time series of the input signal. The value of the short-time fluctuation noise component is smaller than the estimated noise. For this reason, when the formula (3) is compared with the formula (1), the possibility that the formula (3) requires a flooring process can be reduced. Therefore, the noise suppression method according to the present embodiment can reduce the adverse effects due to the flooring coefficient, as compared with the noise suppression method that suppresses the estimated noise as in Expression (1).

  Mismatch between the input signal and the acoustic model due to the remaining noise component can be suppressed by applying noise adaptation to the acoustic model based on the statistics of noise. In addition, by performing noise adaptation based on the statistic of acoustic model noise, the frequency of noise adaptation processing for acoustic models with a large amount of computation can be reduced compared to the method of applying noise adaptation to acoustic models based on estimated noise. It is possible to follow the fluctuating noise while suppressing the noise.

Embodiment 2. FIG.
Next, a second embodiment of the present invention will be described with reference to the drawings. FIG. 9 is a block diagram illustrating a configuration example of the speech recognition apparatus according to the present embodiment. As shown in FIG. 9, the speech recognition apparatus according to the present embodiment is different from the first embodiment shown in FIG. 1 in that it further includes trigger generation means 301.

  The trigger generation means 301 controls the noise adaptation timing of the acoustic model. The trigger generation unit 301 may be realized by, for example, a unit based on a voice detection unit or an input device that can generate a trigger according to a user's intention. In this embodiment, the noise statistic calculation unit 202 and the acoustic model adaptation unit 203 of the acoustic model adaptation unit 200 obtain the operation start timing by the trigger generated by the trigger generation unit 301.

  FIG. 10 is a flowchart showing an operation example of the speech recognition apparatus of the present embodiment. Note that the operation example shown in FIG. 10 is basically the same as that of the first embodiment shown in FIG. However, the timing of the operation of the acoustic model noise adaptation process (step S206) is controlled by the trigger generation unit 301. That is, in this example, in place of step S103 in FIG. 2, it is determined in step S203 whether or not a trigger is generated by the trigger generation unit 301, and the presence or absence of the operation of the acoustic model adaptation unit 200 is determined based on the presence or absence of the trigger. To do. Since other operations are the same as those in the first embodiment, description thereof will be omitted.

  For example, the trigger generation unit 301 uses a voice detection unit (not shown) to perform voice detection based on the input signal obtained by the input signal acquisition unit 101, and generates a trigger mainly when it is determined that it is a silent section. You may let them. In such a case, the noise adaptation process of the acoustic model can be operated during the silent period of the input signal. That is, CPU resources can be preferentially assigned to the voice recognition process in the voiced section.

  For example, the trigger generation unit 301 may generate a trigger based on a user instruction input via an input device (not shown). In such a case, the acoustic model adaptation unit 200 can be operated at a timing when the user wants to adapt the acoustic model.

  FIGS. 11 to 12 are explanatory diagrams illustrating examples of operation timings of the acoustic model noise adaptation process and the noise suppression process for the input signal in the present embodiment, along with examples of noise estimated from the input signal. Note that the operation is basically the same as in the first embodiment except that the operation instruction timing of the acoustic model adaptation unit 200 is not a cycle but an arbitrary timing.

  Note that the example shown in FIG. 11 is basically the same as the example shown in FIG. 8 except that the operation instruction of the acoustic model adaptation unit 200 is not periodic. That is, in this example, the noise statistic calculation unit 202 calculates the noise statistic at the same timing as the timing when the operation instruction of the acoustic model adaptation unit 200 is given (see FIG. 11B). Note that the noise data used at this time is the noise data for t1 held at that time. In addition, if there is an acoustic model that has been adapted for noise, the short-time fluctuation noise component calculation unit 103 acquires the statistical amount of noise used in the acoustic model, and performs noise suppression based on that. Thereafter, the search means 105 performs a speech recognition process using the currently stored acoustic model using the suppression signal obtained thereby as an input signal (see FIG. 11C).

