JP2816163B2 - Speaker verification method - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to the field of speaker recognition, and more particularly to a speaker verification method for identifying a speaker by voice.

[Conventional technology]

In general, in speaker recognition, it is not efficient to directly compare an input speech wave with a registered speech wave, so that the comparison is performed after conversion into acoustic parameters such as a frequency spectrum and a linear prediction coefficient. The acoustic parameters include a fundamental frequency (pitch frequency), voice energy, formant frequency, Percoll coefficient, log cross-sectional area ratio, number of zero crossings, and the like.

These acoustic parameters primarily include phonological information and secondarily personality information. Therefore, when performing speaker verification, new speaker-specific features are additionally obtained from the acoustic parameters. Creating a quantity is done. Conventionally, when creating this feature amount, the input voice is divided into frames every several msec, and various acoustic parameters (spectrum, cepstrum) obtained for each frame are averaged in the time axis direction over the entire voiced sound section, Statistical analysis was performed on the entire obtained acoustic parameters to convert them into speaker-specific features.

As a known document related to speaker verification, for example, JP-A-61-278896 is cited.

[Problems to be solved by the invention]

The method of averaging the parameters in the time axis direction has advantages that the processing is relatively simple and the size of the feature amount (dictionary capacity) is small. However, in order to create a stable feature, a relatively long time voice sample is required, and the generation of the feature by adding the time-series acoustic parameters in the time axis direction compresses the information in the time axis direction. Therefore, it is not suitable for text-dependent speaker verification. Further, in this method, the personality information to be superimposed on the phoneme is also averaged along with the averaging of the phonological information, so that the feature amount does not sufficiently reflect the personality information.

An object of the present invention is to provide a speaker verification method for identifying a speaker by faithfully extracting personality information included in a voice signal.

[Means for solving the problem]

The present invention divides an input audio signal into frames for each predetermined time unit, calculates an acoustic parameter for each frame, detects an audio section from the input audio signal, and converts the audio section on the time axis into an audio section. The method is characterized in that a block is divided into a plurality of blocks according to the value of the parameter, and a characteristic amount is generated by, for example, adding and averaging acoustic parameters in the time axis direction for each block, and collation determination is performed using the characteristic amount. The voice section is divided into one block by extracting the voiced section by discriminating the presence / absence, unvoiced and silent sections using the power of the input voice, the rate of change of the spectrum, the pitch, and the like, and extracting the voiced section.

Further, the present invention has a pre-processing for dividing a uttered voice into blocks in advance at the time of registration based on its power, power dip, spectrum change rate (dynamic scale), and the like, and based on a reference block obtained by this pre-processing. , The input audio signal is divided into blocks.

In addition, the present invention uses a spectrum as an acoustic parameter, and divides a predetermined number of uttered voices into a plurality of blocks according to the value of the first moment of the spectrum in advance when dividing a block at the time of registration. Block is used as a reference block, and all input voices at the time of registration are divided into blocks again based on the reference block. At the time of matching, unknown voices are blocked based on the reference block created at the time of the previous registration. It is characterized in that it is divided.

Further, the present invention is characterized in that a pitch period is added to the feature amount, and a speaker is identified by comprehensive distance determination between the spectrum and the pitch.

In extracting the pitch period, the input audio signal is passed through a low-pass filter to remove high-frequency components, then binarized by a plurality of thresholds, and a transition point in a specific direction of the binarized signal by each threshold. Is measured by a counter, and the pitch period or frequency is estimated from each of the measured time intervals.

[Action]

In the present invention, the speech section of the input speech is divided into several blocks on the time axis, and the feature amount is calculated for each block, so that information on the time change is added to the feature amount, and the personality information is also obtained. Emphasize.

However, in this case alone, since the block division is performed by the power of the voice, the power dip, the spectrum change rate, and the like, the extraction position of the block division is unstable due to the influence of the power fluctuation of the input voice for each utterance, and the Block positions are different. Therefore, at the time of registration as preprocessing of block division, the uttered voice is divided into blocks in advance by its power and the like, and the input audio signal is divided into blocks based on the reference block obtained thereby, thereby stabilizing the block division. Plan.

In particular, by applying the duration control type transition model using the state of the first moment of the speech spectrum to block division, it is possible to roughly associate each block,
As a result, a stable block division ratio can be achieved. Also, the amount of calculation is much smaller than block division by DP matching or the like.

Next, in the present invention, the matching rate is improved by adding a voice pitch period to the feature amount. In order to describe an individual feature, it is important to include pitch in addition to the spectrum which is a phonological parameter, and it is considered that this improves the matching rate. However, using an average value as an acoustic parameter that has individuality in its value and its time change pattern, such as pitch, cannot express variations within a block well. I can't say. Therefore, in the present invention, as a feature of the pitch period,
A histogram of the pitch period for each block is used. As a result, the value of the pitch period and the manner of its time change can be reflected in the feature value itself, and can be converted into a feature value effective for speaker verification, thereby improving the verification rate.

Further, in general, complicated processing is required to accurately determine the pitch.
By applying a simple pitch extraction method using a combination of a low-pass filter and a counter, the system size of the entire speaker verification can be reduced.

In this case, the first moment of the spectrum is used as the feature amount of the block division, and the average spectrum and the pitch histogram are used as the individual feature amounts. However, the LPC cepstrum may be used instead of the spectrum.

〔Example〕

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

FIG. 1 is a block diagram of a first embodiment for explaining the basic principle of the present invention. In this embodiment, the microphone 1,
Acoustic parameter conversion section 11, voice section detection section 12, voiced / silence discrimination section 13, voiced section extraction section 14, feature quantity generation section 15, standard pattern storage section 16, distance calculation section 17, determination section 18, and control It is composed of object 2.

The audio signal input from the microphone 1 is converted by the audio parameter conversion unit 11 into a time series of audio parameters for each frame. As the acoustic parameter conversion unit 11, for example, an input voice is
After cutting the component of more than 1/2 of the sampling frequency, the analog-to-digital converter quantizes it into a discrete signal sequence, cuts it out for each short-time waveform, and adds a hamming window, etc. Then, it may be converted into various feature amounts (spectrum, cepstrum, PARCOR, pitch, etc.), or an acoustic parameter may be obtained as time-series information of the spectrum using a band-pass filter group. Here, an example using a band-pass filter group of 29 channels arranged at intervals of 1/6 octave at a center frequency of 250 to 6300 Hz will be described.

The acoustic parameter conversion unit 11 divides the input audio signal into frames every 10 msec, and sequentially converts them into a time-series pattern f ({f ₁ , f ₂ ,..., F ₂₉ }) of the spectrum. Now, time i
Spectrum f _ij frame in _{the, f ij = (f i1,} f i2, ... f i29) (j = 1,2, ..., 29) is expressed as.

The voice section detection unit 12 detects a voice section from the input voice signal using the time-series pattern obtained by the acoustic parameter conversion unit 11.

The sound / non-speech determining unit 13 and the sound section extracting unit 14 divide the sound section into several blocks (the sound section is regarded as one block) using the time series vector f _ij . FIG. 2 shows a procedure for performing division by the power of the input audio signal. That is, one frame is input (step S1), the 29 spectra obtained for each frame are added to make the power of one frame (step S2), and this power is compared with a preset threshold. Voiced voice section,
It is divided into silent teblocks (steps S3 and S4), and a continuous voiced section is divided into one block (step S3).
Five). Here, the processing of steps S1 to S4 is handled by the voiced / silence discriminating unit 13, and the processing of step S5 is performed by the voiced segment extraction unit 14
Is responsible for.

FIG. 3 shows a specific example of the block division.
(A) shows a time spectrum pattern (TSP) of a certain voice, and (b) shows a relationship between a time change of power and a threshold in (a).

(B) The voice section is divided into blocks of three sound sections I, II, and III according to the threshold values in the figure. (B) When the position of the section shown by the figure is shown on the time spectrum pattern, (c)
It becomes a figure.

The feature value generation unit 15 obtains a feature value for each block extracted by the sound section extraction unit 14. Here, it is assumed that the averaging in the time direction of the intra-block spectrum time series is used as the feature amount. For example, when the spectrum sequence included in the k-th block is fkij = (fki1, fki2,..., Fki29), k: the k-th block, i: the time in the k-th block, j: the channel number of the band-pass filter, k The block feature Xkavr is Becomes Here, m indicates the number of frames in the block, and ▼ (j = 1, 2,..., 29) indicates the average of the spectrum for each channel. In addition, as the feature quantity, Xkavr may be regarded as one spectrum, this least square line may be calculated, and the slope or the like may be added as one element of the feature vector.

