JPH05323992A - Inspection learning system for speech recognition device - Google Patents

Abstract

PURPOSE:To flexibly and efficiently perform inspection and learning by a small scale sound source by automatizing the inspection and learning of a speech recognition device, and generating an inspection signal by combining a main inspection tone with a background tone arbitrarily. CONSTITUTION:This system comprises a tone attribute table on which the fundamental attribute of the main inspection tone and the background tone are registered, a menu table on which at least inspection sequence, the main inspection tone of the tone attribute table used in each inspection sequence, the background tone and its superimposing level of the tone attribute table used in each inspection sequence, a speech output part 17 which supplies the inspection signal, and a connection response detecting part 4 which inputs a response signal from a speech recognition device 1, etc. The tone attribute table is referred according the inspection sequence of the menu table, and the inspection signal for the speech recognition device 1 can be generated, and also, the recognition ratio of the speech recognition device 1 is calculated based on the response signal obtained via the connection response detecting part 4.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a system for inspecting and learning the recognition rate of an apparatus having a voice recognition mechanism.

[0002]

2. Description of the Related Art In recent years, speech recognition has been actively researched and developed in automatic ticket vending machines, machine tools, word processors, etc. at transportation stations. That is, it is intended to perform ticket sales processing, processing processing, or character input processing by voice, which is the simplest and most basic means of communication for humans.

By the way, the voice recognition rate is an important factor for determining whether or not this type of system can be put to practical use.
As a system for inspecting these voice recognition devices, there are Japanese Patent Laid-Open Nos. 59-162598 and 63-63.
No. 149699 is proposed.

[0004]

Each of the above-mentioned prior arts is for inspecting the speech recognition rate or for learning the speech recognition apparatus. However, the speech recognition apparatus has a recognition rate depending on the environment in which it is installed. It is naturally expected that there will be large fluctuations.

Therefore, for the main inspection sound to be provided to such a voice recognition device, is it necessary to prepare in advance a voice recorded in the crowd of a station or a sound source in which the crowd and the main inspection sound are mixed. Otherwise, it was necessary to reproduce the main inspection sound and the background sound by different reproduction devices.

However, there are endless combinations of levels (volumes) of the main inspection sound and the background sound, and it is almost impossible to prepare all the inspection signals with a recording tape or the like. In the case of a system configuration having a reproducing device independent of sound, the reproducing mechanism becomes large and it is not practical.

[0007] Further, little has been considered on the teaching technique for improving the recognition rate for the voice recognition device. The present invention has been made in view of the above problems, and an object thereof is to automate inspection / learning of a voice recognition device, generate an inspection signal by arbitrarily combining a main inspection sound and a background sound, and to perform a small scale. An object of the present invention is to provide an inspection and training system for a voice recognition device that enables flexible and efficient inspection and training with various inspection sound sources.

[0008]

The present invention has the following configuration in a system for providing a main inspection sound to a voice recognition device and evaluating the recognition result.

That is, at least the main inspection sound, the sound attribute table in which the basic attributes of the background sound are registered, at least the inspection order, the main inspection sound of the sound attribute table used in each inspection order, and similarly used in each inspection order. A menu table in which the background sound of the sound attribute table and the superposition level thereof are registered, a voice output unit for providing a voice inspection signal to the voice recognition device, and a control response detection for inputting a response signal from the voice recognition device. Section, referring to the sound attribute table according to the inspection order of the menu table, and generating an inspection signal for the voice recognition device,
And a main control unit that calculates a recognition rate of the voice recognition device based on a response signal obtained from the voice recognition device through the control response detection unit.

[0010]

According to the present invention, by setting the menu table, the type of the main inspection sound, the level (volume) of the main inspection sound, the presence or absence of the background sound for this main inspection sound, the level of the background sound, the main inspection sound and the background. It is possible to arbitrarily select the superimposition ratio with sound.

As a result, for example, a speech recognition apparatus, which is an apparatus to be inspected, is supplied with an inspection signal of a gradually increasing level from a low level to automatically inspect the recognition rate and adjust the sensitivity of the speech recognition apparatus. It becomes possible to teach.

