JP2009276365A - Processor, voice recognition device, voice recognition system and voice recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To favorably divide a reverberation echo pattern into an initial reflection component and a diffused reverberation component. <P>SOLUTION: This processor measures a reverberation pattern from an impulse response, and divides a measured reverberation pattern into an initial reflection component, which is the former half of the reverberation pattern and a diffused reverberation component, which is the latter half of the reverberation pattern, thereby performing voice recognition for input voice. The processor includes: a decay time boundary calculating part 6 for calculating a decay curve of the reverberation pattern, and calculating a decay time boundary showing a temporal boundary between the initial reflection component and the diffused reverberation component based on the decay curve; and a parameter determination part 7 for an acoustic model, which determines an analysis frame length used in the acoustic model, a frame shift and the number of dynamic characteristic amounts. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a speech recognition technique for speech generated by a speaker. The present invention relates to a speech recognition method.

In speech recognition in a real environment, there is a problem that speech generated by a speaker is obscured due to reverberation and recognition performance deteriorates. In particular, sound collection by a non-contact type microphone such as a hands-free microphone is strongly influenced by reverberation due to the shape of the room and the surroundings of the microphone. Conventionally, a technique for performing speech recognition processing in consideration of a reverberation pattern is known (for example, Patent Document 1).
JP 2007-65204 A

  However, in the conventional speech recognition technique, since the acoustic model is learned in consideration of all parts of the reverberation pattern, sufficient recognition performance cannot be obtained. The reverberation component is decomposed into an initial reflection component, which is a primary reflection component, and a diffuse reverberation component including secondary reflection components, tertiary reflection components, and the like, which are subsequent components. These decomposed reverberation patterns have greatly different properties with respect to characteristics such as power and spectral shape. In the initial reflection component mixed with the initial reflection component of the voice generated by the speaker, the acoustic characteristics of the speech to be recognized remain strong, whereas in the acoustic component mixed with the diffuse reverberation component, The acoustic characteristics of the speech to be recognized are already different in nature. Therefore, regarding the acoustic component mixed with the diffuse reverberation component, it is considered that the acoustic characteristics of the speech to be recognized have disappeared. For this reason, when learning reverberation with an acoustic model, such diffuse reverberation components that affect beyond the frame length of the acoustic model hinder accurate phonological learning.

  By the way, according to the patent application of the present applicant (Japanese Patent Application No. 2008-122288), a reverberation pattern is measured in advance from an impulse response, and the reverberation pattern is determined based on a frame length used for learning an acoustic model. A speech recognition method that divides and treats an initial reflection component that is the first half of the first and a diffuse reverberation component after the initial reflection component is disclosed. After the reverberation pattern is divided in this way, when the acoustic model is learned in advance, the initial reflection component is absorbed by the acoustic model learning. Then, when recognizing the input speech, the diffuse reverberation component is removed from the input speech by spectral subtraction, and the input speech is recognized with reference to an acoustic model considering the initial reflection component. That is, the reverberation pattern is divided so that the initial reflection component is reflected in the acoustic model and the diffuse reverberation component is removed from the input speech. By performing the voice recognition process in this way, more accurate voice recognition can be realized.

  However, since the characteristics of reverberation patterns differ depending on the shape of the room and the surroundings of the microphone, the boundary that divides the reverberation pattern into the initial reflection component and the diffuse reverberation component is determined based on the frame length used for learning the acoustic model. As such, it is not always a good boundary for the environment. In other words, if the reverberation pattern is divided into the initial reflection component and the diffuse reverberation component based on the predetermined acoustic model learning frame length, the reverberation pattern cannot be divided well with respect to the environment. was there. In addition, the user who observes the reverberation pattern measurement result determines the boundary of the reverberation pattern, and there is a problem that a determination process occurs every time after the reverberation measurement, which takes time. In this way, the boundary that divides the reverberation pattern into the initial reflection component and the diffuse reverberation component can be determined well with respect to the environment, and the initial reflection component obtained by dividing the reverberation pattern can be learned well. Is needed.

  The present invention has been made to solve such a problem, and a processing device, a speech recognition device, a speech recognition system, and a speech recognition that can divide and process a reverberation pattern into an initial reflection component and a diffuse reverberation component. It aims to provide a method.

  The processing apparatus according to the present invention measures a reverberation pattern from an impulse response, and converts the measured reverberation pattern into an initial reflection component that is the first half of the reverberation pattern and a diffuse reverberation component that is the second half of the reverberation pattern. A processing device for dividing and recognizing input speech, calculating an attenuation curve of the reverberation pattern, and based on the attenuation curve, a temporal boundary between the initial reflection component and the diffuse reverberation component An attenuation time boundary calculating unit for calculating an attenuation time boundary indicating the acoustic frame length, an analysis frame length used in the acoustic model, a frame shift, and the number of dynamic features based on the calculated attenuation time boundary A model parameter determination unit.

  Thereby, the attenuation boundary time for dividing the reverberation pattern is calculated according to the measured attenuation state of the reverberation pattern, and the parameters for the acoustic model (analysis frame length, frame shift, By determining the number of dynamic feature quantities, the reverberation pattern can be favorably divided into the initial reflection component and the diffuse reverberation component.