  FIG. 12 shows an example in which the short-time fluctuation noise component calculation unit 103 performs noise suppression processing from the input signal when the trigger is generated in response to the completion of the noise adaptation processing of the acoustic model (FIG. 12 ( see c)). Note that the calculation timing of the noise statistic may be the same as in FIG. In addition, in this example, in response to the completion of the acoustic model noise adaptation processing, the short-time fluctuation noise component calculation unit 103 performs noise suppression processing from the input signal when the trigger is generated. For example, when a trigger occurs at Ft1 in the time series of the input signal, the process waits for completion of noise adaptation processing of the acoustic model started by the trigger, and the noise used for the acoustic model generated by the processing , And noise suppression may be performed from the input signal Ft1 at the time of trigger utterance. In addition to this, various timing controls are possible. For example, as in the first embodiment, the noise statistic is calculated using data after the trigger utterance, or the acoustic model is applied from the input signal used to calculate the noise statistic. It is also possible to do.

  As described above, according to the present embodiment, noise adaptation processing of the acoustic model can be performed in the silent section of the input signal, or noise adaptation processing of the acoustic model can be performed based on the user's intention. It is possible to avoid an increase in the amount of calculation due to the operation of the acoustic model adaptation. Further, for example, if the noise adaptation processing of the acoustic model is performed only when the noise environment changes according to the user's instruction, it is possible to perform the speech recognition processing that follows the noise that changes more effectively with a small amount of calculation. That is, since the CPU efficiency can be increased, the speed of the voice recognition process is increased.

Embodiment 3. FIG.
Next, a third embodiment of the present invention will be described with reference to the drawings. FIG. 13 is a block diagram illustrating a configuration example of the speech recognition apparatus according to the present embodiment. As shown in FIG. 13, the speech recognition apparatus of this embodiment is different from the first embodiment shown in FIG. 1 in that it further includes a feature amount conversion unit 401 and a feature amount inverse conversion unit 402. Further, the voice recognition unit 100 further includes a feature amount conversion unit 107.

  The feature amount conversion means 401 and 107 respectively convert noise components into feature amounts. In FIG. 13, the feature quantity conversion means 401 and 107 are shown as separate means, but one feature quantity conversion means can also be shared. Further, as described in the first embodiment, when the search unit 105 has a feature amount conversion function, the feature amount conversion unit 107 may be omitted.

  The feature amount reverse conversion unit 402 performs reverse conversion from the noise feature amount to the noise component. Specifically, the average of noise in the feature amount area calculated by the noise statistic calculating means 202 is converted into a noise component.

  The feature amount conversion units 401 and 107 and the feature amount inverse conversion unit 402 are realized by a CPU that operates according to a program, for example.

  Next, the operation of this embodiment will be described. FIG. 14 is a flowchart illustrating an example of the processing flow of the acoustic model noise adaptation processing of the present embodiment. As shown in FIG. 14, in this embodiment, a feature amount conversion step (see step S <b> 122 in FIG. 4) is not necessary in the acoustic model noise adaptation process. In the present embodiment, the noise statistic calculating unit 202 obtains a noise statistic from the estimated noise feature amount held in the estimated noise holding unit 201 (step S221).

  Although not shown, in the present embodiment, in the noise estimation / data holding step (see step S102 in FIG. 4), when the estimated noise data is held in the estimated noise holding means 201, the feature amount conversion means. Through 401, it is converted into a noise feature and held. Further, in the short-time fluctuation noise component calculation step (see step S111 in FIG. 3) of the noise suppression / voice recognition processing, noise is obtained through the feature amount inverse conversion means 402 when obtaining the average of the noise components. Get the average shown as a component. For example, the feature amount inverse transform unit 402 converts the average value of noise in the feature amount region calculated by the noise statistic calculation unit 202 into noise such as a power spectrum for input to the short-time fluctuation noise component calculation unit 103. Convert to component format. The other points are the same as in the first embodiment.