Further, according to the present invention, by storing the contents of the sound attribute table as a sound source file in the storage unit, random access is possible, the speed of sound source reproduction is improved, and the inspection time is shortened. it can.

Further, the target recognition rate is set in the menu table, and the inspection result table is provided to register the recognition result, for example, erroneously recognized sound, so that this can be included in the inspection content again and the lesson for the speech recognition apparatus can be automated. can do.

As a result, the efficiency of automatic learning is improved, and the recognition rate in the voice recognition system can be dramatically improved in a short time. Conventionally, it has been difficult to compare general recognition rates for sound sources used for speech recognition measurement because they are based on individual regulations and not standard. According to the present invention, it is possible to standardize the inspection sound by the type of the sound source, its volume, the type and the volume of the background sound, the inspection sound by its volume, and the conditions of the volume, and different manufacturers Even with the voice recognition device according to the above, it is possible to evaluate with the same standard.

[0015]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a block diagram showing the outline of the system configuration of this embodiment. 2 is an external view thereof, and FIG. 3 is a block diagram showing internal functions of the inspection unit.

As shown in FIG. 2, the system configuration of this embodiment is composed of a voice recognition device 1 which is a device to be inspected and an inspection unit 2. The recognition result output unit (training control input unit) 15 of the voice recognition device 1 is connected to the control response detection unit 4 of the inspection unit 2. Further, a speaker 6 from the inspection unit 2 is arranged opposite to the voice input unit 5 of the voice recognition device 1. That is, the inspection unit 2
From the speaker 6 through the speaker 6, the reaction of the voice recognition device 1 at this time is taken out through the recognition result output unit (training control input unit) 15 and analyzed by the inspection unit 2 to inspect the recognition rate. ..

In FIG. 2, the inspection unit 2 includes a CRT.
Alternatively, a monitor such as the liquid crystal panel 7 and a printer 8 are provided so that the inspection result can be output. Further, the inspection result and the like can be notified to the external computer 10. Further, in the figure, the console 11
The operator has an inspection unit 2 by an instruction switch such as a push button provided on the operation console 11.
Can be controlled.

FIG. 2 shows a configuration having the inspection unit 2 dedicated to the present system. However, as shown in FIG. 7, the inspection unit 2 is replaced with a general-purpose computer 12, and the speaker 6 is replaced with a cassette stereo 13 with a radio. May be realized with.

The voice recognition device 1 is usually provided in a ticket vending machine of transportation means, and has a voice recognition unit 14 and a recognition result output unit (training control input unit) 15. The recognition result output unit (training control input unit) 15 has either a communication interface (RS-232C), ON / OFF digital input / output (Di / o), or a voltage signal output. Although FIG. 1 illustrates the case where all of these three systems are provided, only the communication interface among them may be provided.

The inspection unit 2 includes the voice recognition device 1 described above.
Has a control response detection unit 4 for receiving a detection signal from the recognition result output unit (training control input unit) 15. The control response detection unit 4 has the above-mentioned communication interface (RS-
232C), ON / OFF digital input / output (Di / o)
Further, it has a communication processing section, a data input / output section (Di / o) and an analog / digital conversion section (A / D) corresponding to the voltage signal output. Of course, the voice recognition device 1, the inspection unit 2, and the external computer 10 are connected to a LAN (Loca
l Area Network) or the like.

The control response detector 4 is connected to the central control unit (CPU) by the system bus 16. Central control unit (CPU) that functions as the main control unit
Is configured to analyze the detection signal obtained through the control response detector 4.

The system bus 16 includes a voice output section 1
Have 7. The speaker 6 shown in FIG. 2 is connected to the voice output unit 17 and outputs a test signal to the voice recognition unit 14 of the voice recognition device 1.

As shown in FIG. 3, the audio output unit 17 includes an adder 21 for adding the outputs of the buffer 18 and the superimposed sound source buffer 20, and a digital / analog conversion unit (D / D).
A), a filter 22 that cuts off a certain low and high frequencies, an amplifier 23 that amplifies a signal, and a sound pressure adjusting unit 2
4 and 4.