  In addition, as a method for calculating the decay time boundary by the decay time boundary calculation unit, a time when the reverberation pattern is decreased by a predetermined amount may be set as the decay time boundary, and a change amount of the decay curve of the reverberation pattern is predetermined. The time when the value falls below the threshold may be set as the decay time boundary.

Further, the acoustic model parameter determination unit further includes a storage unit that stores in advance a combination of values of the analysis frame length Tf, the frame shift Ts, and the number of dynamic features N used in the acoustic model. A combination of values of the analysis frame length Tf, the frame shift Ts, and the number N of the dynamic feature quantities satisfying the following expression with respect to the decay time boundary Ta calculated by the decay time boundary calculation unit: May be selected.
Ta≈Tf + N × (2 × Ts)

  The speech recognition apparatus according to the present invention measures a reverberation pattern from an impulse response, and the measured reverberation pattern includes an initial reflection component that is the first half of the reverberation pattern, and a diffuse reverberation component that is the second half of the reverberation pattern. A speech recognition device for recognizing input speech, calculating an attenuation curve of the reverberation pattern, and determining a temporal boundary between the initial reflection component and the diffuse reverberation component based on the attenuation curve. An attenuation time boundary calculation unit for calculating an attenuation time boundary to be shown; and an acoustic model for determining an analysis frame length, a frame shift, and the number of dynamic features used in the acoustic model based on the calculated attenuation time boundary A parameter determination unit for use, a reverberation component extraction unit that divides the measured reverberation pattern at the attenuation time boundary to extract the initial reflection component and the diffuse reverberation component, and the extraction A learning unit that reflects the initial reflection component in the learning speech data and learns an acoustic model based on the determined analysis frame length, frame shift, and number of dynamic features, and the input speech A recognition unit that removes the extracted diffuse reverberation component and recognizes the input speech with reference to the learned acoustic model.

  Thereby, the attenuation boundary time for dividing the reverberation pattern is calculated according to the measured attenuation state of the reverberation pattern, and the parameters for the acoustic model (analysis frame length, frame shift, By determining the number of dynamic feature quantities, the reverberation pattern can be favorably divided into the initial reflection component and the diffuse reverberation component. By reflecting the divided early reflection components in the speech data for learning, a better acoustic model can be constructed, and the recognition rate of the input speech can be improved by removing the diffuse reverberation component from the input speech. Can do.

  A voice recognition apparatus according to the present invention includes the voice recognition apparatus described above and a microphone that receives a sound generated in the environment and outputs a voice signal to the voice recognition apparatus.

  The speech recognition method according to the present invention measures a reverberation pattern from an impulse response, and the measured reverberation pattern includes an initial reflection component that is the first half of the reverberation pattern, and a diffuse reverberation component that is the second half of the reverberation pattern. A speech recognition method for recognizing an input speech by calculating an attenuation curve of the reverberation pattern and determining a temporal boundary between the initial reflection component and the diffuse reverberation component based on the attenuation curve. Calculating an attenuation time boundary to be indicated; determining an analysis frame length used in the acoustic model, a frame shift, and a number of dynamic features based on the calculated attenuation time boundary; Dividing the reverberation pattern at the attenuation time boundary to extract the initial reflection component and the diffuse reverberation component; and extracting the extracted initial reflection component into the learning audio data. Reflecting an acoustic model based on the determined analysis frame length, frame shift, and number of dynamic features, and removing the extracted diffuse reverberation component from the input speech And recognizing the input speech with reference to the learned acoustic model.

  Thereby, a reverberation pattern can be divided | segmented favorably according to the attenuation condition of the measured reverberation pattern. Then, by determining acoustic model parameters (analysis frame length, frame shift, and number of dynamic features) based on the calculated attenuation boundary time, the divided initial reflection components are reflected in the learning speech data. In this case, a more suitable acoustic model parameter can be determined. As a result, a better acoustic model can be constructed, and the recognition rate of the input speech can be improved by removing the diffuse reverberation component from the input speech.

  According to the present invention, it is possible to provide a processing device, a speech recognition device, a speech recognition system, and a speech recognition method that can perform processing by dividing a reverberation pattern into an initial reflection component and a diffuse reverberation component.

  Hereinafter, the best mode for carrying out the present invention will be described in detail with reference to the drawings. For clarity of explanation, the following description and drawings are omitted and simplified as appropriate. In the drawings, components having the same configuration or function and corresponding parts are denoted by the same reference numerals, and description thereof is omitted.

Embodiment 1 of the Invention
The speech recognition system according to the first embodiment measures a reverberation pattern from an impulse response, and based on the measured decay curve of the reverberation pattern, an initial reflection component that is the first half of the measured reverberation pattern, and the latter half It has a function of calculating a decay time boundary indicating a temporal boundary with a certain diffuse reverberation component. The speech recognition system has a function of determining the analysis frame length, the frame shift, and the number of dynamic features used in the acoustic model based on the calculated decay time boundary.

