JPH01309099A - Speech responding device - Google Patents

Abstract

PURPOSE:To maintain a high recognition rate without any deviation between unspecified recognition and specified recognition by providing unspecified standard patterns as to words whose registration contents are already known. CONSTITUTION:Candidates up to high-order (n) of the unspecified recognition are denoted as WI1, WI2... WIn and their similarity data are denoted as SI1, SI2... SIn; and candidates up to high-order (n) of the specified recognition are denotes as WD1, WD2... WDn and their similarity data are denoted as SD1, SD2... SDn. Then, (kXSI1) and SD1 are compared by using a coefficient (k) which is found by experiment in advance, thereby obtaining WI1 when (kXSI1)> SD1 or WD1 when SD1>(kXSI1) as the final recognition result of an unspecified/specified decision part. Consequently, the deviation between the unspecified recognition and specified recognition is eliminated and the high recognition rate is maintained.

Description

[Detailed Description of the Invention] The present invention relates to a voice response device, more specifically a voice recognition device applicable to a voice dialing device, a voice word processor, other voice response devices, etc. The present invention relates to a device in which control of an external peripheral device connected to a control recognition processing section can be realized by a central processing section that performs recognition processing control.

It is known that voice registration in specific voice recognition devices that have been developed and devised in the past is a burdensome task for users. Therefore.

In order to reduce the amount of language used, device command terms and general names are registered in advance as unspecified standard patterns, and recognized at the same time as specific marks and patterns registered by the user. There are possible ways. however,
Current speech recognition technology cannot remove the influence of a person's individuality from a standard pattern of a person. It is difficult to perform recognition by treating standard patterns for unspecified speakers as equivalent, based on the utterances of speakers, and eliminating the influence of individuality. The method of comparing the similarities between the top recognition results is not a comparison of equivalent similarities, so candidates registered in unspecified standard patterns are likely to be selected, or candidates registered in specific standard patterns are likely to be selected. Either unspecific recognition or specific recognition, such as the candidate that is more likely to be selected! ! The drawback was that there was a high probability that the results would be biased towards the results of the study.

Furthermore, conventional voice dialing devices only have standard patterns for specific speakers, and the numbers rllOJ and rl19J are the same no matter where they are used.

r104J, rL'l'iJ, etc., and the corresponding "
? Standard patterns for recognition and synthesis such as ``Yoshiken'', ``Fire Department'', ``Number Guide'', and ``Weather Forecast'' had to be registered for each speaker, which was a hassle.

Further, since the amount of calculation required for recognition processing is enormous, conventional recognition processors have been configured as so-called recognition-only processors that can only implement recognition functions. therefore,
Traditional recognition-only processors with external peripherals, e.g.
If you try to connect a facsimile device or voice dialing device, etc.

Unable to connect directly to recognition-only processor.

A control central processing unit (Cl) that bridges control between the recognition-dedicated processor and these external peripheral devices
It was necessary to intervene.

Further, the conventional speech recognition LSI requires a huge amount of recognition calculations, so it can only perform the recognition function. For example, in the case of matching using dynamic programming (DP), the computation of grid points for matching is huge, so in practical use it cannot be performed by processing only on a general-purpose CPU, so it requires large-scale dedicated hardware. clothing is required. However, this dedicated hardware (or dedicated processor) cannot perform functions other than matching, such as registering dictionary patterns, output recognition results, or perform peripheral control related to recognition, or connect with external peripheral control. Therefore, in these cases, a general-purpose CPU must be connected. Also, in speech recognition LSIs, the recognition algorithms are generally different between LSIs for specific speaker recognition and LSIs for non-specific speaker recognition, and arithmetic processors are required. Both of them have to be dedicated, and it is necessary to select and use the respective recognition LSIs depending on the purpose of specific speaker recognition or non-specific speaker recognition.

In addition, conventionally, the similarity between feature patterns and dictionary patterns (
or distance), - generally, DP
As typified by processors, these are large-scale dedicated processors, and it has been impossible to include a CPU and an arithmetic unit in one LSI.

Furthermore, there is a method of using multi-templates in order to accommodate both specific speakers and non-specific speakers using conventional arithmetic processors. This is a method of having multiple dictionary patterns for one recognition result. However, in order to perform speaker-independent recognition using this method, it is necessary to have about 40 to 60 dictionary patterns for one recognition result, which has the disadvantage that the recognition process takes time.

In addition, the previously developed BTS P (Binar
yTimeSpectruffiPattern) method, the feature patterns are the same in recognition of specific speakers and non-specific speakers, but the dictionary pattern is
Each had a different configuration and could not be calculated using the same calculation unit.