  As described above, in the present embodiment, the estimated noise data is held in the estimated noise holding unit 201 via the feature amount conversion unit, whereby the estimated noise holding unit 201 holds the estimated noise feature amount. It will be. Also, the noise statistic calculation means 202 calculates the noise statistic from the estimated noise feature.

  In general, the data amount of the feature amount is smaller between the data amount of the input signal and the data amount of the feature amount in one frame. Therefore, according to the present embodiment, the storage area of the estimated noise holding unit 201 can be saved, and the calculation amount in the noise statistic calculating unit 202 can be saved.

  In this embodiment, the example in which the feature amount conversion units 401 and 107 and the feature amount inverse conversion unit 402 are added to the first embodiment has been described. However, such addition is added to the second embodiment. It is also possible.

  The outline of the present invention will be described below. FIG. 15 is a block diagram showing an outline of the present invention. The speech recognition apparatus shown in FIG. 15 includes a noise statistic calculation unit 11, a short-time fluctuation noise component calculation unit 12, an acoustic model adaptation unit 13, a noise suppression unit 14, and a speech recognition unit 15.

  The noise statistic calculation means 11 calculates a noise statistic from noise data estimated for a plurality of frames of the input signal. Note that the noise statistic calculation unit 11 is shown as the noise feature amount calculation unit 202 in the above embodiment.

  The short-time fluctuation noise component calculation unit 12 calculates a short-time fluctuation component of noise included in each frame of the input signal based on the noise statistic calculated by the noise statistic calculation unit 11. Note that the short-time fluctuation noise component calculation means 12 is shown as the short-time fluctuation noise component calculation means 103 in the above embodiment.

  The acoustic model adaptation unit 13 adapts the acoustic model to noise using the noise statistic calculated by the noise statistic calculation unit 11. The acoustic model adaptation means 13 is shown as the acoustic model adaptation means 203 in the above embodiment.

  The noise suppression unit 14 suppresses the short-time fluctuation component of the noise included in the frame calculated by the short-time fluctuation noise component calculation unit 12 for each frame of the input signal. Note that the noise suppression unit 14 is shown as the noise suppression unit 104 in the above embodiment.

  The speech recognition unit 15 performs speech recognition on the input signal suppressed by the noise suppression unit 14 using an acoustic model that is noise-adapted by the acoustic model adaptation unit 13. Note that the voice recognition means 15 is shown as the search means 105 in the above embodiment.

  With such a configuration, it is possible to perform speech recognition with high accuracy even in a changing noise environment.

  Further, the short-time fluctuation noise component calculation means subtracts the average of the noise component indicated by the noise statistic calculated by the noise statistic calculation means from the noise component included in each frame of the input signal. A short-time fluctuation component of noise included in each frame may be calculated.

  The speech recognition apparatus according to the present invention further includes estimated noise holding means (for example, estimated noise holding means 201) that sequentially holds noise data estimated for each frame of the input signal, and the noise statistic calculating means includes: The noise statistics may be calculated using noise data held in the estimated noise holding means.

  In addition, the speech recognition apparatus according to the present invention includes acoustic model storage means (for example, acoustic model storage means 106) that stores information on the acoustic model and information on the statistical amount of noise adapted to the acoustic model in association with each other. The short-time fluctuation noise component calculation means calculates the short-time fluctuation component of noise included in the frame based on the noise statistic stored in the acoustic model storage means, and the speech recognition means Speech recognition may be performed on each frame of the input signal using an acoustic model stored in the means.

  The noise statistic calculation means calculates noise statistic for each predetermined time interval in the input signal time interval, and the acoustic model adaptation means converts the acoustic model to noise every time the noise statistic is updated. The short-time fluctuation noise component calculation means is adapted to reduce the noise included in the frame based on the noise statistic calculated using the noise data of the frame in the time interval including the frame to be calculated. The time recognition component is calculated, and the speech recognition means is subjected to noise adaptation for each frame of the input signal using noise statistics calculated using noise data of frames in the time interval including the frame. Speech recognition may be performed using an acoustic model.