  First, with reference to FIG. 1, characteristic components of the speech recognition system according to the first embodiment will be described. FIG. 1 is a block diagram showing the configuration of the speech recognition system according to the first embodiment. The voice recognition system according to the first embodiment includes a microphone 1 (hereinafter referred to as a microphone 1) and a voice recognition device 2. The speech recognition device 2 includes a reverberation processing unit 3 that processes a reverberation pattern, a learning unit 4 that learns an acoustic model, and a recognition unit 5 that recognizes input speech. Details of the reverberation component extraction unit 8, the learning unit 4, and the recognition unit 5 will be described later.

  The microphone 1 is provided in the environment and receives sound generated in the environment. Therefore, the microphone 1 collects the voice spoken by the speaker and outputs a voice signal corresponding to the received voice to the voice recognition device 2. The microphone 1 is installed in a room of a building, for example. The microphone 1 is installed at a predetermined place in the environment.

  The voice recognition device 2 performs voice processing on the voice signal from the microphone 1 to perform voice recognition. The speech recognition device 2 is a processing device having a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), a communication interface, and the like, and performs data processing necessary for speech recognition. Further, the voice recognition device 2 has a removable HDD, optical disk, magneto-optical disk, etc., stores various programs and control parameters, and supplies the programs and data to a memory (not shown) as necessary. To do. For example, the voice recognition device 2 converts a signal from the microphone 1 into a digital signal and performs arithmetic processing. Furthermore, the voice recognition device 2 executes voice recognition processing according to a program stored in the ROM or HDD. That is, the voice recognition device 2 stores a program for voice recognition, and the voice recognition device 2 performs various processes on the digital signal according to the program.

  The reverberation processing unit 3 includes an attenuation time boundary calculation unit 6, an acoustic model parameter determination unit 7, and a reverberation component extraction unit 8. An impulse response is input to the reverberation processing unit 3. The reverberation processing unit 3 outputs acoustic model parameters (analysis frame length, frame shift, and number of dynamic features) used in the acoustic model to the learning unit 4. The reverberation processing unit 3 extracts the initial reflection component and the diffuse reverberation component from the reverberation pattern, and outputs the extracted initial reflection component and the diffuse reverberation component. The extracted initial reflection component is output from the reverberation processing unit 3 to the learning unit 4. The extracted diffuse reverberation component is output from the reverberation processing unit 3 to the recognition unit 5.

  The decay time boundary calculation unit 6 measures a reverberation pattern from the input impulse response. Then, an attenuation curve of the reverberation pattern is calculated, and an attenuation time boundary indicating a temporal boundary between the initial reflection component and the diffuse reverberation component is calculated based on the calculated attenuation curve.

  Here, the initial reflection component and the diffuse reverberation component included in the reverberation pattern will be described with reference to FIGS. FIG. 2 is a diagram schematically illustrating how the sound generated in the room is reflected. FIG. 3 is a diagram illustrating an example of a signal detected by the microphone 1 installed in the environment. In FIG. 3, the horizontal axis indicates time, and the vertical axis indicates signal power. In FIG. 3, the waveform of the measurement signal when the impulse response is measured in the environment is shown discretely.

  FIG. 2 shows an example in which the speech recognition system shown in FIG. As shown in FIG. 2, the sound uttered by the speaker 45 in the room reaches the microphone 1 mounted on the robot 44 and is measured. The uttered sound to be measured may propagate directly to the microphone 1 or may be reflected by the wall surface 43 and propagate to the microphone 1. Of course, the light may be reflected not only by the wall surface 43 but also by a ceiling, floor, desk, or the like. The sound reflected by the wall surface 43 or the like is delayed compared to the sound that directly reaches the microphone 1. That is, the direct sound that has reached the microphone 1 directly differs from the reflected sound that has reached the microphone 1 after being reflected by the wall surface 43, and the timing at which the sound is measured by the microphone 1 is different. Further, among the sounds reflected by the wall surface 43, the sound reflected repeatedly has a further time delay. Thus, the sound measurement timing differs depending on the sound propagation distance and the like. In addition, in a room, sound is reflected not only on the wall surface 43 but also on the ceiling, floor, desk, or the like.

  When an impulse composed of a single pulse having a very narrow width is generated in a room as shown in FIG. 2, the measurement signal has the waveform shown in FIG. In the impulse time response, the direct sound that directly reaches the microphone 1 without being reflected by the wall surface 43 is measured at the earliest time (t = 0). And the reflected sound reflected by the wall surface 43 is measured after the direct sound. Since the reflected sound is absorbed by the wall surface 43 or the like, the power is lower than that of the direct sound. The reflected sound that is repeatedly reflected is measured over time.

  Here, the reverberation pattern of the impulse response is divided into an initial reflection component and a diffuse reverberation component. Therefore, a reverberation component is measured from the impulse response, and the measured reverberation component is divided into an initial reflection component and a diffuse reverberation component. Of the reverberation pattern, the first half is the initial reflection component and the second half is the diffuse reverberation component. Therefore, the diffuse reverberation component follows the initial reflection component. The initial reflection component includes low-order reflection components such as primary reflection and secondary reflection. Further, the diffuse reverberation component includes a high-order reflection component.