The adder 21 may be replaced with a digital signal processor (DSP). In this case, the buffer 18 and the superimposed sound source buffer 20 can be individually amplified. Further, when the background sound and the inspection sound are different from the levels registered in advance in the sound attribute table 30 or the like, an amplification calculation circuit may be provided in the audio output unit 17 so that the level can be changed.

The system bus 16 is further provided with an external storage device interface (I / O) to which an external storage device 25 such as a magnetic disk device or an optical disk device is connected. The external storage device 25 stores large-capacity data such as sound source data.

Further, the system bus 16 is provided with a communication section 26, which enables communication with the external computer 10. This communication unit 26 is the above-mentioned RS-232C.
And the like, or a bus interface.

A memory 27 connected to the system bus 16
The buffer 18, the superimposed sound source buffer 20 and various tables described above are assigned to and set in a predetermined address.

Further, an input / output interface (Di / o) is connected to the system bus 16, and an external input / output device such as the above-mentioned console 11 or liquid crystal panel 7 is connected.
Alternatively, the printer 8 or the like is connected to this.

Next, the tables set in the memory 27 will be described. In the inspection menu table 28, as shown in FIG. 4, the inspection order, the inspection sound symbol, the inspection sound level, the superimposed sound symbol, the superimposed sound level, and the target recognition rate are set. Here, the reason why the target recognition rate registration area is provided is that the inspection reference may be based on whether or not the target recognition rate registered in the inspection menu has been cleared. That is, considering that the voice recognition device 1 may not always require a high recognition rate, the recognition rate can be set in advance as a reference value for inspection.

The central control unit (CPU) sequentially carries out the inspection in accordance with the inspection order of the inspection menu table 28. At this time, the inspection sound registered in the inspection sound symbol of the inspection menu table 28 as the inspection parameter. (Defined in the sound attribute table described later), the volume is inspected at the inspection sound level. Further, when there is a background sound, the inspection is performed at the volume defined by the level of the superimposed sound using the superimposed sound symbol as a parameter.

As shown in FIG. 5, the sound attribute table 30 uses a sound symbol as an ID to discriminate between a check sound and a superposed sound, a sound name, a level, and a check sound attribute (gender distinction,
Age, characteristics of voice quality, accent, etc.), type of background sound, location, etc. are registered.

Various parameters corresponding to phonetic symbols registered in the phonetic attribute table 30 are initial values,
On the other hand, the parameter in the above menu table has priority.

Explaining this figure by way of example, the inspection sound ID is "k.
“001” is an inspection sound, the sound name is “Kyoto”, the level is 50 dB, and the attribute of the inspection sound is 45.
The voice of a talented husky man, defined as having an Aomori accent.

The inspection sound ID "c001" is a background sound superimposed on the inspection sound, and the crowd of the ticket office in the station is defined as 50 dB. here,
In the inspection menu table 28 shown in FIG. 4, when the first inspection is performed, the inspection sound symbol is "k001", so the attribute parameter of "k001" is read from the sound attribute table 30. The inspection sound level is 50 dB, which is the same level as the sound attribute table 30 in this case, and therefore the level of the sound attribute table 30 is output as it is without being changed. Further, since the superimposed sound symbol is "c001", when this attribute parameter is read from the sound attribute table 30, it is 50 dB, but it is amplified to 60 dB of the superimposed sound level defined by the inspection menu table 28 and output.

As a result, 1 of the inspection menu table 28 is displayed.
The number of inspection signals is 60 dB due to the crowd in the station yard as background sound.
In the state set by, the voice of a 45-year-old man with a husky Aomori accent is said to be “Kyoto” at 50 dB.

FIG. 3 shows the more detailed functions of the inspection unit 2 centering on the central control unit (CPU). In the figure, the inspection menu table 28 and the sound attribute table 30 are shown together with the IC memory 31 and the buffer 18.
A superposed sound source buffer 20 is provided, which are set in the memory 27 shown in FIG.