At the time of registration, the feature amount obtained by the feature amount generation unit 15 is added to the standard pattern storage unit 16 with the file name of the speaker and stored as a standard feature amount (dictionary feature amount). On the other hand, at the time of matching, the distance calculation unit 17 calculates the distance between the feature amount of the unknown speaker obtained by the feature amount generation unit 15 and the feature amount of the registered speaker stored in the standard pattern storage unit 16. The calculation is performed for each block, and the calculated value is obtained by averaging the load over all blocks. Here, the distance may be determined using a Euclidean distance, a Mahalanobis distance, or the like. Further, a process of reducing the number of dimensions of the feature amount using an F ratio (inter-speaker / in-speaker variance ratio) or the like may be performed in advance.

The determination unit 18 determines a speaker by comparing a threshold value preset for each speaker with the distance obtained by the distance calculation unit 17. The control target 2 is controlled based on this determination result. Examples of the control target 2 include a banking service system, an entry management system using voice keys, and a response device such as a toy.

FIG. 4 shows a block diagram of a second embodiment of the present invention. In the present embodiment, only a voiced sound section is extracted from a voice section of an input voice signal detected by the voice section detection section 22 by using a voiced / unvoiced / silent discrimination section 23 and a voiced section extraction section 24. This is one block, and the rest is the same as FIG. Note that power, spectrum change rate (dynamic scale), pitch, and the like are used for voiced, unvoiced, and silent section discrimination.

FIG. 5 shows the relationship between voice patterns divided into voiced, unvoiced, and silent, and extracted blocks. In FIG.
V is a voiced sound section, Vl is an unvoiced sound section, S is a silent section, and the section of V is a block.

FIG. 6 shows a block diagram of a third embodiment of the present invention. In FIGS. 1 and 4, only the voiced or voiced sound section of the voice section is extracted and only that section is taken into account. In the present embodiment, however, the power dip of the input voice signal is detected and this power dip is detected. Is compared with a preset threshold value to divide the entire voice section into several blocks. That is, the power dip is detected by the power dip detector 33, and the block detector 34 first compares the power dip with a threshold value set in advance by the threshold value setting unit 39. Next, the audio section is divided at a position where the power dip is lower than the threshold, and the divided section is defined as a block. Otherwise, it is the same as FIG.

FIG. 7 shows a specific example. In FIG. 7A, four power dips are detected, but three power dips are selected according to the threshold value. Therefore, the voice section is divided according to the positions of these three dips. FIG. 7 (b) shows the divided blocks.

FIG. 8 shows a block diagram of a fourth embodiment of the present invention. In FIG. 6, the entire voice section is divided into blocks by detecting the power dip of the input voice and comparing the power dip with a preset threshold. On the other hand, in this embodiment, the spectrum change is performed instead of the power dip. By detecting the local peak using a rate (dynamic scale), block division of the entire voice section is performed. That is, first, the spectrum change rate calculator 43 calculates the change rate of the spectrum of the input audio signal, and the local peak is detected by the local peak detector 44. In the block dividing unit 45, the detected local peak is compared with a predetermined threshold value by the threshold value setting unit 50, and a voice section is divided using a position where the local peak is higher than the threshold value as a reference point, and the divided section is defined as a block. Set. The other processes related to the feature amount are the same as those in FIG.

FIG. 9 shows a specific example. Here, countless local peaks are detected, but six are selected according to the threshold value. Therefore, the voice section is divided into seven sections using the positions of these six peaks as reference points. In FIG. 9, I to VII indicate divided blocks.

FIG. 10 shows a block diagram of a fifth embodiment of the present invention. The embodiment of FIGS. 6 and 8 detects the power dip of the input voice and the local peak of the rate of change of the spectrum, compares them with a predetermined threshold value, and puts a break into the voice section. The voice section was divided into blocks. On the other hand, in the present embodiment, in the threshold adjusting unit 60, the number of blocks is set in advance, and the threshold is made variable so as to become the number of blocks.
Other processes are the same as those in FIG.

FIG. 11 shows a processing procedure of the threshold adjustment unit 60. Threshold adjuster
For example, in the case of 60, 7 (anb =
It is assumed that 7) is set. In the block division unit 55,
Block division is performed by comparing the detected local peak of the spectrum change rate with a predetermined threshold (Th).
The threshold adjusting unit 60 determines that the number of divided blocks (nb) is 7
When the number is larger than 7, the threshold is increased, and when the number of blocks is smaller than 7, the threshold is decreased. The threshold value is changed by this operation, and the process is repeated until the number of blocks becomes seven.

FIG. 12 shows a specific example of this block division. That is,
12A, the number of blocks becomes 12 by the initial threshold value, but in FIG. 10B, the number of blocks is obtained as 7 by changing the threshold value as shown in the figure.

Note that a power dip may be used instead of the spectrum change rate. In this case, when the number of divided blocks is larger than the predetermined number of blocks, the threshold value is decreased, and when the number of blocks is smaller than the predetermined number of blocks, the threshold value is increased.

FIG. 13 shows a block diagram of a sixth embodiment of the present invention. In the embodiments described above, since the block division is performed based on the power of the voice, the power dip, the spectrum change rate, and the like, the extraction position of the block division is unstable due to the influence of the power fluctuation of the input voice for each voice, and the voice The position of each block is different. Therefore, in the present embodiment, at the time of registration as preprocessing, the uttered voice is divided into blocks in advance based on its power and the like, and the input audio signal is divided into blocks based on the reference block obtained thereby, thereby achieving block division. It is intended to stabilize.

The audio signal input from the microphone 1 is converted into a time series of audio parameters for each frame by an audio parameter conversion unit 61, and the audio section detection unit 62 uses these feature amounts to convert the audio section from the input audio signal. To detect. Up to this point, the operation is the same as in the previous embodiments.

The block dividing unit 64 divides the voice section into several blocks on the time axis. The block division is performed, for example, by adding the 29 spectra obtained for each frame using the time series vector f _ij of the above spectrum to obtain a power of one frame, and matching this power with a preset reference block. Divide the voice section. On the other hand, in the block division preprocessing unit 63, at the time of registration as preprocessing, similarly, the input audio signal is divided into blocks in advance by its power and the like, and this is reflected as a reference block in the subsequent block division of the block division unit 64. .

The feature amount generation unit 65 obtains a feature amount for each block extracted by the block division unit 64. Here, it is assumed that the averaging in the time direction of the intra-block spectrum time series is used as the feature amount. For example, when the spectrum sequence included in the k-th block is fkij = (fki1, fki2,..., Fki29), k: the k-th block, i: the time in the k-th block, j: the channel number of the band-pass filter, k The block feature Xkavr is Becomes Here, m indicates the number of frames in the block, and ▼ (j = 1, 2,..., 29) indicates the spectrum average for each channel. In addition, as the feature quantity, it is also possible to regard Vkavr as one spectrum, calculate this least-square straight line, and add the slope or the like as an element of the feature vector.

At the time of registration, the feature amount obtained by the feature amount generation unit 65 is added to the standard pattern storage unit 66 with the file name of the speaker and stored as a standard feature amount (dictionary). On the other hand,
The distance calculator 67 calculates the distance between the feature amount of the unknown speaker obtained by the feature amount generation unit 65 and the feature amount of the registered speaker stored in the standard pattern storage unit 66 for each corresponding block. This is obtained by averaging the load over all blocks. Here, the distance may be determined using a Euclidean distance, a Mahalanobis distance, or the like. Further, a process of reducing the number of dimensions of the feature amount using an F ratio (inter-speaker / in-speaker variance ratio) or the like may be performed in advance.

The determination unit 68 determines a speaker by comparing a threshold value preset for each speaker with the distance obtained by the distance calculation unit 67. The control target 2 is controlled based on this determination result.

Hereinafter, the block division pre-processing unit 63 and the block division unit 64
Will be described in detail.

FIG. 14 shows the procedure of one embodiment of the pre-block division processing and the block division. First, the pre-block division processing unit 63 divides the first utterance into blocks using the power, power dip, spectrum change rate, and the like of the voice as pre-processing for block division, and obtains a block division point. . The block division unit 64 performs DP matching on the second and subsequent input voices using the first voice as a dictionary pattern to obtain a frame position corresponding to a block division point, thereby performing block division.