  Here, a temporal boundary that divides the initial reflection component and the diffuse reverberation component is defined as an attenuation time boundary. Therefore, the component from the time when the direct sound is measured by the microphone 1 to the attenuation time boundary is the initial reflection component, and the component after the attenuation time boundary is the diffuse reverberation component. For example, if the decay time boundary is 65 msec, the data at t = 0 is a direct sound, the data in the range of 0 to 65 msec (not including t = 0, t = 65) is the initial reflection component, and the data after 65 msec is Diffuse reverberation component.

  As shown in FIG. 4, the decay time boundary calculation unit 6 can calculate, for example, the time t when the amplitude level (power P) of the reverberation pattern decreases by a predetermined amount x as the decay time boundary Ta. In the figure, the time t of the power P (t) when the power P (0) at time 0 is attenuated by x [db] is Ta (that is, P (0) −P (t)> x is satisfied). Time t to be treated as Ta). Here, the predetermined amount x is, for example, 20 db. The attenuation time boundary calculation unit 6 may calculate the attenuation time boundary according to the amount of decrease in the amplitude level (power P) of the reverberation pattern as described above, or the amount of change in the attenuation curve of the reverberation pattern is predetermined. The time when the value falls below the threshold may be calculated as the decay time boundary.

Based on the decay time boundary Ta calculated by the decay time boundary calculation unit 6, the acoustic model parameter determination unit 7 analyzes the analysis frame length Tf used in the acoustic model, the frame shift Ts, and the number N of dynamic features Δ. , Determine. More specifically, the acoustic model parameter determination unit 7 determines the analysis frame length Tf, the frame shift Ts, and the motion that satisfy the following expression with respect to the decay time boundary Ta calculated by the decay time boundary calculation unit 6. It can be determined by calculating a combination of the number N of the characteristic feature amounts.
Ta≈Tf + N × (2 × Ts)

  The acoustic model parameter determination unit 7 calculates a combination of values of the analysis frame length Tf, the frame shift Ts, and the number N of dynamic feature quantities that satisfies the above formula as described above. Alternatively, a combination of values of the analysis frame length Tf, the frame shift Ts, and the number N of dynamic features is stored in advance, and the above equation is obtained from the stored combinations of these values. You may make it determine by selecting the combination of the value of the analysis frame length Tf, the frame shift Ts, and the number N of dynamic feature-values to be satisfied. For example, {25, 2, 10}, {32, 3, 16}, {20, 2, 5} may be used as the combination of values {Tf, N, Ts} stored in advance. it can.

  Thus, the attenuation boundary time for dividing the reverberation pattern is calculated according to the measured attenuation state of the reverberation pattern, and the parameters for the acoustic model (analysis frame length, frame shift, and By determining the number of dynamic feature quantities, the reverberation pattern can be favorably divided into the initial reflection component and the diffuse reverberation component.

  The reverberation component extraction unit 8 divides and extracts the reverberation pattern into the initial reflection component and the diffuse reverberation component with the decay time Ta calculated in this way as a boundary. Then, the learning unit 4 to be described later convolves the divided initial reflection component with the learning speech data. The learning unit 4 learns the convolved data according to the acoustic model parameters (analysis frame length Tf, frame shift Ts, and number N of dynamic feature quantities) determined by the acoustic model parameter determination unit 7. Thereby, the learning of the acoustic model is executed by reflecting the initial reflection component. Further, the recognizing unit 5 recognizes the input speech by removing the diffuse reverberation component extracted from the input speech and referring to the acoustic model learned by reflecting the initial reflection component.

  Next, details of processing by the reverberation component extraction unit 8, the learning unit 4, and the recognition unit 5 will be described. When the speech recognition device 2 recognizes the input speech, the recognition unit 5 removes the diffuse reverberation component divided by the reverberation processing unit 3 from the input speech uttered by the speaker and causes the learning unit 5 to learn in advance. The input speech is recognized with reference to the placed acoustic model.

  FIG. 5 is a block diagram showing a detailed configuration of the voice recognition system. The voice recognition system includes a microphone 1 and a voice recognition device 2. The speech recognition device 2 includes a reverberation component extraction unit 8, a learning unit 4, and a recognition unit 5. In FIG. 5, the reverberation processing unit 3, the decay time boundary calculation unit 6, and the acoustic model parameter determination unit 7 are not shown.

  The reverberation component extraction unit 8 includes an initial reflection component extraction processing unit 11 and a diffuse reverberation component extraction processing unit 21. The learning unit 4 includes a convolution processing unit 12, a learning speech database 13, an acoustic model learning processing unit 14, and an acoustic model 15. The recognition unit 5 includes a spectrum conversion processing unit 22, a filter creation unit 23, a spectrum conversion processing unit 31, a spectrum subtraction processing unit 32, a speech recognition feature value conversion unit 33, and a pattern matching processing unit 34. ing.

  The initial reflection component extraction processing unit 11, the convolution processing unit 12, the learning speech database 13, and the acoustic model learning processing unit 14 perform processing for creating an acoustic model 15 necessary for speech recognition. Thereby, the acoustic model 15 reflecting the initial reflection component of the reverberation pattern of the audio signal is created. Here, a hidden Markov model (HMM) is used as the acoustic model 15. The processing here is performed offline in advance. That is, the acoustic model 15 is created in advance before detecting a voice signal for voice recognition.