The present invention has been made in view of the above-mentioned circumstances.
The invention described in claims 1 and 2 of the present application uses speaker-independent speech recognition for terms such as command terms and general names, and uses speaker-specific speech recognition for other words. In some cases, candidates registered in unspecified standard patterns are more likely to be selected, or candidates registered in specific marks or patterns are more likely to be selected. This eliminates the bias between unspecified recognition and specific recognition, and maintains a high recognition rate. This was done with the purpose of providing a speech recognition device that can perform the following tasks. Furthermore, the invention described in claim 3 is made for the purpose of saving the trouble of registering public institution names that everyone uses in a voice dialing device that dials a telephone by voice. It is. Furthermore, in the invention described in claims 4 and 5, when recognizing the characteristic pattern of input data, an external peripheral device is directly connected to the control recognition processing unit without going through a control central processing unit (CPU). This was done for the purpose of providing a recognition processing device that can be connected. Furthermore, in speech recognition, whether specific speaker recognition or non-specific speaker recognition is performed,
The control recognition processing unit can be shared and calculations can be executed at high speed! 5! providing an Ia device;
The present invention further provides a recognition processing device in which detection of a valid section of a characteristic pattern of input data can be variably controlled from outside the control recognition processing section.

In the recognition processing device, this is done for the purpose of reducing data loss when comparing characteristic patterns of input data and lightening the calculation load on the central processing unit.

Figure 1 is a configuration diagram for explaining one embodiment of the present invention.
FIG. 2 is a flowchart for explaining the operation, and this embodiment relates to a device applicable to a voice recognition device, such as a voice dialing device, a voice word processor, and other voice response devices.

It is known that voice registration in specific voice recognition devices that have been developed and devised in the past is a burdensome task for users. Therefore, in order to reduce the number of registered terms as much as possible, it is recommended that device command terms and general names be registered in advance as unspecified standard patterns, and then recognized at the same time as the specific standard patterns registered by the user. Conceivable. However, current speech recognition technology cannot remove the influence of a person's individuality from a person's standard pattern. Therefore, the standard pattern for a specific speaker, which includes the influence of individuality, and the standard pattern for non-specific speakers, which is based on the utterances of many speakers and eliminates the influence of individuality, are treated as equivalent. It is difficult to perform recognition, and furthermore, unspecified speech recognition and specific speech recognition are performed separately.

How to compare the similarity between the top recognition results?

Because it is not an equivalent comparison of similarity, candidates registered in unspecified standard patterns are more likely to be selected, or candidates registered in specific standard patterns are more likely to be selected. The disadvantage is that there is a high probability that the recognition result will be biased towards one of the two.

In this example, when using speaker-independent voice recognition for words such as command terms and general names, and using speaker-specific voice recognition for other words, the candidates registered in the unspecified standard pattern are The purpose of this invention is to provide a speech recognition device that can maintain a high recognition rate by eliminating bias between unspecified recognition and specific recognition, such as candidates being more likely to be selected or candidates registered with specific standard patterns being more likely to be selected. It is something that

The invention set forth in claims 1 and 2 of the present application is for achieving the above object, and the invention set forth in claim 1 is a feature extraction means for extracting feature amounts of input speech. and,
unspecified standard pattern storage means for storing preregistered unspecified speaker standard patterns; specific standard pattern storage means for storing standard patterns registered by the user; and the unspecified standard patterns and the specified standard patterns. recognition means for performing recognition by comparing each of the voice features with the feature amount of the voice extracted by the feature extraction means; recognition result confirmation means for making the comparison by the recognition means and informing the user of the recognized recognition results; a coefficient storage means for storing at least one predetermined predetermined coefficient; a value corresponding to the degree of similarity to either the unspecified standard pattern or the specific standard pattern obtained by the recognition means; The leather for which the coefficient has been calculated is compared with the value corresponding to the degree of similarity to the standard pattern for which the calculation has not been performed among the unspecified standard pattern or the specified standard pattern, and the value corresponding to the maximum value of the degree of similarity is compared. a that yields the value
and recognition means for recognizing the quasi-pattern. The invention of claim 2 also provides a voice recognition method in which the first-order St
ep, that is, the feature amount of the input voice is extracted by the feature amount extraction means; the unspecified speaker standard pattern registered in advance and the specific standard pattern registered by the user, and the feature amount extraction means The extracted voice features are compared and recognized by a recognition means; a value is predetermined as a value corresponding to the degree of similarity to either the unspecified standard pattern or the specific standard pattern obtained by the recognition means. and at least one predetermined coefficient; calculate the calculated value and a value corresponding to the degree of similarity to the standard pattern on which the calculation was not performed among the unspecified standard pattern or the specified standard pattern; Comparing: constitutes recognizing a standard pattern for which a value corresponding to the maximum similarity is obtained by the comparison. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a configuration diagram for explaining one embodiment of the present invention,
Fig. 2 is a flowchart for explaining the operation. In Fig. 1, 1 is a microphone, 2 is a feature extraction section, 3 is a non-specific speaker recognition section, 4 is a specific speaker recognition section, and 5 is a non-specific speaker recognition section. standard pattern storage section.

6 is a specific speaker standard pattern storage section, 7 is an unspecified/specific speaker determining section, 8 is a keyboard, and 9 is a coefficient section.

10 is a speech synthesis section, 11 is a speech synthesis dictionary section, 12 is a speaker, and the feature amount of the voice extracted by the specific extraction section 2 corresponds to an unspecified speaker registered in advance in the unspecified standard pattern storage section 5. Pattern recognition is performed with the unspecified standard pattern, and the result becomes the result of the unspecified recognition unit. Further, in parallel with this, pattern recognition is performed on the feature amount with the specific standard pattern registered by the user in the specific standard pattern storage section 4, and the result becomes the result of the specific recognition section.