FIG. 15 shows another embodiment for block division. First, the pre-block division processing unit 63 divides the first utterance into blocks using the power, power dip, spectrum change rate, and the like of the speech as pre-processing for block division, and sets the acoustic parameters in the block. A feature amount is obtained for each block by performing averaging in the time axis direction. The block dividing unit 64 obtains a frame position corresponding to the reference block using the feature amount of each first audio block as a dictionary pattern and a duration control type state transition model or the like for the second and subsequent input voices. Block division. FIG. 16 shows the correspondence between the feature amount of each reference block and the second and subsequent audio frames. Here, represents a feature amount in the block of the standard pattern, and I to I represent blocks of the input voice.

FIG. 17 shows still another embodiment for block division. First, the block division preprocessing unit 63 divides the first utterance into blocks using the power, power dip, spectrum change rate, and the like of the voice as preprocessing for block division, as in the case of FIG. Then, the acoustic parameters in the block are averaged in the time axis direction or the like, and the feature amount is obtained for each block. In the block division unit 64, when dividing a second or subsequent input audio signal into blocks, first, the first audio is divided into blocks using the power, power dip, spectrum change rate, and the like of the input audio signal. Divide into blocks. Next, a feature amount is obtained by averaging the acoustic parameters in the semi-block in the time axis direction or the like, and the feature amount of each first audio block is used as a dictionary pattern, and DP matching with the feature amount of each semi-block is performed. Is performed to obtain the position of the semi-block corresponding to the reference block, and the block is divided based on the position. FIG. 18 shows the correspondence between the feature amount of each reference block and the second and subsequent semi-blocks. Here, 〜 indicates a feature amount in a block of a standard pattern, and 〜 to を indicate a feature amount in a semi-block of an input voice.

FIG. 19 shows another embodiment for block division. As preprocessing of block division in the block division preprocessing unit 63, first, all voices uttered at the time of registration are divided into blocks using their power, power dip, spectrum change rate (dynamic scale), and the like. Then, the representative number of blocks is obtained by averaging the number of blocks for each utterance. The block division unit 64 performs block division of the input voice at the time of registration and at the time of collation based on the block of the voice pattern indicating the representative number of blocks.

In this case, the positioning of the blocks may be performed by DP matching with the sound of the reference block, or the acoustic amount of the reference block may be averaged in the time axis direction to obtain a feature amount for each block. A frame position corresponding to the reference block may be obtained by using a state model or the like with a duration control type using the feature amount as a dictionary pattern, and block division may be performed using this. In addition, voices that do not have a representative number of blocks are divided into semi-blocks in advance, and feature amounts obtained by averaging acoustic parameters in the semi-blocks in the time axis direction, and feature amounts in the reference block. By performing the DP matching or the like, the position of the semi-block corresponding to the reference block may be obtained, and the block division may be performed using this.

 FIG. 20 is a block diagram showing a seventh embodiment of the present invention.

Generally, when block division is performed using audio parameters such as audio power, power dip, and spectrum change rate, the processing is complicated, and it is difficult to always reliably obtain a feature amount of audio that changes for each utterance. At the same time, it is difficult to divide all uttered voices into blocks at the same position when creating a feature at the time of registration and collation. Therefore, it is conceivable to associate block positions by linear expansion / contraction or DP matching. However, these methods aim at the correspondence in the frame unit to the last,
It is unsuitable for association such as finding a section position of a series of block areas, and the amount of calculation increases.

Therefore, in the present embodiment, the first moment is not determined by using the first moment as a criterion, instead of dividing the voice section into voiced, unvoiced, or unvoiced by using the first moment of the spectrum of the voice frame divided every predetermined time. Is divided into one identical section (block),
This is performed by setting a division point unique to the speaker and the uttered voice. Further, the same voice is associated with each block by setting the average of the first moments in the reference block set in advance as the state of the block, and applying the duration control type transition model to obtain the voice of the input voice. This is performed by dividing the first moment of the section into a plurality of blocks. According to this method, the association between blocks is the association between a state and a frame, and the number of operations is determined by the number of grid points. Therefore, processing can be performed with a very small number of operations. In addition, the calculation at the grid point requires only one subtraction of the first moment, and the calculation amount is very small as compared with the DP matching or the like in which the normal vector calculation is performed as the calculation of the grid point.

Hereinafter, the embodiment of FIG. 20 will be described. Even here
The acoustic parameter conversion unit 71 divides the input audio signal from the microphone 1 into frames every 10 msec and sequentially converts them into a time series pattern f ({f ₁ , f ₂ ,..., F ₂₉ }) of the spectrum. In this case, the spectrum f _ij of the frame at the time i is expressed as f _ij = (f _i1 , f _i2 ,... F _i29 ) (j = 1, 2,..., 29).

The voice section detection unit 72 detects a voice section from the input voice signal using the time-series pattern of the acoustic parameter conversion unit 71. Here, detection of a voice section from the input signal includes calculating a power spectrum Pwi for each frame, comparing this value with a predetermined threshold Thpw, and detecting a section in which the power exceeds the threshold Thpw as a voice section. Let's do it. Here
The audio power Pwi at time i is It is expressed as Next, the first moment calculator 73 calculates the first moment of the spectrum of each frame. Here, the first moment of the spectrum at time i is Is calculated by This mi indicates the position of the center of gravity of the spectrum of each frame, and takes a value between 1 and 29.

By the way, in general, the spectrum of voiced voices is concentrated in the low frequency band, and unvoiced voices are concentrated only in the high frequency band. Therefore, the value of the first moment is small in voiced sound periods and large in non-voiced sound periods. Will take. In a silent section, a value near the center (in the vicinity of 15 in terms of value) is taken. However, a phoneme and a silent section in which spectral power is concentrated in the center of the channel have similar first moment values and cannot be clearly distinguished. Further, the noise in the silent section has a large change in the first moment according to the above equation, and the first moment lacks stability. Therefore,
A predetermined frequency band is weighted in advance to stabilize the first moment. As a measure for this, a first moment may be calculated by adding a weight to a predetermined channel out of the 29 channels, or another temporary frequency band may be set above the high band and below the low band. Then, a predetermined load may be applied thereto to calculate the first moment. Here, a temporary channel is set on the high frequency channel, and a predetermined load W is added to the channel to obtain a first moment. At this time, the spectrum of the frame at time i is f _ij = (f _i1 , f _i2 ,..., F _i29 , W) (j = 1, 2,..., 29). Becomes FIG. 21 shows the spectrum and first-moment moment change over time when [Akjunkan] is uttered.

At the time of registration, the pre-block division processing unit 74 creates a reference block. First, all voices uttered at the time of registration are divided into a plurality of blocks by comparing the first moment of the spectrum with a predetermined threshold. The change in the first moment is relatively stable, depending on the individual and the spoken word. In the example of FIG. 21, the value of the first moment of the section considered to be a voiced sound section is small, and is large in the unvoiced and silent sections. Therefore, by setting the predetermined threshold value Th, the first moment can be divided into a plurality of blocks in which the first moment is stable.

FIG. 22 shows a flowchart for explaining an example of the setting method of the division point. After performing initial settings i = 0 and n = 0 (step S1), set i = 1 + 1 (step S2),
First, a position where the first moment mi becomes smaller than the threshold Th is detected (step S3). When mi becomes smaller than Th,
This position is stored in sp [n] of the memory as the start point position of the n-th block section (step S4). Next, mi is Th
If it is smaller, the value of i is increased (step S5), and when it becomes larger than Th, the value of i at that time is stored in ep [n] of the memory as the end position of the n-th block section (step S5). S6, S7). The value of n is increased, and the same processing is repeated until the end of the voice section, and block division is performed for all registered voices (steps S8 and S9).

After the pre-block division processing is completed for all registered voices in this way, the reference block setting unit 75 selects a reference block. As the selection criterion, a maximum value, a minimum value, a median value (median), or the like of the number of blocks may be used, or a comparison is made for every corresponding block among all voices, and for example, an average value of block length is obtained Then, a block of a sound sample having a block length closest to the average value may be selected. Here, the block having the smallest number of block divisions is set as the reference block.