In the IC memory 31, the sound source defined in the sound attribute table 30 is binarized or multi-valued and registered as a sound source file. Central controller (C
The main function of the PU is, firstly, as described above, the inspection parameters are taken out from the inspection menu table 28 and the IC
The actual sound source data (main inspection sound and background sound) stored in the memory 31 is developed in the buffer 18 and the superposed sound source buffer 20, and these are superposed using the adder 21, and the audio output unit 17 outputs the inspection signal. It is to output through the speaker 6.

Secondly, the response signal from the voice recognition unit 14 based on the output from the speaker 6 is input to the control response detection unit 4 via the recognition result output unit (training control input unit) 15 and the response is sent. The purpose is to analyze the signal and output the analysis result to the liquid crystal panel 7 or the printer 8.

The flow chart shown in FIG. 6 shows the processing of the central control unit (CPU) in order. First,
When the inspection is started, an inspection menu is displayed on the liquid crystal panel 7 (step 601).

Next, the instruction from the push button from the console 11 is detected, and the inspection menu table 28 is loaded (602). At this time, a plurality of inspection menu tables 28 may be registered so that they can be arbitrarily selected by operating the console 11.

Next, the inspection parameters defined in the inspection menu table 28 are sequentially loaded into the buffer 18 and the superimposed sound source buffer 20 according to the sound attribute table 30 (603, 604).

Next, the main inspection sound stored in the buffer 18 and the background sound stored in the superimposed sound source buffer 20 are added (superposed) by the adder 21 and reproduced by the sound output unit 17 through the speaker 6. ..

In the voice recognition unit 14 of the voice recognition device 1, the inspection unit 2 outputs the response signal to the reproduced sound from the speaker 6 through the recognition result output unit (training control input unit) 15.
Is sent to the control response detection unit 4.

The central control unit (CPU) analyzes the response signal obtained through the control response detection unit 4, and if the response signal in the voice recognition device 1 is different from the expected recognition response, the response content , The inspection sound name, the inspection sound pressure level, the superimposed sound name, the superimposed sound pressure level, etc. are stored and output to the liquid crystal panel 7 or the printer 8 or the like through the input / output interface (Di / o) (604).

After repeating the processing of step 604 in the order of the inspection menu table, the recognition rate, the target recognition rate and the quality of recognition are recorded in the memory 27 and output to the liquid crystal panel 7 (605).

In the above description of the embodiment, FIGS.
In addition to the method of registering the level of the inspection sound in advance, as a method of changing the level of the inspection sound, the inspection menu table (training menu table) 28 includes a plurality of files having different levels. The audio output unit 17 may be prepared in advance or a programmable gain amplifier or the like may be provided to perform automatic gain adjustment. Then, the level is gradually increased, and the level and the like when there is a recognition reaction in the voice recognition device 1 are recorded.

In addition to the level change, a tone change or the like may be added. Further, as the level change, the gain may be gradually increased from a small level, or the gain may be decreased from a high level.
Further, the level of the main inspection sound may be fixed and the level of the background sound that is the superimposed sound may be changed.

In the above, the inspection process for the voice recognition device 1 was mainly focused, but the process mainly for the lesson for the voice recognition device 1 will be described below. As the hardware configuration, the same system configuration as that shown in FIGS. 1 to 3 can be used.

FIG. 8 shows an inspection result table 32. The inspection result table 32 is the above-mentioned inspection menu table (here, it is used as a training menu table).
It is set in the memory as well as.

As a parameter of the inspection result table 32, an item of "erroneous recognition sound" is added to the parameter of the inspection menu table (training menu table) 28.

That is, as a result of analyzing the response signal from the voice recognition device 1 to the inspection signal, if this is an erroneously recognized sound, an inspection sound corresponding to this erroneously recognized sound is searched from the sound attribute table 30 and is searched. Register in the inspection new table (training menu table) 28.

That is, in the concrete example of FIG. 8, the note name represented by the inspection sound symbol "k001" is originally "Kyouto" as compared with FIG. However, when the voice recognition device 1 erroneously recognizes that “hiding” is uttered, the sound attribute table 30 is searched for the inspection sound corresponding to “hiding”. Here, according to the sound attribute table 30 of FIG. 5, "hidari" which is a misrecognized sound.
It is understood that is the inspection sound represented by the inspection symbol "k002".