Here, the top n candidates for unspecified recognition are WI 1
, WI 2. -, WIn, and their similarities are respectively SII and SI2. −, SIn, and the top n candidates for specific recognition are WD 1,
WD 2 ,−, WD n their similarity, respectively.

S D 1 , S D 2 , − , S D n , using a coefficient k determined experimentally in advance, (
A comparison is made between kXs I l) and SDl, and (
If kXS I 1)>S01, WII is SDI>
(kXsll), WDl becomes the final recognition result of the unspecified/specific discriminator.

(Step ~ 4) Also, when the next candidate is required, similarly, from among the unspecified top n similarity values multiplied by the coefficient k, and the specific top n similarity values, , are selected in descending order.

Note that if the coefficient k is one fixed value, the recognition result of the unspecified/specific discriminator may vary depending on factors such as the registration environment of the specific standard pattern, the recognition environment, the characteristics of the user's voice, and vocalization fluctuations. It may be biased toward specific or specific candidates. Therefore, the coefficient actt? li [Several coefficients are prepared, and in order to correct this bias, the user manually sets the coefficient k to an appropriate value using a switch or a command.

In addition, in the embodiment described below, the unspecified degree of similarity is multiplied by the coefficient k and compared with the specified degree of similarity, but conversely,
Even if a specific similarity is multiplied by a coefficient and compared with an unspecified similarity, the result is the same. In this case, 5DiXk and SD
Since the magnitude with i will be compared, the relationship between 7 and k will be /=1/.

Further, the recognition result confirmation means notifies the user of the recognition result by displaying it or by voice synthesis.

When the user performs an action indicating that the recognition result is incorrect by key input or voice input (in Fig. 1, the voice synthesis section, the voice synthesis section, and the keyboard are (The speech synthesis dictionary unit may use each recognition standard pattern.) The device outputs the next candidate. In this recognition result confirmation means, when the user cancels the recognition result included in the unspecified standard pattern, the next candidate is further output. The process is executed sequentially by canceling and outputting the next candidate, and the desired candidate is selected (Steps 5 to 7), but in this case,
Only when the recognition result included in the unspecified standard pattern is canceled and the recognition result included in the specific standard pattern is adopted, the value of the coefficient is set to a certain experimentally determined value,
Or, if the value calculated by a certain formula is reduced, and conversely, the recognition result included in the specified standard pattern is canceled, and the recognition result included in the unspecified standard pattern is adopted. automatically by increasing it by some experimentally determined value or by a value calculated by some experimentally determined formula.
Always keep the coefficient k at an appropriate value.

The next candidate is output, and the user cancels the recognition result included in the unspecified standard pattern and uses the next candidate again.
When the recognition result included in the unspecified standard pattern is adopted, the value of the coefficient is not changed. Conversely, even if the next candidate is output and the user cancels the recognition result included in the specific standard pattern and then adopts the recognition result included in the specific standard pattern again in the next candidate, the coefficients will not be affected. Do not change the value.

As is clear from the above description, according to the embodiment shown in FIGS. 1 and 2, +3 words (or
By providing an unspecified standard pattern for other utterance units), it saves the user the trouble of registering the word.

To maintain a high recognition rate by eliminating bias between unspecified recognition and specific recognition, such as candidates registered in unspecified standard patterns being more likely to be selected, or candidates registered in specific standard patterns being more likely to be selected. I can do it.

FIG. 3 is a diagram for explaining another embodiment of the present invention,
This embodiment relates to a voice dialing device. Conventionally,
Voice dialing devices only have a standard pattern for a specific speaker, and the rll number is the same no matter where it is used.
oJ, rl19”.

rl, 04J, r177J, etc., and the corresponding standard patterns for recognition and synthesis such as "police", "fire department", "directory assistance", "weather forecast", etc. must also be registered for each speaker. was there.

This embodiment was developed in view of the above-mentioned circumstances, and was designed to save the trouble of registering public institution names used by everyone in a voice dialing device that performs telephone dialing by voice. It is something.

The invention of claim 3 of the present application is for achieving the above object, and the invention of claim 3 is a feature extraction method for extracting feature amounts of voice input from a handset for transmitting and receiving telephone calls. means, standard pattern storage means for storing the feature extracted by the feature extraction means, number storage means for storing a dial number corresponding to input voice, and features extracted by the feature extraction means at the time of voice input. a pattern matching means for comparing the quantity with a feature quantity stored in the standard pattern storage means stored in advance, and recognizing which standard pattern the input voice corresponds to;

dial signal output means for reading a dial number corresponding to the recognition result recognized by the pattern matching means from the number storage means and outputting a dial signal; and an unspecified standard pattern of a voice of a name corresponding to a specific telephone number A standard pattern storage means for unspecified speakers is provided as a standard pattern for speakers.

An explanation will be given below based on one embodiment of the present invention.