In the re-division processing unit 76, based on the average first moment in the reference block, all the input speech including the speech for which the reference block was created is again divided into blocks. As a method at this time, for example, a state length control type state transition model may be used. FIG. 23 is a diagram for explaining this, in which the vertical axis represents the state of the reference block (subscript j), and the horizontal axis represents the temporal change of the first moment of the voice section of the input voice (subscript i). The vertical axis and the horizontal axis are associated with each other using a duration control state transition model. FIG. 24 shows the result of subdividing all the sounds at the time of registration for the example of FIG.

The feature quantity creation unit 78 calculates the feature quantity of the speaker for each block. The calculation of the feature amount here is the same as that of the feature amount generation unit 65 in FIG. 13, and for example, an average spectrum is calculated for each block. At the time of registration, the characteristic amount creation unit 78 stores the calculated characteristic amount in the standard pattern storage unit 81 with the file name of the speaker added. That is, as the personal dictionary pattern at the time of registration, an average spectrum within a block calculated for all voices at the time of registration is collected for each corresponding block, and for example, an average value and a variance thereof are calculated.

At the time of matching, the block division unit 77 controls the duration of the first moment of the speech section of the unknown speech calculated by the first moment calculation unit 73 and the state of the reference block of the registered speaker registered in the standard pattern storage unit 81 in advance. Pattern matching is performed using the type state transition model to determine a division point. The feature amount generation unit 78 calculates an average spectrum for each of the obtained blocks to obtain a feature amount. Then, the distance calculator 79 calculates the distance between the feature amount of the unknown speaker and the feature amount of the registered speaker.

The determination unit 80 compares the distance obtained as described above with a predetermined threshold value set in advance, and determines whether the user is the person or the impostor.

FIG. 25 shows a block diagram of the eighth embodiment of the present invention. The operations of the acoustic parameter conversion unit 81, the voice section detection unit 82, and the block division unit 83 are the same as in the previous embodiments.

The feature value generation unit 84 calculates the feature value of each block obtained by the block division unit 83. The feature value generation unit 84 includes an averaging spectrum calculation unit 84-1 and a least square line calculation unit.
84-2.

Again, the Fourem spectrum sequence that enters the k-th block is represented by fkij = (fki1, fki2, ..., fki29) k: k-th block i: time in k-th block j: channel number of bandpass filter and I do.

Using the averaging in the time direction of the intra-block spectrum time series as the feature amount and letting this feature amount be Xavr, Is as described with reference to FIG. 13, for example.

The feature value Xavr is calculated by the averaging spectrum calculation unit 84. Further, the feature Xavr is regarded as one spectrum, the least square straight line is calculated by the least square straight line calculator 84-2, and the slope Xa is also used as the feature.

The feature amount integration unit 85 integrates the feature amounts Xavr and Xa. That is, if the entire feature amount of the k blocks is Xk, then Xk = (Xavr, Xa).

At the time of registration, the feature amount (Xa) integrated by the feature amount integration unit 85 is stored as a standard feature amount (dictionary pattern) with the file name of the speaker added to the standard pattern storage unit 86. On the other hand, at the time of collation, the distance calculation
Features of unknown speakers obtained through 85 and features of registered speakers stored in the standard pattern storage unit 86 (dictionary patterns)
Is calculated for each corresponding block.

Now, assuming that the dictionary pattern is X and the unknown voice pattern is Z, the distance d (X, Z) is Ask for. Here, dis (Xk, Zk) is the sum of the distances between the corresponding features of the corresponding block, and dis (Xk, Zk) = w1 × dist (Xavr, Zavr) + w2 × dist (Xa,
Za). Calculation of distance between corresponding features dist (a, b)
May use the Euclidean distance, the Mahalanobis distance, or the like. Here, w1 to w2 indicate weights for the respective distances, and are set in advance using an F ratio (inter-speaker / in-speaker separation ratio) or the like.

The determination unit 88 determines a speaker by comparing a threshold value preset for each speaker with the distance obtained by the distance calculation unit 87. The control target 2 is controlled based on this determination result.

FIG. 26 is a block diagram showing a ninth embodiment of the present invention. This is because the feature amount generation unit 94 is added to the averaged spectrum calculation unit 94-
1. It is composed of a least squares straight line calculation unit 94-2 and a difference calculation unit 94-3, and calculates a least squares straight line from the average spectrum in a block for each block as a feature amount in the block.
The remaining spectrum outline obtained by drawing a least-squares straight line from the average spectrum is used, and other than that is the same as the embodiment of FIG.

FIG. 27 is a block diagram of a tenth embodiment of the present invention. This means that the feature amount generation unit 104
-1, a least-squares straight line calculation unit 104-2 and a difference spectrum calculation unit 104-3, which calculates a least square line of the spectrum outline for each frame in the block, and subtracts the least square line from the spectrum outline. The obtained spectrum is used as personality information in the frame, and is added and averaged in the time axis direction to obtain a feature amount in the block.

FIG. 28 is a block diagram showing an eleventh embodiment of the present invention. In order to add a fine structure of a spectrum for each frame to a feature amount, position information of peaks (peaks) and bottoms (valleys) of the spectrum is introduced as a method for efficiently describing the feature. The feature amount generation unit 114 includes a peak-to-valley pattern conversion unit 114-1, a peak-to-valley position extraction unit 114-2 of a spectral shape, and an averaging unit 114-3.

First, it is determined which channel has a peak in the spectrum of the k-th block in the k-th block.
If the first channel is a peak, “1” is used, otherwise “0” is used to express the position of the peak of the spectrum of each frame as follows.

TOPkij = (TOPki1, TOPki2, ..., TOPki29) TOPkil 1: When the j-th channel is at a peak 0: When the j-th channel is at a non-peak k: The k-th block i: The time at the k-th block j: Frequency Of the channel number block of the TO
The averaging in the time direction of Pkij is used. If this is TOPavr, Becomes Here, m is the number of frames in the block, ▲
▼ is the average value in the block where the peak of the spectrum appears on the channel.

The same applies to the bottom (valley) of the spectrum.
That is, for the i-th spectrum in the k-th block, it is checked which channel has a valley. BTMkij = (BTMki1, BTMki2,..., BTMki29) This is the frequency channel number.

Here, BTMkij contains “1” if the j-th channel is a valley, and “0” otherwise.

The feature value in the block is BT obtained for each frame.
Mkij averaging in the time direction is used. If this is BTMavr, Becomes

Summarizing the above features, the overall feature Xk of k blocks
Then, Xk = (XTOPavr, XBTMavr).

When performing the distance calculation, the distance between the corresponding blocks between the dictionary pattern and the unknown pattern is calculated using the above feature amounts. Here, if the dictionary pattern is X and the unknown voice pattern is Z, the distance d (X, Z) is Ask for. Here, dis (Xk, Zk) is the sum of the distances between corresponding feature values between corresponding blocks, and dis (Xk, Zk) = w3 × dist (XTOPavr, ZTOPavr) + w4 × dis
Expressed as t (XBTMavr, ZBTMavr). Calculation of distance between corresponding features dist (a, b)
May use the Euclidean distance, the Mahalanobis distance, or the like. Here, w3 to w4 indicate weighting for each distance, and are set in advance using the F ratio or the like.

In this case, only the position information of the peaks and valleys of the spectrum is used as the feature vector of each block, but the average spectrum in the block described in the previous embodiment, the slope thereof, and the like are added, and Xk = (Xavr, Xa , XTOPavr, XBTMavr).

In this case, the sum of the distances between corresponding blocks, dis (Xk,
Zk) is obtained by: dis (Xk, Zk) = wl × dist (Xavr, Zavr) + w2 × dist (Xa,
Za) + w3 × dist (XTOPavr, ZTOPavr) + w4 × dist (XBTMa
vr, ZBTMavr). Calculation of distance between corresponding features dist (a, b)
May use the Euclidean distance, the Mahalanobis distance, or the like. Here, w1 to w4 indicate weighting for each distance, and are set in advance using the F ratio or the like.

FIG. 29 is a block diagram showing a twelfth embodiment of the present invention.

In the embodiments described above, the input sound is divided into frames every several msec, the sound section is divided into a plurality of blocks using the sound parameters obtained for each frame, and the averaging of the sound parameters in each block is performed. Basically, values and variances are used as individual feature values. By the way, in order to describe an individual feature, it is important to actively use the pitch in addition to the spectrum, which is a phonological parameter, and it is thought that the collation rate is improved. However, using an average value as an acoustic parameter that has individuality in its value and its time change pattern, such as pitch, cannot express variations within a block well. I can't say. The embodiment of FIG. 29 uses a pitch cycle histogram for each block as the pitch cycle feature quantity, so that the value of the pitch cycle and the state of its time change can be reflected in the feature quantity itself. The purpose is to improve the matching rate.