  The diffusion reverberation component extraction processing unit 21, the spectrum conversion processing unit 22, and the filter creation unit 23 perform processing for removing the diffusion reverberation component. Thereby, a subtraction filter for subtracting the diffuse reverberation component is created. The processing here is performed offline in advance. That is, a subtraction filter is created in advance before detecting a voice signal for voice recognition.

  The spectrum conversion processing unit 31, the spectrum subtraction processing unit 32, the speech recognition feature value conversion unit 33, and the pattern matching processing unit 34 perform speech recognition processing on the input speech. The speech recognition process is performed using the subtraction filter and the acoustic model 15 described above. These processes are performed online with respect to the input voice, so that the voice is recognized as needed.

  Next, learning of the acoustic model 15 using the initial reflection component will be described with reference to FIGS. FIG. 6 is a diagram illustrating a learning flow of the acoustic model. The process shown in FIG. 6 is performed offline. That is, the acoustic model 15 is created by the processing flow shown in FIG. 6 before acquiring the speech signal to be recognized.

As shown in FIG. 5, the initial reflection component extraction processing unit 11 extracts the initial reflection component from which the diffuse reverberation component is removed from the impulse response input. That is, as described above, the impulse response is measured by the microphone 1, and data before the attenuation time boundary is extracted as the initial reflection component from the reverberation component of the measured impulse response. As shown in FIG. 6, the initial reflection component and h _E. Convolution processing unit 12 performs convolution processing by using the initial reflection components h _E.

The learning speech database 13 stores clean learning speech data. For example, the learning speech database 13 stores speech data in units of phonemes as a database. The voice data is measured in a place where there is no noise or reverberation. For example, a conversation for one hour is used as a corpus. Each phoneme included in the corpus is labeled “A” or “I”. Thus, clean speech data for phonemes is stored in the learning speech database 13. Then, the convolution processing unit 12 convolves the initial reflection component h _E with the clean speech data s stored in the learning speech database 13. Thus, the initial reflection components h _E is is reflected convolved data x _E is generated. By convolving the initial reflection component h _E with each speech data s in units of phonemes, convolution data x _E for each phoneme is calculated.

Acoustic model learning processing unit 14 based on the convolution data x _E initial reflection components is reflected, performs an acoustic model learning process. When the acoustic model 15 is an HMM, the acoustic model learning processing unit 14 performs HMM learning. When performing HMM learning, learning is performed based on the analysis frame length Tf, the frame shift Ts, and the number N of dynamic feature amounts Δ described above. More specifically, to extract a feature from the convolution data x _E. And the feature-value of a phoneme unit is memorize | stored as a database. That is, the feature quantity vector for each phoneme becomes a template model. The feature vector is extracted for each analysis length, for example.

Specifically, it converted into spectral data by FFT convolution data x _E (Fast Fourier Transform) or the like. Then, the spectral data is logarithmically converted using a filter matched to human auditory characteristics, and further converted into time data by IFFT (Inverse Fast Fourier Transform). In this way, a mel cepstrum is required. In the mel cepstrum space, the spectral envelope appears in the lower order and the fine vibrations appear in the higher order. Then, the low order part is taken out and the MFCC is calculated. Here, a 12-dimensional MFCC is calculated. Further, the primary difference and the power primary difference are extracted as feature quantities. In this case, the feature quantity vector has 25 dimensions (12 + 12 + 1). Of course, the process for extracting the feature quantity is not limited to this.

  Then, learning is performed using the MFCC data group. It should be noted that by processing the speech data s included in a large amount of corpus, the feature amount for one phoneme has an average and a variance. The acoustic model 15 holds average and variance values. And the acoustic model learning process part 14 determines the state transition probability, output probability, etc. of HMM according to the average and dispersion | distribution of a feature-value. The acoustic model learning processing unit 14 learns the HMM using, for example, an EM algorithm. Of course, a known algorithm other than the EM algorithm may be used. In this way, the acoustic model 15 is learned.

  The acoustic model 15 learned by the acoustic model learning processing unit 14 is stored as a database. The acoustic model 15 takes into consideration the initial reflection component of the reverberation pattern. That is, the initial reflection component is modeled and estimated by the HMM. Thereby, the acoustic model 15 having learned the initial reflection component is constructed. By using this acoustic model 15, the influence of the initial reflection component contained in the audio signal can be reduced, and the recognition rate can be improved.

  Next, filter creation processing using a diffuse reverberation component will be described with reference to FIGS. FIG. 7 is a conceptual diagram for explaining approximate calculation for creating a filter. FIG. 8 is a diagram showing a processing flow for creating a filter.

  As shown in FIG. 5, the diffusion reverberation component extraction processing unit 21 performs diffusion reverberation component extraction processing on the impulse response input. Thereby, the diffuse reverberation component from which the initial reflection component is removed from the reverberation pattern of the impulse response is extracted. That is, among the reverberation components of the impulse response measured by the microphone 1, the diffuse reverberation components are extracted from the data after the attenuation time boundary. The spectrum conversion processing unit 22 converts impulse response time data into spectrum data. That is, the time domain diffuse reverberation component data is converted to frequency domain data. Here, the data of the diffuse reverberation component is converted using Fourier transform or the like. That is, it is converted into frequency domain data by FFT (Fast Fourier Transform) or the like. Note that the spectrum conversion processing unit 22 performs framing processing according to the analysis frame length and the frame shift before conversion into spectrum data.