FIG. 3 is a block diagram for explaining another embodiment of the present invention, in which 21 is a transmitter, 22 is a receiver, 23 is a communication circuit, 24 is a hook switch, 25 is a feature extractor, 26 is a keyboard, 27 is a control unit, 28 is a pattern matching unit, 2
9 is a standard pattern storage unit for recognition; 30 is a speech synthesis unit;
31 is a standard pattern storage unit for synthesis, 32 is a telephone number storage unit, 33 is a calling (4 control unit, 34 is an incoming call control unit, 35 is a line control unit, 3G is a central office line, and the voice input from the transmitter is The feature h (is extracted by the feature extractor. Depending on the state of the keyboard, the controller determines whether to use this feature as a standard pattern for recognition or to perform pattern matching.4) The marker pattern registration button on the keyboard When this feature is not pressed, matching with the recognition mark I (I! pattern) stored in the pattern matching section is performed.
The one with the highest degree of similarity is selected as the recognition result. This result is sent to the call control section, and the telephone number, which has been registered in advance with the keyboard, is read out from the number storage section and a call is made. In addition, the above recognition results are sent to the speech synthesis section,
A standard pattern for synthesis, which has been associated and registered in advance, is read out and output to the receiver as a synthesized sound. At this time, a standard pattern for recognition and a standard pattern for synthesis corresponding to it,
The number storage section has a structure as shown in FIG.

In FIG. 4, (A) is a standard pattern storage unit for synthesis;
(B) is a standard pattern storage area for recognition, (C) is a phone number storage area, (1) is a non-writable area, and (■) is a writable area. , public institution telephone numbers riioJ, rL19, etc. are stored, and a plurality of corresponding standard patterns for unspecified speakers are stored for one telephone number, as well as standard patterns for synthesis. are stored one by one, and this area cannot be written to. Each device also has a writable area, where the user's registered pattern and phone number are written.

Since some standard patterns for non-specific use may not be used by users, those that are not to be matched are selected using the keyboard.

Next, the time of registration will be explained. When the standard pattern registration button on the keyboard is pressed, the feature amount of the input voice extracted by the feature extraction section is registered as a standard pattern for recognition and a standard pattern for synthesis.

At this time, the telephone number corresponding to this input voice is registered in the number storage section using the keyboard. Registration of a telephone number is performed by inputting the telephone number into the number storage section 32 using the numeric keys on the keyboard. This input operation using the numeric keypad is also a key dial operation performed when making a call.

Therefore, the telephone number registration operation (work) must be performed in an on-hook state to prevent the number from being called to the central office line.The on-hook state means that the oscillation output of the telephone number information is connected to the central office line. It is in a state where it has not been
The telephone detects an incoming call from the central office line and is ready to make a call. The registration modes include a recognition standard pattern registration mode, a synthesis standard pattern registration mode, and a telephone number registration mode. The operation will be explained with reference to FIGS. 3 and 4. When the system is set to the registration mode by operating a registration mode key for setting one of the registration modes, the system enters an on-footer state. These registration modes may be set independently for each key, or multiple modes may be set simultaneously using a single key.

The part having speech synthesis including the synthesis memory 31 uses a method that stores the sound as it is and plays it back, a so-called decomposition and synthesis method as in this embodiment, and a method that compresses and stores the sound waveform and then decompresses and plays it back. Either method can be applied.

As is clear from the above explanation, according to the embodiment shown in FIGS. 3 and 4, it is possible to save the trouble of registering telephone numbers of public institutions, etc. that everyone uses, or standard voice patterns. Registration work is reduced.

5 to 10 are diagrams for explaining another embodiment of the present invention, in which a central processing unit performs recognition processing control for controlling external peripheral devices connected to a control recognition processing unit. This invention relates to a device that can be realized in the department.

Conventional recognition processors require a huge amount of calculation for recognition processing. It was configured as a so-called recognition-only processor that could only implement the i1 function.

Therefore, if you try to connect an external peripheral device, such as a facsimile machine or a voice dialing device, to a conventional recognition-only processor, it will not be possible to connect it directly to the recognition-only processor, and the connection between the recognition-only processor and these external peripherals will be limited. It was necessary to interpose a central processing unit (CPU) for control in order to bridge control between the two.

Furthermore, since the amount of recognition calculation required by conventional speech recognition LSIs is enormous, only the recognition function can be executed. For example, in the case of matching using dynamic programming (DI).

Since the computation of grid points for matching is huge, in practice it cannot be performed only by general-purpose CI'lU processing.
Requires extensive specialized hardware.

However, this dedicated hardware (or dedicated processor)
In addition, it is impossible to perform functions other than matching, such as dictionary pattern registration, output of recognition results, or peripheral control related to recognition, or connection with external peripheral control.
In these cases, a general-purpose CPU must be connected,
In addition, in LSI for voice recognition, LS for specific speaker recognition
I and LSI for speaker-independent recognition generally have different recognition algorithms, and the processors for both must be dedicated, depending on the purpose of specific speaker recognition or speaker-independent recognition. Each recognition LSI must be selected and used.

In addition, conventionally, the similarity between feature patterns and dictionary patterns (
or distance), - generally, DP
As represented by processors: - They are large-scale dedicated processors, and it was impossible to include a CPU and an arithmetic unit in one LSI.