This will be described below with reference to the embodiment shown in FIG. Microphone 1
Is divided into frames every several milliseconds, and the spectrum extraction unit 121-1 and the pitch extraction unit 121-1
At -2, it is converted into a time series of the spectrum and the pitch period. As the spectrum extraction in the spectrum extraction unit 121-1, for example, after the input sound is cut by a low-pass filter (LPF) to cut off a component equal to or more than サ ン プ リ ン グ of the sampling frequency, the analog-to-digital converter converts the input sound into a discrete signal sequence. Then, this is cut out for each short-time waveform, and a Hamming window or the like is multiply windowed, and converted into a spectrum, or converted into a spectrum using a band-pass filter group to obtain time-series information. May be. Here, as in the previous embodiments, the center frequency is 250 to 630.
An example will be described in which bandpass filters of 29 channels arranged at intervals of 1/6 octave at 0 Hz are used. Further, the pitch extraction in the pitch extraction unit 121-2 may be performed by a numerical calculation such as an autocorrelation method, or by a dedicated hard board using a simple extraction method such as a combination of a low-pass filter and a counter. May be detected. The specific configuration of the pitch extraction will be described later.

 Now, the spectrum fij of the frame at the time i is fij = (fi1, fi2,..., Fi29) (j = 1, 2,..., 29), and the pitch period is represented by pi.

The voice section detection unit 122 detects a voice section from the input voice using these feature amounts. As a method of detecting a voice section from an input signal, for example, a power spectrum for each frame is calculated, this value is compared with a predetermined threshold, and a section in which power exceeds the threshold may be detected as a voice section.

The block dividing unit 123 divides each frame into a plurality of blocks using the first moment or the like of the spectrum.
In addition, block association is also performed.

Next, the feature quantity creation unit 124 calculates the feature quantity of the speaker for each block. As the feature amount, an average spectrum calculated for each block is used for the spectrum. As described above, when the spectrum of the k-th block is fkij = (fki1, fki2,..., Fki29) k: k-th block i: time in k-th block j: channel number of bandpass filter , The feature quantity Xkavr for the spectrum of the kth block is Can be expressed as Here, m indicates the number of frames in the block, and ▼ (j = 1, 2,..., 29) indicates the average spectrum for each channel.

As for the pitch, a histogram of the pitch in each block is used. When the pitch cycle of the k-th block at time i is Pki, the pitch histogram of this block counts the number of the same pitch cycle in the k-th block and sets this as a feature point PKL. Here, L is a variable corresponding to the pitch period. PKL is k
In the second block, the pitch period is L and the number is shown.

At the time of registration, the in-block average spectrum calculated for all voices is grouped for each corresponding block, and the average value and variance thereof are calculated to obtain a spectrum feature amount. As for the pitch histogram, PKL is added and averaged for each corresponding block and used as a feature amount.
The generated feature amount is added to the speaker's file name and stored in the standard pattern storage unit 125 as a standard feature amount (dictionary feature amount).

At the time of matching, the first moment of the speech section of the unknown speech and the state of the reference block of the registered speaker registered in advance are pattern-matched by a state length control type state transition model to obtain a division point, and the average is obtained for each block. A feature amount is obtained by calculating a histogram of the spectrum and the pitch period.

The distance calculation unit 126 calculates the distance between the feature amount of the unknown speaker generated by the feature amount generation unit 124 and the dictionary feature amount of the registered speaker stored in the standard pattern storage unit 125. The calculation here uses the Euclidean distance and the like.

The determination unit 127 compares the distance obtained as described above with a predetermined threshold value set in advance, and determines whether the user is the person or the impostor. Based on this determination result, the control target 2
Control.

FIG. 30 shows a processing flow of another embodiment when the pitch histogram is used as the feature in the feature generator 124. First, the extracted pitch is input (step 131), and a histogram is created for each block (step 132). Thereafter, at the time of registration, the sum of the histograms of the pitches in the corresponding blocks of all the uttered voice samples is calculated (step 133), smoothed in the pitch cycle direction (step 134), and further normalized by the total power. (Step 135) The obtained result is used as a dictionary feature amount relating to the pitch. At the time of matching, the histogram of the pitch in the block is smoothed in the pitch cycle direction (step 136), and the result is normalized by the entire frequency (step 137). That is, in the present embodiment, when the pitch histogram is used as the dictionary feature, smoothing and normalization of the histogram are introduced in order to provide stability as a dictionary.

FIG. 31 shows a specific process. This is (1) ~
This is an example in which a dictionary is created by three utterances of (3). In the figure, I
Indicates a change in spectrum, and II indicates a change in pitch.
The graph of in each utterance is a histogram of the pitch corresponding to each block. For example, focusing on a block, a process of creating a feature amount of the pitch cycle will be described. First, only a block of each utterance is extracted and added for each corresponding pitch cycle. This result is indicated by.
Next, this is smoothed in the pitch period direction (horizontal axis direction). As the smoothing process, for example, the averaging may be performed every three to five points. Further, normalization processing is performed. For example, the total pitch number may be obtained by adding the frequencies of all pitch periods, and the value may be normalized. The result is as follows, and this is used as the dictionary feature amount of the pitch.

FIG. 32 is a block diagram showing a thirteenth embodiment of the present invention.

When the distance between the feature amount of the unknown speaker and the feature amount of the registered speaker is compared with a predetermined threshold value for all blocks (such as adding the distance of each block over all blocks), Is not reflected. Further, when calculating a distance using two different dimensions, that is, a spectrum and a pitch period, each parameter is normalized by its variance, the dimension is eliminated, and then these distances are added to calculate the distance. Since both the spectrum and the pitch are used with the same weight, even if the spectral distance is large, the pitch is not so large, or conversely, the spectral distance is small and the pitch distance is large. Can not be distinguished. Therefore, in the embodiment of FIG. 32, in the calculation of the distance between the unknown speaker and the registered speaker, the spectrum and the pitch are separately compared with predetermined threshold values, and the results are integrated into a score (score). The present invention provides a method of comprehensively determining the success or failure of collation by comparing a result obtained by adding the scores over all blocks with a predetermined threshold.

In FIG. 32, a spectrum / pitch extraction section 141-1,1
The processing from 41-2 to the feature amount generation unit 144 is the same as that in FIG. 29, and the processing after the distance calculation unit 146 is different. At the time of comparison, the distance between the spectrum and the pitch is calculated by the distance calculation unit 146, and the spectrum and the pitch preset for each block are compared with a predetermined threshold of the pitch. Then, a score as shown in FIG. 33 is given depending on whether or not the spectrum and the pitch each exceed the threshold value. In FIG. 33, a mark “以内” indicates a value within the threshold value (accepted), and a mark “X” indicates a value greater than the threshold value (rejection). ◎, △, Δ, and × are given as scores. In particular,
For example, ◎ indicates 2 points, ○ indicates 1 point, Δ indicates 0 points, × indicates -10 points, and the like. For example, when the spectral distance of a certain block is within the threshold value and the pitch distance is equal to or greater than the threshold value, the score is ○. The score calculation unit 158 calculates a score for each block. The overall determination unit 159 makes a determination based on the scota for each block. At this time, all the scores are added, and if it is larger than a predetermined value, it is accepted, otherwise, it may be rejected. If one or more blocks exceed 90% of the total number of blocks, Accepted, otherwise rejected.

FIG. 34 is a block diagram of a fourteenth embodiment of the present invention. The present embodiment is basically the same as FIG. 32, except that a plurality of thresholds are set for the distance between the spectrum and the pitch for each block, and the thresholds are given a range, so that errors near the thresholds can be reduced. It is intended to be reduced. That is, when a single threshold is set for each block, the score of each block is determined by the value of the threshold, so that a pattern that fluctuates before and after the threshold is erroneously recognized. Therefore, the present embodiment aims to reduce erroneous recognition by setting a plurality of thresholds to allow a certain degree of freedom for the distance score.