  The filter creation unit 23 creates a subtraction filter for removing diffuse reverberation using the data of the diffuse reverberation component. First, approximate calculation for creating a filter will be described with reference to FIG. FIG. 7 shows online processing for voice recognition.

As shown in FIG. 7, a speech signal by speech spoken by the speaker as an input x, the diffuse reverberation components in the impulse response a rear impulse response h _L. Spectral subtraction processing is performed to remove the back diffuse reverberation x _L for the input x from the input x. After subtracting the spectrum, it is converted into a feature value, and voice recognition is performed by pattern matching.

However, the back diffuse reverberation x _L for the input x cannot be observed directly. In other words, it is not possible to observe only the rear diffusion reverberation x _L. Therefore, the rear diffuse reverberation x _L is approximated using the rear impulse response h _L observed in advance. That is, if x ′ _L (= x * h _L ) can be approximated to x _L , the spectrum component of the diffuse reverberation component can be subtracted. Therefore, to create a filter that what convolving a rear impulse response to an input x can be approximated to the rear spreading reverberation x _L.

The off-line processing for approximating in this way will be described with reference to FIG. Here, an impulse response is measured, and a filter δ is created from clean learning speech data s. The rear impulse response h _L (t) is convolved with the voice data s stored in the learning voice database 13. Thus, the rear diffusion reverberation x _L is generated. Further, the impulse response h is convoluted with the voice data s stored in the learning voice database 13. That is, the entire impulse response h is convoluted with the audio data s. As a result, the input x in the case where a clean voice is emitted is generated. Further, the rear impulse response h _L (t) is convolved with the input x. That is, after convolution of the impulse response h with the audio data s, the rear impulse response h _L (t) is further convolved with the data. This rear impulse response h _L (t) is the same as the rear impulse response h _L (t) convolved with clean audio data.

The above processing is performed for each of the voice data s included in the learning voice database 13. Then, the calculated rear diffuse reverberations x _L and x _'L estimates the filter δ as close. That is, a coefficient that satisfies x _L ≈δx ′ _L is calculated. Here, the filter δ is estimated by the least square error calculation. That, x _L performs processing so as to minimize the error function between .delta.x _'L. As a result, δ such that δx ′ _L is closest to x _L can be calculated. Here, the optimum coefficient differs in the frequency band. Therefore, the filter δ is estimated for each frequency band. As shown in the upper right of FIG. 8, an optimum coefficient is calculated for each frequency band. Specifically, a 12-dimensional filter δ (δ ₁ , δ ₂ , δ ₃ , δ ₄ ,... Δ ₁₂ ) is estimated. By using this filter δ, spectral subtraction can be performed to remove the diffuse reverberation component from the audio signal. That is, the filter δ is a subtraction filter that can subtract the diffuse reverberation component.

  Next, online speech recognition processing will be described with reference to FIGS. FIG. 9 is a diagram showing a processing flow of voice recognition. First, the input voice detected by the microphone 1 is input to the voice recognition device 2. In FIG. 9, the input voice is input x. The spectrum conversion processing unit 31 converts the input x into spectrum data. That is, time domain data is converted to frequency domain data by FFT or the like. The spectrum conversion processing unit 31 performs framing processing according to the analysis frame length and the frame shift before conversion into spectrum data.

  The spectrum subtraction processing unit 32 subtracts the diffuse reverberation component from the spectrum data using the filter δ. By performing the spectrum subtraction process using the filter δ in this way, the influence of the diffuse reverberation component is removed from the audio signal. Based on the subtraction data obtained by subtracting the spectrum of the diffuse reverberation component, the speech is recognized as follows.

  The voice recognition feature value conversion unit 33 converts the spectrum data into a voice recognition feature value. The speech recognition feature value conversion unit 33 extracts a feature value based on the subtraction data obtained by subtracting the diffuse reverberation component. As the feature quantity, for example, a 12-dimensional Mel Frequency Cepstrum Coefficient (MFCC) can be used. Therefore, a filter bank analysis using a mel filter is performed. Then, logarithmic transformation (Log transformation) is performed, and discrete cosine transformation (DCT) is performed to calculate the MFCC. Here, as described above, a 25-dimensional feature vector including the MFCC primary difference and the power primary difference is calculated.

  When MFCC is used as a feature amount for speech recognition, the recognition rate can be further improved. That is, nonlinear processing such as spectral subtraction causes distortion when returning to an audio signal, but does not pose any problem when converted to MFCC. That is, since the spectrum data from which the diffuse reverberation component is removed is directly converted to MFCC without returning to the audio signal, the occurrence of distortion can be prevented.

  And the pattern matching process part 34 performs a pattern matching process using the feature-value vector of the acoustic model 15. FIG. As a result, the phoneme having the pattern closest to the feature vector for the detected speech signal is recognized. That is, the pattern matching processing unit 34 is a recognition processing unit that performs voice recognition processing with reference to the acoustic model 15.