Furthermore, there is a method of using multi-templates in order to accommodate both specific speakers and non-specific speakers using conventional arithmetic processors. This is a method of having multiple dictionary patterns for one recognition result. However, in order to perform speaker-independent recognition using this method, it is necessary to have about 40 to 60 dictionary patterns for one recognition result, which has the disadvantage that the recognition process takes time.

In addition, the previously developed BTS P (Binar
Even in the calculation of the y TimeSpectra+Pattern method, the feature patterns are the same in recognition of specific speakers and non-specific speakers, but the dictionary pattern is
Each had a different configuration and could not be calculated using the same calculation unit.

This embodiment has been developed in view of the above-mentioned actual situation, and especially when recognizing characteristic patterns of input data,
An external peripheral device is connected to the control recognition processing section by the control central processing section (C
The purpose of this invention is to provide a recognition processing device that can be directly connected without going through a PU. Furthermore, it is an object of the present invention to provide a recognition device which can share a control recognition processing section and can perform calculations at high speed when performing either specific speaker recognition or non-specific speaker recognition in speech recognition. Furthermore, in the recognition processing device, it is possible to variably control detection of a valid section of a characteristic pattern of input data from outside the control recognition processing section.

In the recognition processing device, this is done for the purpose of reducing the amount of data when comparing characteristic patterns of input data and lightening the calculation load on the central processing section.

The invention described in claims 4 and 5 of the present application is for achieving the above object, and the invention of claim 4 extracts feature amounts of input data and generates a feature pattern. and a feature pattern generation unit.

a matching unit that matches the feature pattern with a dictionary pattern;
and a control recognition processing unit having a central processing unit (CPU) capable of controlling the generation unit and the verification unit as well as controlling required external peripheral devices in conjunction with the generation unit, and controlling the central processing unit. and a data storage section that stores data necessary for collation and control, and each of the generation section, control recognition processing section, program storage section, and data storage section It is configured to be removably connected to the transfer path (CPU BUS) of the central processing unit. Further, the invention according to claim 5 uses a feature pattern input section for inputting a feature pattern, a dictionary pattern input section for inputting a dictionary pattern, and output data of the feature pattern input section and the dictionary pattern input section. a calculation unit that calculates at least one element necessary for pattern matching; a frame data register that stores one frame worth of the element necessary for pattern matching; and a data register that doubles the data stored in the dictionary pattern input unit. and a frame data addition unit that adds the data of the calculation unit and the data of the calculation unit.

The pattern matching circuit includes a word data register that stores values of elements necessary for the pattern matching, and a word data adder that adds values of one frame of the elements stored in the word data register.

Hereinafter, the present invention will be explained based on examples.

) A system using processor 1 and processor 2.
1. FIG. 5 is a diagram showing an example of the configuration of a basic speech recognition system, which consists of a processor 41, a processor 42, a program ROM 43, and a template RAM 44.
They are connected via U bus 45.

The processor 41 uses an LS for extracting features of the voice.
I outputs the power spectrum and BTSP at a sampling period of 10 m5. First, the functions of each part within the processor 41 will be explained.

- The microphone amplifier section 51 amplifies the audio signal input from the microphone 47.

- The low-pass filter section 52 cuts high frequencies to remove sampling aliasing noise.

- AGG (auto gain control) amplifier and pre-emphasis block 53 adjusts the audio level to an appropriate range and further emphasizes the high range to compensate for the power of the high range of the audio.

・Band pass filter, detector, and low pass filter block 55 have sharpness Q=6 and middle 6 frequency 250
1/3 oct from Hz to 6.35KHz, interval (15
channel) to find the power spectrum of the audio.

- The SCFC controller 54 is composed of a switched capacitor filter (SCF), and controls a block 55 of a bandpass filter, a detector, and a lowpass filter.

- The A/D converter 56 converts the audio power spectrum into an 8-bit digital value.

Register 57 stores digitized power spectrum data.

- The LOG converter 58 converts the power spectrum to the LOG axis.

- The LSFL filter unit 59 performs correction using a least squares approximation straight line (LSFL) in order to normalize the speaker's sound source characteristics.

- The binary converter 60 gives 1 to channels having a value of 1/2 or more of each local peak among the 15 channels, and gives O to channels that do not.

These LOG converters 58. To explain the LSFL filter section 59 and the binary conversion section 60 in more detail, first of all, if the frequency characteristics of the human vocal cord sound source are attenuated in the high range, individual differences are even greater. Then, this is corrected by a method using a least square error approximation straight line. A straight line giving the least square error is found for the audio frequency data obtained every 101 ISs, this is subtracted from each data, and then it is binarized. If the threshold value for converting from TSr' to BTSP is set too low, the "1" part will become broad and the pattern difference between words will be poor, whereas if it is set too high, even the slightest frequency fluctuation of the peak pattern will not be absorbed. Here, the threshold value was determined to be 50% of the peak through a simple experiment.

- The timer unit 61 outputs pulses with a constant period (ioms) and sends an interrupt signal to the CPU inside the processor 42, etc.

The processor 42 is an LSI that performs voice registration processing, recognition processing, etc., and has a parallel port 76 for communicating with one peripheral device W46. Next, the functions of each part within the processor 42 will be explained.

- The CPU section 71 is a general-purpose 16-bit CPU core,
Execute the program in the external ROM.