Here, the score when the number of thresholds is two will be described. Now, two thresholds, a strong threshold and a gentle threshold, are set for both the spectrum and the pitch. For example, when the distance between the spectrum and the pitch is within the intense threshold, it is determined as ◎. Combinations of determination for obtaining a spectrum and a pitch in each block are as shown in FIG. The score at that time was also ◎ (2 points), ○ (1 point), △ (0 point),
It is assumed that there are four stages of × (−10 points). Now, the distance calculator 156 calculates the distance between the spectrum and the pitch,
As a result of comparing the spectrum for each block with the pitch threshold, when the spectrum is ○ and the pitch is ◎, the score becomes よ り from FIG. 35. The comprehensive determination over the entire block by the comprehensive determination unit 159 is performed in the same manner as in the case of FIG.

In the above, the description has been given of the case where a plurality of blocks are formed. Although the first moment of the spectrum is used as the feature amount of the block division, and the average spectrum and the pitch histogram are used as the individual feature amounts, an LPC cepstrum may be used instead of the spectrum.

Next, the pitch extraction unit in FIG. 29, FIG. 32, and FIG.
A specific configuration example of 121-2, 141-2, 151-2 will be described.

As mentioned earlier, the spectrum, LP
If a pitch cycle is used in addition to the C cepstrum or the like, the matching accuracy is improved. However, in general, accurate extraction of pitch requires complicated processing such as calculation of the autocorrelation coefficient of the audio signal. Therefore, if complicated processing is required, the ratio of pitch extraction in the calculation of the entire speaker verification increases, and the system scale of the entire speaker verification also increases. Therefore, a simple pitch extraction mechanism that enables highly accurate pitch extraction with a simple hardware configuration will be described. By using the simple pitch extraction mechanism, the system scale of the entire speaker verification can be made relatively small.

FIG. 36 shows an embodiment of the simple pitch extracting mechanism. This includes a plurality of comparator circuits for binarizing an audio waveform from which a high-frequency component has been removed by a low-pass filter using a preset amplitude threshold, and a circuit for generating a clock of a predetermined frequency;
A counting circuit that counts the clock circuit output and outputs the number of clocks in a section where the binarized signal rises from 0 to 1, and estimates a pitch frequency in a predetermined section using a plurality of measured count values. Things.

Hereinafter, the operation of FIG. 36 will be described. Audio signal input from the microphone, after the cut-off frequency is removing high frequency components by the low pass filter 161 of about 300 Hz, a plurality of comparators 162 _1, 162 _2, ... are input to 162 _n, the threshold Th respectively different _1, Th _2, ..., it is binarized by Th _n. Half of these comparators 162 _{1 to} 162 _n (the upper half in the figure) work on the positive side of the amplitude of the signal waveform, and the other half (lower half in the figure) work on the negative side of the amplitude. It works. The binarized signal from the positive comparator is
It becomes "1" when the amplitude exceeds the threshold value in the positive direction. The binary signal from the negative comparator becomes "1" when the amplitude exceeds the threshold value in the negative direction.

The output signals of the comparators 162 _{1 to} 162 _n , that is, the binarized signals based on the respective thresholds are respectively supplied to the synchronization circuits 163 ₁ , 163 ₂ ,
... are input to 163 _n, the reference clock synchronized generated by the clock generation circuit 164, the binary signal after synchronization counting circuit 165 _1, 165 _2, ..., and inputs to 165 _n. Each of the counting circuits 165 _{1 to} 165 _n starts counting the reference clock after clearing zero at a transition point from “0” to “1” of the binary signal input from each of the synchronization circuits 163 _{1 to} 163 _n , When the binarized signal transitions from "1" to "0", counting stops. That is, each of the counting circuits 165 _{1 to} 165 _n measures a time interval of a transition point from “1” to “0” of each binarized signal after synchronization.

The count values of the respective counting circuits 165 _{1 to} 165 _n are determined by the judging circuits 166 ₁ , 16
6 _2, ..., and inputs to the estimation circuit 167 via a 166 _n. Each of the determination circuits 166 _{1 to} 166 _n determines whether or not the count value of the corresponding counting circuit 165 _{1 to} 165 _n is within an appropriate range, and notifies the estimation result to the estimation circuit 167. For example, if the pitch period of a human voice is approximately 100 Hz to 250 Hz, and the frequency of the reference clock is 16 KHz, a valid count value is between 64 and 160, so a count value within the range is determined to be valid. be able to. Alternatively, the differences between the count values can be determined, and the count values whose differences are equal to or less than a predetermined threshold value can be determined to be appropriate. The criterion for the determination is not limited to this, but in any case, the valid range of the count value can be predicted in advance, so that the validity can be easily determined without performing complicated calculations.

The estimating circuit 167 is a circuit for estimating the pitch period or frequency of the input audio signal using the count value determined as appropriate by each of the determining circuits 166 _{1 to} 166 _n . Various algorithms can be used for this estimation. For example, an average value or a minimum value of a set of valid count values is used as the pitch period.

FIG. 37 shows a signal waveform diagram of the present embodiment. (A) shows the waveform of the reference clock, and (b) shows the waveform of the audio signal after passing through the low-pass filter 161. (C) the waveform of the binary signal by a comparator 162 _i of positive ,, (d) is a waveform after synchronization of the binary signal. (E) the waveform of the binary signal by the negative side of the comparator 162 _j, a (f) is the waveform after the synchronization. The counting circuits 165 _i and 165 _j measure the time in the section (a) to (a) of the waveform (d) or the section (c) to (d) in the waveform (f).

The configuration shown in FIG. 36 can be changed as follows. That is, the output from the low-pass filter 161 is separated into two, positive and negative, by a half-wave rectifier circuit, smoothed on each of the positive and negative sides, and its effective value (voltage) is obtained. Multiple positive and negative comparators
162 _{1 to} 162 _n are set as the amplitude thresholds, and the amplitude threshold is made variable depending on the magnitude of the input signal. As a result, the stability of pitch extraction with respect to fluctuations in the amplitude of the input voice can be improved. Further, an execution value test circuit is provided for determining whether the execution values on the positive side and the negative side obtained by the half-wave rectification circuit are appropriate, and the execution value on the positive side or the negative side becomes equal to or less than a predetermined threshold. At this time, it may be instructed to the estimation circuit 167 to prohibit the pitch extraction of the corresponding section.

FIG. 38 shows another embodiment of pitch extraction. When the input audio signal is passed through a low-pass filter to remove high-frequency components other than the pitch frequency, the cut-off frequency of the low-pass filter is
It is appropriate to set it near the pitch frequency of the input voice. However, it is difficult to uniquely determine the cut-off frequency of the low-pass filter due to the phenomenon that the pitch frequency of voice changes from time to time and the pitch frequencies of male and female voices are greatly different. In particular, in a system that is greatly affected by the cutoff frequency of the low-pass filter such as the number of zero crossings, extraction errors such as double pitch and half pitch increase. In the embodiment of FIG. 38, the cutoff frequency of the low-pass filter is dynamically changed on the time axis, thereby reducing extraction errors such as double pitch and half pitch, and enabling highly accurate pitch extraction. Things.

Hereinafter, the operation of FIG. 38 will be described. One of the input audio signals is input to a low-pass filter 171 to remove high-frequency components. The other is input to the power detection circuit (I) 172,
Find the short-term average power. Here, low-pass filter
Reference numeral 171 denotes a component whose cutoff frequency changes depending on the value of the control signal E, and can be realized by a voltage control filter by voltage control, a clocked variable switched capacitor filter, or the like.

The audio signal from which the high-frequency component has been removed by the low-pass filter 171 is input to the pitch extraction circuit 175, where the pitch frequency is obtained. This pitch extraction circuit 175 corresponds to the part except for the low-pass filter 161 in FIG. The other of the audio signals from which the high-frequency components have been removed by the low-pass filter 171 is input to the power detection circuit (II) 173, and the short-time average power is obtained. In the control circuit 174, when the output of the power detection circuit (I) 172 is A and the output of the power detection circuit (II) 173 is B, B / A becomes smaller than a predetermined B / A target value C. The control signal E is output so that the cut-off frequency of the low-pass filter 171 is increased when B / A becomes greater than C when B / A becomes greater than C.

As the control circuit 174, in the case of a low-pass filter in which the cut-off frequency increases when the control signal E of the low-pass filter 171 is increased, as shown in FIG. And the result of passing the output signal D through the integration circuit 182 may be used as the control signal E of the low-pass filter 171.

〔The invention's effect〕

According to the speaker verification method of the present invention, the following effects can be expected.