  In addition, the speech recognition apparatus according to the present invention may include trigger generation means (for example, trigger generation means 301) for controlling the timing of noise adaptation of the acoustic model.

  The trigger generation means may generate a trigger for causing the acoustic model adaptation means to start the acoustic model adaptation processing when it is determined that the input signal is a silent section.

  The trigger generation means may generate a trigger for causing the acoustic model adaptation means to start the acoustic model adaptation processing in response to an instruction from the user.

  In addition, the speech recognition apparatus according to the present invention includes a feature amount conversion unit (for example, feature amount conversion units 401 and 107) for obtaining a noise statistic in the feature amount region, and a noise statistic calculated in the feature amount region. Feature amount reverse conversion means (for example, feature amount reverse conversion means 402) for obtaining a short-time fluctuation component of noise from the image data.

  Further, the speech recognition method according to the present invention generates a trigger for starting the adaptive processing of the acoustic model, and calculates noise statistics from the noise data estimated for a plurality of frames of the input signal by the trigger. Then, the acoustic model may be adapted to noise using the calculated noise statistic.

  In addition, the speech recognition method according to the present invention converts the noise data into feature amounts and holds the noise data when sequentially storing the estimated noise data for each frame of the input signal. A statistic in the noise feature quantity region may be calculated from the quantity data.

  In addition, the speech recognition program according to the present invention may cause the computer to execute a process for generating a trigger for controlling the timing of noise adaptation of the acoustic model.

  In addition, the speech recognition program according to the present invention allows a computer to perform a feature amount conversion process for obtaining a noise statistic in the feature amount region, and a short-time fluctuation component of noise from the noise statistic calculated in the feature amount region. You may perform the feature-value reverse transformation process for obtaining.

  The present invention is not limited to speech recognition, and can be applied to uses for obtaining desired data by suppressing noise with respect to input speech.

DESCRIPTION OF SYMBOLS 11 Noise statistic calculation means 12 Short-time fluctuation noise component calculation means 13 Acoustic model adaptation means 14 Noise suppression means 15 Speech recognition means 101 Input signal acquisition means 102 Noise estimation means 103 Short-time fluctuation noise component calculation means 104 Noise suppression means 105 Search Means 106 Acoustic model storage means 201 Estimated noise holding means 202 Noise statistic acquisition means 203 Acoustic model adaptation means 301 Trigger generation means 401, 107 Feature quantity conversion means 402 Feature quantity reverse conversion means

Claims (15)