- The bus control unit 72 controls the internal bus 7 inside the LSI.
3 and controls the CPU bus 45 outside the LSI.

- The memory controller section 74 provides chip selection of the external ROM 3.

- The interrupt controller 1-roller unit 75 receives timer interrupt signals and the like from the processor 41 and controls interrupts.

- The parallel port section 76 communicates with the four peripheral devices 46.

・The control signal generator section 77 is based on this LSI
2. Generates various internal control signals.

- The clock generator section 78 generates a baud rate clock etc. when a serial interface is externally attached.

・The 5ECU unit (similarity element calculation unit) 79 is a BT
The unknown input cover pattern and template pattern expressed in SP are input by memory transfer, and Pd and P in equation (2), which are elements for calculating the similarity shown in equation (1), described later.
Calculate y, Pv, Pi, etc. at high speed. 5ECU part 1
9 corresponds to both specific speaker recognition and non-specific speaker recognition.

The speech recognition program resides in the program ROM and performs the following functions. (Each function is.

May be executed in parallel. ) - When an interrupt signal comes from the processor 41, the BTSP
and load the power spectrum.

・Detect the voice section and use the BTSP during that period as an unknown input cover turn. So when matching two patterns,
Considering the width of frequency fluctuations and time fluctuations, one pattern is set wide, and the other pattern is set to approximately the same width using the thinning method, which is a method of extracting line characteristics from a line figure with a certain width. Verification is performed after extracting a point or center line near the center. At this time, it is desirable to narrow the width in the 1-time axis direction as well. By doing this, even if both time and frequency axes of one pattern fluctuate, matching can be achieved without the thin line pattern protruding from the wide pattern. Various thinning processes can be considered here (for example, IEICE study group material PRL-75
-66, pp. 49 to 56), basically, points touching the boundary of a binary figure are erased one by one while maintaining the connectivity of the figure, and then the points near the center or the center line are removed. In the above explanation, one of the patterns has been made into a thin line, but this is to ensure matching without overlapping when comparing the two patterns, so one should be made into a thick line while retaining the characteristics of the line shape. Alternatively, one may be made into a thin line and the other may be made into a thick line.

・5EC of unknown cover pattern and each template pattern
The memory is transferred to the U section, and the resulting elements for determining the degree of similarity are read out.

The degree of similarity of each template to the unknown input pattern is calculated from the above factors, and the template that gives the highest degree of similarity is taken as the recognition result.

Furthermore, since the program ROM is external, it is not only possible to replace the specific speaker recognition program with the non-specific speaker recognition program, but also to enable voice registration, recognition, and other applications.

It can be stored in the same ROM as a program such as voice dialing and executed by the CPU inside the processor 42. It is also possible to upgrade the recognition program and expand the number of recognized words. There are advantages such as.

The template RAM is external, and not only can templates be easily replaced, but also the number of recognized words can be expanded by adding a RAM area.

The degree of similarity Syi between the unknown cover pattern y and the template pattern Peng i is determined as follows.

5yi=Pv-Pi-... (1)
Pd - Pv Py - Pi where: In the above formula, * indicates multiplication, and n indicates logical product (1 when y=1.■i>0, 0 otherwise). Also, f represents the frequency,
Corresponds to each channel, and represents time, which corresponds to each frame.

FIG. 6 is a flowchart 1- for explaining another embodiment of the present invention. In this embodiment, as shown in FIG. This is a case of forming a standard pattern (step 6), and when forming a standard 2D pattern, while comparing this with the 2D pattern (self) imported as data (step 5), each dimension axis direction of the former is The pattern width is determined to be different from the latter pattern in each corresponding dimension axis direction.
→5→6→3). However, in this case, the conversion method must be such that the width of the standard pattern to be formed in each dimensional axis direction is always thinner or thicker than the corresponding width of the data pattern. In this case, there is no need to compare and judge whether the widths of both patterns are different (step 5).

Of the above, pattern creation is done by the CPU in the processor 42.
and the CPU storage area in the external template RAM (w
ork RAM) and is executed by a program in an external ROM.

FIG. 7 is a flowchart for explaining another embodiment of the present invention. This embodiment imports a two-dimensional pattern of data (stapl→2→3) as shown in FIG. Form as a dimensional pattern (step 4)
It is a matter of the case. Generally, the two-dimensional pattern of data is compared with a standard quadratic pattern formed (s
top 6) The former pattern width in each dimension axis direction is determined to be different from the latter pattern width in the corresponding dimension axis direction ((step → 2 → 3 → 5). However, each dimension axis of the data pattern If the conversion trend is such that the width of the standard pattern is always thinner or thicker than the corresponding width of the standard pattern, then it is possible to compare and judge whether the widths of the two patterns are different (step 6). ) is no longer needed.

Of the above, pattern creation is performed by the CP in the processor 42.
CPU storage area in U and external template RAM (
The program is executed using the external ROM (work RAM).

FIG. 8 shows a configuration diagram of a speech recognition system using a personal computer 80.

In a general personal computer 80, CP
Since the bus 100 of U81 comes out to the expansion slot, voice Li fi'l can be executed by inserting the voice recognition board 90 including the processor 91 into this expansion slot.