(1) By dividing a voice section into blocks, creating feature amounts using various acoustic parameters, and performing speaker verification, it is possible to incorporate information amounts of personality into feature amounts more than ever. , The matching rate is improved.

(2) As preprocessing of block division, at the time of registration, an uttered voice is divided into blocks in advance by its power and the like, and an input audio signal is divided into blocks based on a reference block obtained by this preprocessing. We can expect stable blocking and a higher matching rate.

(3) By applying the duration control type state transition model using the state of the first moment of the spectrum of the speech frame to block division, it is possible to roughly associate each block, and thereby to achieve stable block division. I can do it. Also, the amount of calculation is much smaller than block division by DP matching or the like.

(4) Since a pitch period is used in addition to a spectrum or the like as a feature amount, and the histogram is used as the feature amount of the pitch period, it is necessary to reflect the pitch period value and the state of its time change in the feature amount itself. It can be converted into an effective feature amount for speaker verification, so that the verification rate can be improved.
Further, since the histogram is smoothed and normalized as the pitch period dictionary feature, the pitch period feature can be stabilized, and the matching rate can be improved.

(5) When comparing the feature amount of the unknown speaker with the registered dictionary feature amount and determining that the unknown speaker is the same as the registered speaker when the similarity exceeds a predetermined threshold, each block is used. A predetermined threshold value is set for each distance between the spectrum and the pitch for each block, and the case is divided depending on whether the spectrum and the pitch for each block exceed the threshold value, and a score (score) is given to each of them. By taking the blocks into consideration, the spectrum and pitch distance of each block can be reflected in the overall judgment. Furthermore, a plurality of predetermined thresholds are set for each spectrum and pitch for each block, and the spectrum and the pitch are divided into cases according to the ranges of the thresholds, and a score (score) is given. Misrecognition can be reduced.

(6) By using a simple pitch extraction method using a low-frequency extraction filter and a counter for the pitch extraction, the matching rate can be improved without increasing the system scale. Also, in this case, by dynamically changing the cutoff frequency of the low-pass filter on the time axis, extraction errors such as double pitch and half pitch can be reduced, and highly accurate pitch extraction is possible. Therefore, the collation rate can be further improved.

[Brief description of the drawings]

FIG. 1 is a block diagram of a first embodiment of the present invention, FIG. 2 is a processing flowchart of a portion related to block division in FIG. 1, FIG. 3 is a diagram showing a specific example of block division according to FIG. FIG. 4 is a block diagram of a second embodiment of the present invention,
FIG. 5 is a diagram showing a specific example of block division according to FIG. 4,
6 is a block diagram of a third embodiment of the present invention, FIG. 7 is a diagram showing a specific example of block division according to FIG. 6, FIG. 8 is a block diagram of a fourth embodiment of the present invention, FIG. FIG. 8 is a diagram showing a specific example of block division according to FIG. 8, and FIG.
FIG. 11 is a block diagram of an embodiment, FIG. 11 is a processing flowchart of a threshold adjusting unit in FIG. 10, FIG. 12 is a diagram showing a specific example of block division according to FIG. 10, and FIG. 13 is a diagram of a sixth embodiment of the present invention. FIG. 14 is a block diagram showing a processing procedure of an embodiment of the block division pre-processing and block division in FIG. 13;
15 is a diagram showing a processing procedure of another embodiment, FIG. 16 is a diagram showing a specific example corresponding to FIG. 15, FIG. 17 is a diagram showing a processing procedure of still another embodiment, FIG. 18 is a diagram showing a specific example corresponding to FIG. 17, FIG. 19 is a diagram showing a processing procedure of still another embodiment, FIG. 20 is a block diagram of the seventh embodiment, and FIG. FIG. 22 is a diagram showing a specific example of block division according to the drawing, FIG. 22 is a flowchart showing an example of a setting procedure of a division point in FIG. 20, and FIG. 23 is a duration length control type state transition using a first moment of a spectrum of an input voice. FIG. 24 is a diagram illustrating a model, FIG. 24 is a diagram showing a result of subdivision with respect to FIG. 21, FIG. 25 is a block diagram of an eighth embodiment of the present invention, and FIG. 26 is a block diagram of a ninth embodiment of the present invention. FIG. 27 is a block diagram of a tenth embodiment of the present invention, FIG. 28 is a block diagram of an eleventh embodiment of the present invention, and FIG. FIG. 30 is a block diagram of the embodiment, FIG. 30 is a processing flow chart for explaining another embodiment of the feature amount generation unit in FIG. 29, FIG. 31 is a diagram showing a specific example of the processing in FIG. FIG. 33 is a block diagram of a thirteenth embodiment of the present invention, FIG. 33 is a diagram showing a specific example of scores in FIG. 32, FIG. 34 is a block diagram of a fourteenth embodiment of the present invention, and FIG.
FIG. 34 is a diagram showing a specific example of the score in FIG. 34, FIG. 36 is a diagram showing an embodiment of the pitch extraction, FIG. 37 is a signal waveform diagram of each part in FIG. 36, and FIG. FIG. 39 is a diagram showing a specific example of the control circuit in FIG. 38. 10 microphone, 2 control object, 11 acoustic parameter conversion unit, 12 voice section detection unit, 13 voiced / silent discrimination unit,
14: voiced section extraction unit (block division unit), 15: feature amount generation unit, 16: standard pattern storage unit, 17: distance calculation unit, 18: determination unit, 63, 74: pre-block division processing unit, 75 ... Reference block setting unit, 76 ... re-division processing unit, 121-1, 141-1, 151-1
... Spectrum extractor, 121-2, 141-2, 151-2. Pitch extractor, 161,171... Low-pass filter.

──────────────────────────────────────────────────続 き Continued on front page (31) Priority claim number Japanese Patent Application No. 63-217585 (32) Priority date August 31, 1988 (33) Priority claim country Japan (JP) (31) Priority Claim Number Japanese Patent Application No. 63-265054 (32) Priority Date October 20, 1988 (33) Priority Country Japan (JP) (56) References JP-A-60-48098 (JP, A) JP-A-56-138798 (JP, A) JP-A-61-138299 (JP, A) JP-A-56-158386 (JP, A) JP-A-60-238897 (JP, A) JP-A-61-21499 (JP) JP, A) JP-A-56-119198 (JP, A) Yasunaga Niimi, "Information Science Course E.19.3 Speech Recognition," Kyoritsu Shuppan Co., Ltd. (Showa 54) 211, l. 9 (58) Field surveyed (Int.Cl. ⁶ , DB name) G10L 3/00 531 G10L 3/00 513 G10L 3/00 515 G10L 9/00 301 JICST file (JOIS)

Claims (21)