  As described above, since the acoustic model 15 in which the initial reflection component is reflected is used when executing the speech recognition processing, a more excellent acoustic model 15 can be constructed. Since the initial reflection component from which the diffuse reverberation component (high-order reflection component) that affects beyond the analysis length to be learned is removed is used for learning, accurate phonological learning can be performed. Since the influence of the initial reflection component can be absorbed by the HMM learning, the speech recognition rate can be improved.

  Furthermore, the diffuse reverberation component is used as a filter δ for spectral subtraction. For this reason, the diffuse reverberation component of the input speech can be removed. Thereby, the influence of a diffuse reverberation component can be reduced and the speech recognition rate can be improved.

  In the present embodiment, the impulse response is measured in the same environment as the environment in which the voice signal that is actually recognized is acquired, and the initial reflection component and the diffuse reverberation component are extracted from the reverberation pattern of the measured impulse response. Here, impulse response measurement is performed in a room where the microphone 1 is installed. The reverberation of the room and the shape around the microphone can be almost the same as long as there is no significant change such as moving from room to room. Therefore, if the environment is the same, the diffuse reverberation component can be regarded as almost constant regardless of the direct sound. In other words, the diffuse reverberation component is substantially constant regardless of the voice that is spoken. After determining the method of installing the microphone, it is possible to estimate the initial reflection component and the diffuse reverberation component separately by measuring the reverberation for the impulse response of the room only once.

  That is, the impulse response is measured in advance in the environment, and the initial reflection component and the diffuse reflection component are extracted. The acoustic model 15 reflecting the initial reflection component and the filter δ created based on the diffuse reflection component are repeatedly used for voice recognition in the environment. That is, the same filter δ and the acoustic model 15 are used for audio signals detected in the same environment. Since the impulse response only needs to be measured once in advance, learning of the acoustic model 15 and creation of the filter δ can be performed easily. In addition, since the acoustic model 15 and the filter δ created in advance are used, the amount of processing online can be reduced. Therefore, speech recognition with a high recognition rate can be performed with simple processing.

  When the environment changes due to the speaker 5 moving from room to room, impulse response measurement is performed once in that environment. Then, the acoustic model 15 is learned and the filter δ is created by the same processing. The recognition rate can be improved by performing model learning and filter creation according to the environment. Alternatively, when the microphone 1 is replaced, impulse response measurement is performed with the replaced microphone 1 and the same processing is performed. Of course, the environment is not limited to indoors, but may be in a car or outdoors. For example, a voice recognition system may be installed in a car navigation system.

  The acoustic model 15 may be an acoustic model other than the HMM. That is, the initial reflection component may be used for learning the acoustic model 15 other than the HMM. In addition, since the reverberation can be removed by one microphone 1, the system configuration can be simplified.

  Furthermore, each process may be performed by different computers. For example, a computer that performs acoustic model learning and filter creation processing may be physically different from a computer that performs speech recognition. In this case, online processing and offline processing are performed by different devices.

  Specifically, the initial reflection component extraction processing unit 11, the convolution processing unit 12, the learning speech database 13, the acoustic model learning processing unit 14, the diffuse reverberation component extraction processing unit 21, the spectrum conversion processing unit 22, and the filter creation unit 23. The acoustic model 15 and the filter δ are created in advance. Then, the created acoustic model 15 and the filter δ are stored in advance in a speech recognition device having the spectrum conversion processing unit 31, the spectrum subtraction processing unit 32, the speech recognition feature value conversion unit 33, and the pattern matching processing unit 34. Then, a voice signal is detected by the microphone 1 connected to the voice recognition device 2, and the above processing is performed on the voice signal. Even in this case, speech recognition processing with a high recognition rate can be easily performed. Alternatively, a computer that performs speech recognition may perform processing with reference to the acoustic model 15 and the filter δ stored in another computer such as a processing device.

  Furthermore, the computer that performs acoustic model learning and the computer that performs filter creation may be physically different. Further, different impulse response measurement results may be used between filter creation and acoustic model learning. That is, the initial reflection component and the diffuse reverberation component may be extracted from different impulse response measurements. For example, impulse response measurement may be performed twice, an initial reflection component may be extracted based on one impulse response measurement, and a diffuse reverberation component may be extracted based on the other impulse response measurement. By mounting the voice recognition system on a voice response type robot, an accurate voice response can be performed. In addition, when the audio | voice signal by a continuous audio | voice is input, you may recognize a audio | voice further using a language model.

It is a figure which shows the structure of the speech recognition system concerning embodiment of this invention. It is a figure which shows a mode that the sound generated in the environment reflects. It is a figure which shows typically the audio | voice signal detected with the audio | voice recognition system concerning embodiment of this invention. It is a figure which shows typically the audio | voice signal detected with the audio | voice recognition system concerning embodiment of this invention. It is a figure which shows the detailed structure of the speech recognition system concerning embodiment of this invention. It is a figure which shows the learning process flow in the speech recognition system concerning embodiment of this invention. It is a figure which shows the approximate calculation of the filter creation process in the speech recognition system concerning embodiment of this invention. It is a figure which shows the processing flow of filter creation in the speech recognition system concerning embodiment of this invention. It is a figure which shows the processing flow in the speech recognition system concerning embodiment of this invention.