The voice recognition board 90 requires only a processor 91 and a decoder 92 for address decoding, and is connected to the personal computer bus 80. Also, the input of the microphone 101 is connected from the outside.

In the example shown in FIG. 5, the functions executed by the CPU 71 of the processor 42 are
U81 will do everything. Programs and template data can be stored on hard disks, floppy disks, etc.102,10
3, and is transferred to the RAM 82 of the personal computer before voice recognition processing is performed.

(In the voice recognition system using the 35ECU structure BTSP, one frame of unknown input cover pattern can be expressed with 2 bytes because each channel is expressed with 1 bit.) In the template pattern of 1 frame, each channel is n If it is a bit, it can be expressed as n × 2 bytes. Therefore, let the unknown input cover pattern of l frame be U (16 bits), and using the same structure as U, 1
The collection of the most significant bits of each channel of the frame template pattern can be expressed as Tn-1, the collection of the next bits can be expressed as Tn-2,1, and the collection of the most significant bits can be expressed as TO. 5ECU part is this U
, Tn-1, Tn-2, , and the similarity elements Pd and Py shown in equation (2) using the data of TO.

Calculate Pv and Pi.

FIG. 9 shows a configuration diagram of the 5ECU section 79.

- The clear controller 79a clears the frame data register 79g and word data register 79i.

- The unknown data register 79a is a 16-bit register for inputting one frame of unknown input cover pattern U.

- The template data register 79b is a 16-bit register for inputting one frame of unknown input cover patterns TO, Tl, , , Tn-1.

- The logical operation unit 79e performs a logical operation on the data in the unknown data register 99C and the template data register 79b for each bit, and the data for P Y t P i is
Send to frame data register 79g.

The data for Pd and Pv are sent to the frame data adder 7.
Send to 9f.

- The frame data register 79g is a register for storing each element of similarity for one frame.

- The frame data adder 79f adds the data temporarily stored in the template data register 79b shifted upward by 1 bit (doubled) for Pd and Pv, and the data of the logic operation unit 79e. .

- The word data register 79i is a register for storing each element of similarity for one template.

- The word data adder 79f adds each similarity element stored in the word data register 79i and each similarity element stored in the frame data register 79g.

・The output sofa 79d is the word data register 79i.
This is a buffer for sending each similarity element stored in the data bus 73 to the data bus 73.

FIG. 10 is a flowchart for calculating each element of similarity between an unknown input cover pattern and one template using the 5ECU 79. First, the word data register 79i is cleared.

・Clear frame data register 79g and at the same time
Transfer.

- Send template data from the higher level (Tn-1).

・Repeat this procedure until TO.

- Repeat this procedure until the final frame.

・Read the results.

Through the above procedure, the 5ECU section 79 performs the following first-order calculation.

In equation (2), y(f, t)1mi(f, t)
Since ny(f, t) takes a value of 1 bit, Py is equal to 4 frames.

The value of Pi, ie, is sent to the frame register 79g immediately after the transfer of 'I'n-1y Il'ro.

Also, ml(LLL mu(ft)・y(Lt) is n
Since bit values are taken, the values of Pd and Pv for one frame, that is, the values of Tn-1, .

Each element for four frames is a frame data register 79g.
When all of the elements are complete, the data for each element is added to the accumulated data up to that point.

The operation of this 5ECU unit 79 is very simple in each part, so high-speed calculation is possible, and the CPU can immediately obtain the similarity element for one template after transferring the data of the last frame of one template. I can do it.

This calculation method using 5ECU units calculates the template data from the top, so it is possible to obtain the similarity element without being influenced by the value of the number of bits n representing each channel of the template pattern of one frame. can. Therefore, it is applicable to both the BTSP method of specific speaker speech recognition and speaker-independent speech recognition.

As is clear from the above explanation, according to the embodiments shown in FIGS. 5 to 10, the calculation of the BTSP method is simple;
Because the hardware is small, it has not only recognition calculations and processing, but also peripheral control and its hardware.
Applications including external control can also be executed.

Also, I! In addition to the simple calculation in the 1TSP method, using the same feature amount and the same binary pattern,
By slightly changing the recognition process, specific and non-specific recognition can be realized using software. Therefore.

By changing the external ROM, the number of words and specific/unspecific changes can be made.

Effects As is clear from the above explanation, according to the inventions stated in claims 1 and 2 of the present application, the implementation shown in FIGS. For example, by providing an unspecified standard pattern for a term (or other unit of utterance) whose registered content is known in advance, such as a command term for a device or a general name, the user can In addition to saving the trouble of registration, candidates registered in unspecified quasi-patterns are easily selected, or
It is possible to maintain a high recognition rate by eliminating bias between non-specific recognition and specific recognition, such as candidates registered in specific standard patterns being more likely to be selected.

Further, according to the invention as claimed in claim 3, it is possible to save the trouble of registering the telephone number of a public institution or the like that everyone uses, or the standard pattern of voice, and the registration work is reduced.

Further, according to the invention of claims 4 and 5.