(57) [Claims] 1. A feature amount registered by a registered speaker is compared with a feature amount input by an unknown speaker, and when the similarity exceeds a certain threshold, the unknown speaker is identified as the registered speaker. In the speaker verification method, the input voice signal is divided into frames every predetermined time unit,
A means for converting to a sound parameter for each frame; a means for detecting a voice section from an input voice signal and dividing the voice section into a plurality of blocks on the time axis according to the value of the voice parameter; Means for generating a speaker-specific feature amount, registering the feature amount for each speaker at the time of registration, and comparing the registered feature amount with the feature amount of an unknown speaker at the time of collation; Is divided into voiced, unvoiced, and unvoiced sections by using the power, the rate of change of the spectrum, and the pitch rate of the input voice to extract a voiced section, and divides this into one block. Matching method. Means for dividing an input audio signal into frames for each predetermined time unit and converting the input audio signal into audio parameters for each frame; detecting an audio section from the input audio signal; Means for dividing into a plurality of blocks, and a speaker-specific feature amount is generated for each block, and the feature amount is registered for each speaker at the time of registration, and the feature amount of an unknown speaker and the registered feature are registered at the time of verification. Means for comparing the feature amount registered by the registered speaker with the feature amount input by the unknown speaker when the similarity exceeds a certain threshold. In the speaker verification method which determines that the voice is the same as the speaker, at the time of registration, there is provided a pre-processing means for dividing the uttered voice into blocks in advance according to its power, power dip, spectrum change rate, etc. A speaker verification method wherein input speech is divided into blocks based on a reference block. 3. The method according to claim 1, wherein only the first voice input at the time of registration is divided into blocks using power, power dip, spectrum change rate, and the like of the voice.
3. The speaker verification method according to claim 2, wherein the voice segment of the second and subsequent input voices is divided into blocks. 4. Using the first voice at the time of registration as a reference pattern, perform DP matching with the second and subsequent voices, and perform voice sections of the second and subsequent voices corresponding to the first voice block. 4. The speaker verification method according to claim 3, wherein is divided as a block. 5. A feature amount for each block is generated by, for example, adding and averaging acoustic parameters in each block of a first speech at the time of registration in a time axis direction, and using this as a reference pattern,
Using a duration control type transition model with the second and subsequent voices and the like, dividing the voice section of the second and subsequent voices corresponding to the block of the first voice into blocks. The speaker verification method described in (3). 6. The second and subsequent voices are divided in advance into a plurality of semi-blocks according to the power, power dip, spectrum change rate, etc. of the voice, and the feature amount in the semi-block is calculated. 4. The speaker verification method according to claim 3, wherein the block division is performed by DP matching with the in-block feature amount of the first speech. 7. All the voices uttered at the time of registration are divided into respective blocks by using their power, power dip, spectrum change rate, etc., the representative number of blocks is obtained, and registration is performed in accordance with the number of blocks. 3. The speaker verification method according to claim 2, wherein the number and the input voice at the time of verification are divided into blocks. 8. A method for performing DP matching with a voice having a representative number of blocks as a reference pattern and a voice having a number of other blocks to divide a voice section corresponding to a representative voice block into blocks. The speaker verification method according to claim 7, characterized in that: 9. A feature amount for each block is generated by, for example, adding and averaging acoustic parameters in each block of a voice indicating a representative number of blocks in a time direction, and using these as a reference pattern, a duration time between the block and another voice. The speaker verification method according to claim 7, wherein block division corresponding to a representative block is performed using a length control type state transition model or the like. 10. An input speech is divided in advance into a plurality of semi-blocks according to speech power, power dip, spectrum change rate, etc., and a feature amount in the semi-block is calculated. The speaker verification method according to claim 7, wherein the block division is performed by DP matching with a feature amount for each block of the indicated voice. 11. A means for dividing an input audio signal into frames for each predetermined time unit, converting the input audio signal into acoustic parameters for each frame, detecting an audio section from the input audio signal,
A means for dividing this voice section into a plurality of blocks on the time axis, a speaker-specific feature quantity is generated for each block, the feature quantity is registered for each speaker at the time of registration, and an unknown speaker at the time of verification. And a means for comparing the registered feature quantity with the feature quantity registered by the registered speaker, and when the similarity between the feature quantity input by the unknown speaker exceeds a certain threshold. In a speaker verification method in which an unknown speaker is determined to be the same as the above registered speaker, a spectrum is used as an acoustic parameter. , Each block is divided into a plurality of blocks, and the block of the optimum voice is used as a reference block, and all the input voices at the time of registration are re-based based on the reference block. A speaker verification method characterized by dividing an unknown voice into blocks based on a reference block created at the time of previous registration, at the time of verification, in addition to block division. 12. As a reference block at the time of registration, all voices uttered at the time of registration are divided into a plurality of blocks by comparing the first moment of the spectrum with a predetermined threshold,
The speaker verification method according to claim 11, wherein the block of the voice having the smallest number of divisions is used as a reference block. 13. An average value of the first moment of each of the reference blocks adopted at the time of registration is obtained, and the average of the first moment is used as the state of each block. The speaker verification system according to claim 11, wherein the speaker is divided into blocks according to a model. 14. An input audio signal is divided into frames for each predetermined time unit, and means for converting the input audio signal into acoustic parameters for each frame;
A means for dividing a voice section into a plurality of blocks on a time axis, a feature amount unique to a speaker is generated for each block, the feature amount is registered for each speaker at the time of registration, and an unknown speaker is And a means for comparing the registered characteristic amount with the registered characteristic amount for each corresponding block. The similarity between the characteristic amount registered by the registered speaker and the characteristic amount input by the unknown speaker is determined by a threshold value. In the speaker verification method in which it is judged that the unknown speaker is the same as the unknown speaker when the number exceeds, an average value in the time axis direction of the spectrum for each frame in the block (average spectrum in the block) and a least square of the average spectrum A speaker verification method characterized in that the inclination of a straight line is used as a feature value of a block. 15. A means for dividing an input audio signal into frames for each predetermined time unit and converting the input audio signal into acoustic parameters for each frame, detecting an audio section from the input audio signal,
A means for dividing a voice section into a plurality of blocks on a time axis, a feature amount having a speaker tag for each block is generated, the feature amount is registered for each speaker at the time of registration, and an unknown Means for comparing the feature amount of the speaker and the registered feature amount for each corresponding block, and the similarity between the feature amount registered by the registered speaker and the feature amount input by the unknown speaker, In the speaker verification method of determining that the unknown speaker is the same as the unknown speaker when the threshold value is exceeded, the positions of the peaks and valleys of the spectral outline of each frame in the block are extracted, and the number of appearances of the peaks and valleys in the block is determined. A speaker verification method characterized in that the accumulated number is obtained for each frequency and the accumulated number is used as a feature amount in a block. 16. An input audio signal is divided into frames for each predetermined time unit and converted into acoustic parameters (spectrum, pitch period, etc.) for each frame, and an audio section is detected from the input audio signal. Means for dividing the voice section into a plurality of blocks on the time axis, and generating a speaker-specific feature amount using the acoustic parameters for each block;
At the time of registration, the feature amount is registered as a dictionary for each speaker, and at the time of matching, there is means for comparing the feature amount of the unknown speaker with the registered dictionary feature amount, and the feature amount input by the unknown speaker is provided. When the similarity between the registered speaker and the dictionary feature registered by the registered speaker exceeds a predetermined threshold, the pitch period is used in a speaker verification method in which the unknown speaker is determined to be the same as the registered speaker. A speaker verification method using a pitch period histogram for each block as a speaker-specific feature amount. 17. When creating a feature quantity relating to a pitch cycle for each block, at the time of registration, a sum of histograms of pitches in the corresponding blocks of all uttered speech samples is calculated, and this is smoothed in the pitch cycle direction. 17. The method according to claim 16, wherein a value normalized by the entire power is used as a dictionary feature amount relating to the pitch, and at the time of matching, a histogram of the pitch in the block is smoothed in a pitch period direction. Speaker verification method. 18. An input audio signal is divided into frames for each predetermined time unit, a means for converting the input audio signal into acoustic parameters for each frame, a voice section is detected from the input voice signal, and the voice section is detected on the time axis. A means for dividing into a plurality of blocks and a speaker-specific feature amount are generated using the acoustic parameter for each block, and the feature amount is registered as a dictionary for each speaker at the time of registration, and an unknown speaker is registered at the time of verification. And a means for comparing the registered feature with the dictionary feature registered by the unknown speaker. The similarity between the feature input by the unknown speaker and the dictionary feature registered by the registered speaker has a predetermined threshold value. In the speaker verification method in which the unknown speaker is judged to be the same as the registered speaker when it exceeds, using the spectrum and pitch as speaker-specific features, the features of the unknown speaker and the registered dictionary features Quantity and compare When it is determined that the unknown speaker is the same as the registered speaker when the similarity exceeds a predetermined threshold, a predetermined threshold is set for each distance between the spectrum and the pitch for each block, and the spectrum for each block is set. And the pitches exceed the threshold, respectively, and a score (score) is given to each of them, and this score is individually compared with a predetermined threshold, or a value added over all blocks is added to a predetermined threshold. Or comparing the unknown speaker with the registered speaker to determine whether the unknown speaker is the same as the registered speaker. 19. A plurality of thresholds are set for the distance between the spectrum and the pitch for each block, and a score is given for each block depending on which range of the plurality of thresholds the distance between the spectrum and the pitch is. The speaker according to claim 18, wherein the score is added over all blocks and compared with a predetermined threshold to determine whether the unknown speaker is the same as the registered speaker. Matching method. 20. A method of extracting a pitch period, comprising the steps of: passing an input audio signal through a low-pass filter to remove high-frequency components; binarizing the input audio signal with a plurality of thresholds; The speaker verification method according to any one of claims (16) to (19), wherein the time interval between points is measured, and the pitch period or the frequency is estimated from the measured time interval. 21. When extracting a pitch period, the low-pass filter is cut off so that the ratio between the short-time average power of the input audio waveform and the short-time average power of the audio waveform passed through the low-pass filter becomes constant. The speaker verification method according to claim 20, wherein the frequency is changed.