A noise statistic calculating means for calculating a noise statistic from noise data estimated for a plurality of frames of the input signal;
A short-time fluctuation noise component calculation means for calculating a short-time fluctuation component of noise included in each frame of the input signal based on the noise statistics calculated by the noise statistics calculation means;
Acoustic model adaptation means for adapting the acoustic model to noise using the noise statistics calculated by the noise statistics calculation means;
Noise suppression means for suppressing, for each frame of the input signal, the short-time fluctuation component of the noise included in the frame calculated by the short-time fluctuation noise component calculation means;
A speech recognition apparatus comprising speech recognition means for performing speech recognition on an input signal suppressed by the noise suppression means using an acoustic model noise-adapted by the acoustic model adaptation means. The short-time fluctuation noise component calculation means subtracts the average of noise components indicated by the noise statistic calculated by the noise statistic calculation means from the noise components included in each frame of the input signal, thereby The speech recognition device according to claim 1, wherein a short-time variation component of noise included in the noise is calculated. Estimated noise holding means for sequentially holding noise data estimated for each frame of the input signal,
The speech recognition apparatus according to claim 1, wherein the noise statistic calculation unit calculates a noise statistic using noise data held in the estimated noise holding unit. Acoustic model storage means for storing the information of the acoustic model and the information of the statistical amount of noise adapted to the acoustic model in association with each other;
The short-time fluctuation noise component calculation means calculates a short-time fluctuation component of noise included in the frame based on a noise statistic stored in the acoustic model storage means,
The voice according to any one of claims 1 to 3, wherein the voice recognition means performs voice recognition on each frame of the input signal using an acoustic model stored in the acoustic model storage means. Recognition device. The noise statistic calculating means calculates a noise statistic for each predetermined time interval in the time interval of the input signal,
The acoustic model adaptation means adapts the acoustic model to noise every time the noise statistic is updated,
The short-time fluctuation noise component calculation means is configured to calculate a short-time fluctuation of noise included in a frame based on a noise statistic calculated using noise data of the frame in the time interval including the frame to be calculated. Calculate the ingredients,
The speech recognition means performs speech for each frame of the input signal using an acoustic model that is noise-adapted using a noise statistic calculated using noise data of the frame in the time interval including the frame. The speech recognition apparatus according to any one of claims 1 to 4, wherein recognition is performed. The speech recognition apparatus according to any one of claims 1 to 4, further comprising trigger generation means for controlling a timing at which the acoustic model is subjected to noise adaptation. The speech recognition apparatus according to claim 6, wherein the trigger generation unit generates a trigger for causing the acoustic model adaptation unit to start the acoustic model adaptation process when it is determined that the input signal is a silent section. The speech recognition apparatus according to claim 6, wherein the trigger generation unit generates a trigger for causing the acoustic model adaptation unit to start an acoustic model adaptation process in response to an instruction from a user. A feature value conversion means for obtaining noise statistics in the feature value region;
The speech recognition according to any one of claims 1 to 8, further comprising a feature amount inverse transform unit for obtaining a short-time fluctuation component of noise from a noise statistic calculated in the feature amount region. apparatus. Calculate noise statistics from noise data estimated for multiple frames of the input signal,
Adapting the acoustic model to noise using the calculated noise statistic,
For each frame of the input signal, suppress the short-time fluctuation component of the noise calculated based on the noise statistics used in the noise-adapted acoustic model,
A speech recognition method, wherein speech recognition is performed on the input signal in which the short-time fluctuation component of the noise is suppressed using the noise-adapted acoustic model. Generate a trigger to start the adaptive processing of the acoustic model,
From the noise data estimated for multiple frames of the input signal by the trigger, the noise statistics are calculated,
The speech recognition method according to claim 10, wherein an acoustic model is adapted to noise using the calculated noise statistic. When sequentially storing the estimated noise data for each frame of the input signal, the noise data is converted into feature values and stored,
The speech recognition method according to claim 10, wherein a statistic in a noise feature amount region is calculated from the retained noise feature amount data. On the computer,
A noise statistic calculation process for calculating a noise statistic from noise data estimated for a plurality of frames of the input signal;
Based on the statistics of the noise calculated by the noise statistic calculation processing, a short time variation noise component calculating process of calculating a short-time variation component of the noise contained in each frame of the input signal,
An acoustic model adaptation process for adapting the acoustic model to noise using the noise statistic calculated by the noise statistic calculation process ;
For each frame of the input signal, a noise suppression process for suppressing the short-time fluctuation component of the noise included in the frame calculated by the short-time fluctuation noise component calculation process ;
A speech recognition program for executing speech recognition processing for performing speech recognition on an input signal suppressed by the noise suppression processing using an acoustic model noise-adapted by the acoustic model adaptation processing . On the computer,
The speech recognition program according to claim 13, wherein a process for generating a trigger for controlling a timing for applying noise to the acoustic model is executed. On the computer,
Feature amount conversion processing for obtaining noise statistics in the feature amount region,
The speech recognition program according to claim 13 or 14, wherein a feature amount inverse transform process for obtaining a short-time fluctuation component of noise is executed from a noise statistic calculated in a feature amount region.

Images