Since the calculations of the BTSP method are simple and require small hardware, it has not only recognition calculations and processing, but also peripheral control and its hardware, and can also execute applications including external control. Also, B
In addition to simple calculations in the TSP method, specific and non-specific recognition can be realized by software by using the same feature amount and the same binary pattern and by slightly changing the recognition process. Therefore, by changing the external ROM, the number of words and specific/unspecified changes can be made.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram for explaining one embodiment of the present invention.
Figure 2 is a flowchart 1 for explaining the operation.
, FIG. 3 is a diagram for explaining another embodiment of the present invention,
FIG. 4 is a diagram showing the configuration of a standard pattern storage unit for synthesis, a standard pattern storage unit for recognition, and a telephone number storage unit, and FIGS. 5 to 10 explain other embodiments of the present invention. This is a diagram for l...Microphone, 2...Feature extraction unit, 3...
・Unspecified speaker recognition unit, 4...Specific speaker recognition unit, 5...
・Unspecified speaker standard pattern storage unit, 6...Specific speaker a
71#! Pattern storage section, 7... Unspecified/specific speaker determination section, 8... Keyboard, 9... Coefficient section, 10...
・Speech synthesis section. 11... Speech synthesis dictionary section, 12... Speaker, 21
... Transmitter, 22... Receiver, 23... Call circuit, 24... Hook switch, 25... Feature extraction unit,
26...Keyboard, 27...Control unit, 28...
Pattern matching section, 29... Standard pattern storage section for recognition, 30... Speech synthesis section, 31... Standard pattern storage section for synthesis, 32... Telephone number storage section, 33... Call control section , 34... Incoming call control unit, 35... Line control unit, 36... Office line, 41.42... Processor, 43
...ROM, 44...RAM, 45...bus, 4
6... Peripheral device, 47... Microphone. Figure 1 Figure 2 Σ Figure 3 Figure 4 Figure 5 Mei 6 Σ Figure 7 Figure 8? 0 shellfish 9 Σ bus

Claims (1)

[Claims] 1. Feature extracting means for extracting feature quantities of input speech; unspecified standard pattern storage means for storing preregistered unspecified speaker standard patterns; specific standard pattern storage means for storing a standard pattern for performing
A recognition means that performs recognition by comparing each of the unspecified standard pattern and the specific standard pattern with the feature amount of the voice extracted by the feature extraction means, and using the recognition result obtained by the comparison and recognition by the recognition means. a means for confirming the recognition results to inform the person, and at least one predetermined method.
coefficient storage means for storing predetermined coefficients, and a value corresponding to the degree of similarity to either the unspecified standard pattern or the specific standard pattern obtained by the recognition means and the coefficients. A standard for which a value corresponding to the maximum similarity is obtained by comparing the calculated value with a value corresponding to the degree of similarity to a standard pattern to which the calculation has not been performed among the unspecified standard pattern or the specified standard pattern. A speech recognition device comprising recognition means for recognizing a pattern. 2. In the speech recognition method, the feature amount of the input voice is extracted by the feature amount extraction means, which includes the following steps; and the voice feature extracted by the feature extracting means and recognized by the recognition means; the degree of similarity of either the unspecified standard pattern or the specific standard pattern obtained by the recognition means to the standard pattern; A calculation is performed on a value corresponding to the value and at least one predetermined coefficient determined in advance; Compare with the value corresponding to the degree of similarity;
A standard pattern from which a value corresponding to the maximum similarity value is obtained through the comparison is recognized. 3. Feature extracting means for extracting feature quantities of voice input from a handset for transmitting and receiving telephone calls; standard pattern storage means for storing feature quantities extracted by the feature extracting means; and a dial number corresponding to the input voice. and a number storage means for storing the number, and compares the feature amount extracted by the feature amount extraction means at the time of voice input with the feature amount stored in the standard pattern storage means stored in advance, and determines which standard pattern the input voice corresponds to. dial signal output means for reading a dial number corresponding to the recognition result recognized by the pattern matching means from the number storage means and outputting a dial signal;
A voice dialing device comprising a standard pattern storage means for unspecified speakers, which stores a standard pattern of a voice of a name corresponding to a particular telephone number as a standard pattern for unspecified speakers. 4. A feature pattern generation unit that extracts feature amounts of input data and generates a feature pattern; a matching unit that matches the feature pattern with a dictionary pattern; and a control unit that controls the generation unit and the matching unit. A control recognition processing section having a central processing section (CPU) capable of controlling the external peripheral devices required in concert, a program storage section that supports the control of the central processing section, and a storage section necessary for collation and control. and a data storage section for storing data, and each section of the generation section, control recognition processing section, program storage section, and data storage section is removably connected to a transfer path (CPUBUS) of the central processing section. A recognition processing device characterized by having the following configuration. 5. A feature pattern input section for inputting a feature pattern, a dictionary pattern input section for inputting a dictionary pattern, and at least one element necessary for pattern matching using the output data of the feature pattern input section and the dictionary pattern input section. and one of the elements necessary for the pattern matching.
a frame data register that stores values for a frame; a frame data addition unit that adds double the data stored in the dictionary pattern input unit to the data of the calculation unit; and values of elements necessary for pattern matching. 1. A recognition processing device comprising a pattern matching circuit comprising a word data register that stores a value of the element, and a word data adder that adds one frame worth of the element stored in the word data register.