Therefore, the central control unit (CPU) registers the inspection sound symbol “k001” which should have been originally recognized and the inspection sound number “k002” which is erroneously recognized in the inspection menu table (training menu table) 28. by this,
The correct inspection sound and the inspection sound that was erroneously recognized the previous time are output again to the voice recognition device 1 to teach.

The flow chart shown in FIG. 9 shows the creation of the inspection menu table (training menu table) 28. In the figure, the output condition (including the background sound) of the inspection sound (x: the original sound name to be recognized) is first registered in each item of the inspection result table 32 from the inspection result based on the response signal, and then the erroneous recognition is performed. Sounds (y: erroneously recognized sound names) are registered in the "misrecognized sounds" item of the inspection result table 32 to complete the inspection result table 32 (step 901).

Next, an inspection sound of the same condition as the inspection sound x is retrieved from the sound attribute table 30 and the inspection sound symbol is registered in the inspection menu table (training menu table) 28 (902).

Subsequently, the inspection sound of the same condition as the inspection sound y is searched from the sound attribute table 30, and the inspection sound symbol is searched for in the inspection menu table (training menu table) 28.
Register with (903). As a result, the inspection sound x and the inspection sound y are output again to the voice recognition device 1, and the training of the voice recognition device 1 can be realized.

FIG. 10 shows a training processing procedure based on the examination menu table (training menu table) 28 thus created. First, when the inspection is started, an inspection menu (training menu) is displayed on the liquid crystal panel 7 (step 1001).

Next, an instruction by the push button from the console 11 is detected, and the inspection menu table (training menu table) 28 is loaded (1002). Next, the inspection parameters defined in the inspection menu table (training menu table) 28 are converted into the sound attribute table 30 described above.
The buffer 18 and the superimposed sound source buffer 20 are sequentially loaded in accordance with the above. At this time, the voice recognition device 1 is set to the learning mode by the manual operation or the control from the inspection unit 2 (1004-1).

Next, the inspection sound symbol (recognition symbol) of the inspection signal (voiced sound) is notified to the voice recognition device 1 through the response detection unit 4 (1004-2). Next, after notifying the voice recognition device 1 of the start of utterance, that is, the start of output of the inspection signal, the sound source is reproduced from the speaker 6 (1004-3).

By the steps 1004-1 to 1004-3, the learning sounds such as "Kyoto" and "Hidari" are taught to the voice recognition apparatus 1 together with the corresponding test sound symbols "k001" and "k002". It

Next, the voice recognition apparatus 1 is set to the inspection mode, and the inspection is repeated 100 times using the same sound source as the sound source taught in the above (1004-4). Next, the recognition rate at 100 times is calculated, and if the set target recognition rate is not reached, the steps from 1004-2 onward are repeated.

After the learning → inspection is performed in this way, the recognition rate, the target recognition rate, and the quality of the result are recorded in the memory 27 and output to the liquid crystal panel 7 (100
5). According to the present embodiment, the sound source can be randomly accessed according to the inspection menu table (training menu table) 28 and the sound attribute table 30. Therefore, the speed of sound source reproduction is faster than that when the inspection sound is recorded on a recording tape or the like. It is possible to shorten the inspection time when a plurality of inspections are performed.

Further, since it is easy to repeat the lessons / inspections until the recognition rate specified for each voice recognition device 1 is satisfied, it is possible to automatically teach the voice recognition device 1.
Further, since the main inspection sound and the background sound can be reproduced at the same time, different reproducing apparatuses are not required. Moreover, since the main inspection sound and the background sound can be arbitrarily selected and the ratio of their levels can be arbitrarily set, various inspections can be performed, and the accuracy of calculation of the recognition rate using statistical processing is improved.