Explanation of symbols

1 microphone, 2 speech recognition device,
3 reverberation processing unit, 4 learning unit, 5 recognition unit,
6 Decay time boundary calculation unit, 7 Acoustic model parameter determination unit,
8 reverberation component extraction unit, 11 initial reflection component extraction processing unit,
12 convolution processing unit, 13 learning speech database,
14 acoustic model learning processing unit, 15 acoustic model database,
21 diffuse reverberation component extraction processing unit, 22 spectrum conversion processing unit,
23 filter creation unit, 31 spectrum conversion processing unit,
32 spectrum subtraction processing unit, 33 speech recognition feature amount conversion unit,
34 Pattern matching processing section

Claims (10)

The reverberation pattern is measured from the impulse response, and the measured reverberation pattern is divided into an initial reflection component that is the first half of the reverberation pattern and a diffuse reverberation component that is the second half of the reverberation pattern, A processing device for performing recognition,
An attenuation time boundary calculation unit that calculates an attenuation curve of the reverberation pattern and calculates an attenuation time boundary indicating a temporal boundary between the initial reflection component and the diffuse reverberation component based on the attenuation curve;
An acoustic model parameter determination unit that determines an analysis frame length, a frame shift, and the number of dynamic features used in the acoustic model based on the calculated decay time boundary;
A processing apparatus comprising: The decay time boundary calculation unit includes:
The processing apparatus according to claim 1, wherein a time when the reverberation pattern decreases by a predetermined amount is set as the decay time boundary. The decay time boundary calculation unit includes:
The processing apparatus according to claim 1, wherein a time when the amount of change in the decay curve of the reverberation pattern falls below a predetermined threshold is set as the decay time boundary. A storage unit that stores in advance a combination of values of the analysis frame length Tf, the frame shift Ts, and the number N of dynamic features used in the acoustic model;
The acoustic model parameter determination unit
A combination of values of the analysis frame length Tf, the frame shift Ts, and the number N of the dynamic feature quantities satisfying the following expression with respect to the decay time boundary Ta calculated by the decay time boundary calculation unit: Ta≈Tf + N × (2 × Ts)
The processing apparatus according to any one of claims 1 to 3. Reverberation pattern is measured from impulse response, and the measured reverberation pattern is divided into an initial reflection component that is the first half of the reverberation pattern and a diffuse reverberation component that is the second half of the reverberation pattern to recognize input speech A voice recognition device that
An attenuation time boundary calculation unit that calculates an attenuation curve of the reverberation pattern, and calculates an attenuation time boundary indicating a temporal boundary between the initial reflection component and the diffuse reverberation component based on the attenuation curve;
An acoustic model parameter determination unit that determines an analysis frame length, a frame shift, and the number of dynamic features used in the acoustic model based on the calculated decay time boundary;
Dividing the measured reverberation pattern at the decay time boundary to extract the initial reflection component and the diffuse reverberation component;
A learning unit that reflects the extracted initial reflection component in learning audio data and learns an acoustic model based on the determined analysis frame length, frame shift, and number of dynamic features,
A recognition unit that removes the extracted diffuse reverberation component from the input speech and recognizes the input speech with reference to the learned acoustic model;
A speech recognition apparatus comprising: A voice recognition device according to claim 5;
A microphone that receives sound generated in the environment and outputs a voice signal to the voice recognition device;
A speech recognition system. Reverberation pattern is measured from impulse response, and the measured reverberation pattern is divided into an initial reflection component that is the first half of the reverberation pattern and a diffuse reverberation component that is the second half of the reverberation pattern to recognize input speech A voice recognition method for
Calculating an attenuation curve of the reverberation pattern, and calculating an attenuation time boundary indicating a temporal boundary between the initial reflection component and the diffuse reverberation component based on the attenuation curve;
Determining an analysis frame length, a frame shift, and a number of dynamic features to be used in the acoustic model based on the calculated decay time boundary;
Dividing the measured reverberation pattern at the decay time boundary to extract the initial reflection component and the diffuse reverberation component;
Reflecting the extracted initial reflection component in learning speech data, and learning an acoustic model based on the determined analysis frame length, frame shift, and number of dynamic features;
Removing the extracted diffuse reverberation component from the input speech and recognizing the input speech with reference to the learned acoustic model;
A speech recognition method comprising: The speech recognition method according to claim 7, wherein an attenuation curve of the reverberation pattern is calculated, and a time when the calculated attenuation curve decreases by a predetermined amount is set as the attenuation time boundary. The speech recognition method according to claim 7 or 8, wherein an attenuation curve of the reverberation pattern is calculated, and a time when a change amount of the calculated attenuation curve falls below a predetermined threshold is used as the attenuation time boundary. . A pre-stored analysis of combinations of values of the analysis frame length Tf, the frame shift Ts, and the number N of dynamic feature quantities satisfying the following expression with respect to the calculated decay time boundary Ta Determined by selecting from combinations of values of the frame length Tf, the frame shift Ts, and the number N of dynamic feature quantities Ta≈Tf + N × (2 × Ts)
The speech recognition method according to claim 7, wherein:

Images