Further, by providing an inspection result table 32,
Efficient automatic learning is performed by registering an erroneously recognized sound in the case of erroneous recognition, retrieving an inspection sound corresponding to this erroneously recognized sound from the sound attribute table 30 and adding it to the inspection menu table (training menu table) 28. Is possible.

[0065]

According to the present invention, an inspection signal can be generated by arbitrarily combining a main inspection sound and a background sound, and flexible inspection / training can be performed with a small inspection sound source.

In addition, the efficiency of automatic learning can be improved,
The recognition rate in the voice recognition system can be dramatically improved in a short time.

[Brief description of drawings]

FIG. 1 is a block diagram showing an outline of a system configuration of an embodiment of the present invention.

FIG. 2 is an explanatory diagram showing the external appearance of the system configuration of the embodiment of the present invention.

FIG. 3 is a block diagram showing an internal configuration of an inspection unit according to an embodiment of the present invention.

FIG. 4 is an explanatory diagram showing the contents of an inspection menu table (training menu table) in the embodiment of the present invention.

FIG. 5 is an explanatory diagram showing the contents of a sound attribute table in the embodiment of the present invention.

FIG. 6 is a flowchart showing an inspection process in the embodiment of the present invention.

FIG. 7 is a diagram showing a modification of the system configuration in the embodiment of the present invention.

FIG. 8 shows the contents of the inspection result table and the inspection menu table (training menu table) in the embodiment of the present invention.
Explanatory diagram showing the relationship with

FIG. 9 is a flowchart showing a process for creating a lesson menu according to the embodiment of the present invention.

FIG. 10 is a flowchart showing the inspection training process in the embodiment of the present invention.

[Explanation of symbols]

1 ... Voice recognition device 2 ... Inspection unit 4 ... Control response detection unit 5 ... Voice input unit 6 ... Speaker 7 ... Liquid crystal panel 8 ... Printer 10 ... External computer 11 ... Operating console 12 ... General-purpose computer 13 ・ Cassette stereo with radio 14 ・ ・ Voice recognition unit 15 ・ ・ Recognition result output unit (training control input unit) 16 ・ System bus 17 ・ Voice output unit 18 ・ ・ Buffer 20 ・ ・ Superposed sound source buffer 21 .. Adder 22 .. Filter 23 .. Amplifier 24 .. Sound pressure adjusting unit 25 .. External storage device 26 .. Communication unit 27 .. Memory 28 .. Inspection menu table (training menu table) 30 .. Sound attribute Table 31 ... IC memory 32 ... Inspection result table CPU ... Central control unit

Claims (4)

[Claims] 1. A system for providing a voice recognition device with a main inspection sound and, if necessary, a background sound, and evaluating the recognition result, wherein at least a main inspection sound and a sound attribute in which basic attributes of the background sound are registered are registered. A table, a menu table in which at least an inspection order, a main inspection sound of the sound attribute table used in each inspection order, a background sound of the sound attribute table also used in each inspection order, and a superimposition level thereof are registered; A voice output unit that provides a voice inspection signal to the device, a control response detection unit that inputs a response signal from the voice recognition device, and a reference to the sound attribute table according to the inspection order of the menu table. The inspection signal for the recognition device is generated, and based on the response signal obtained from the voice recognition device through the control response detection unit, Inspection training system for a speech recognition apparatus and a main control unit for calculating a recognition rate of the speech recognition device. 2. The main control unit includes a memory for registering a sound source file, refers to the sound attribute table based on the menu table, and stores the sound source from the sound source file in a buffer. An inspection training system for a voice recognition device according to claim 1. 3. The voice recognition device according to claim 1, wherein the menu table is provided with an area for registering a target recognition rate for each main inspection sound, and the inspection reference is based on whether or not the target recognition rate is exceeded. Inspection training system. 4. A training menu table for registering a target recognition rate and an inspection result table for registering recognition results are provided. The inspection result table searches for false recognition sounds of the recognition results from the sound attribute table. The main control unit registers an inspection sound including an erroneously recognized sound from the inspection result table in the lesson menu table and re-executes the lesson content to perform automatic lesson to the voice recognition device. Item 1. An inspection training system for a voice recognition device according to Item 1.