JP2005173157A - Parameter setting device, parameter setting method, program and storage medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a parameter setting device which can determine a parameter set so that other conditions fall within ranges to which a user can give permission even in the case where the parameter set is determined so that at least one condition out of a plurality of operating conditions becomes optimum. <P>SOLUTION: In the parameter setting device, a voice recognizing part 204 performs voice recognition using a parameter set selected by a parameter selecting part 202. Results of operations obtained as results of having performed voice recognition in respective trials, namely, recognition rates, recognition periods, and memory usages are held. A vector quantization part 206 performs vector quantization by considering the results of operations as vectors and results of the vector quantization are displayed by a display processing part 208. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a parameter setting device, a method, a program, and a storage medium for appropriately setting parameters for speech recognition.

  In order to use speech recognition in a real environment, various parameters need to be tuned. Here, the parameters include arbitrary parameters related to speech recognition, such as a beam width, an acoustic model, and a VAD (speech extraction) threshold. The parameter tuning is generally performed by a method in which an engineer familiar with speech recognition determines the parameter through trial and error. Therefore, it can be said that parameter tuning is an operation that can be performed only by engineers who have deep knowledge and experience regarding speech recognition.

On the other hand, a method for automatically tuning these parameters has been proposed (see, for example, Patent Document 1). This method gives an evaluation set (speech data and recognition grammar), operating conditions (requirements such as recognition rate, recognition time, memory usage, etc.), search conditions (step size and range) for each parameter, and values for each parameter The speech recognition is repeatedly tried while changing the search conditions in various ways, and the operation results (recognition rate, recognition time, memory usage) in each trial are recorded, and the optimum operation conditions are satisfied from the recorded operation results. The final set of parameters.
JP 2002-328696A

  However, in the above-described method, it is impossible to express the operating condition by AND of each necessary condition such as the recognition rate, the recognition time, and the memory usage. The operating conditions that the user really wants to set are determined in consideration of the recognition rate, the recognition time, and the memory usage. That is, even if the operation condition gives priority to the recognition rate, such an operation condition cannot be actually used if the recognition time and the amount of memory used are too large. If the recognition rate is somewhat inferior but the memory usage can be halved, you should be willing to adopt it. For example, when a recognition rate of 90% or more and a memory usage of 2 MB or less are set as operating conditions, if a parameter set with a recognition rate of 90% and a memory usage of 2 MB is selected as a result of the trial, it was not selected. Some parameter sets have a recognition rate of 89.9% but may have a memory usage of 1.2 MB. Of course, this parameter set cannot be obtained as a solution because it does not satisfy the requirements regarding the recognition rate. However, in general, the effect of reducing the amount of memory used obtained on the contrary to 0.8 MB is more attractive than the recognition rate of 0.1%. In such a case, the latter parameter set is desired to be selected, but in the above method, the former parameter set is selected.

  Moreover, if only two conditions of the recognition rate and the memory usage are set as described above, the recognition time may be very long in the optimum parameter set obtained.

  The present invention has been made in view of the above problems, and even when at least one of a plurality of operating conditions is determined to be optimal, other conditions are within a range acceptable by the user. It is an object of the present invention to provide a parameter setting device capable of determining a parameter set so as to fall within the range, a method thereof, a program realizing the control method, and a storage medium.

  In order to achieve the above object, the present invention provides a parameter setting device for speech recognition, and trial means for trying speech recognition using each of a plurality of parameter sets prepared in advance for operating conditions. And an operation result holding means for holding an operation result including a recognition rate for each trial together with a parameter set corresponding thereto, and a grouping means for grouping the operation results for each trial.

  In order to achieve the above object, the present invention is a parameter setting method for speech recognition, and a trial step of trying speech recognition using each of a plurality of parameter sets prepared in advance for operating conditions. And an operation result holding step for holding an operation result including a recognition rate for each trial together with a parameter set corresponding thereto, and a grouping step for grouping the operation results for each trial.

  In order to achieve the above object, the present invention is a program for realizing a parameter setting method for speech recognition, and uses a plurality of parameter sets prepared in advance for operating conditions to perform speech. A trial module for attempting recognition; an operation result holding module for holding an operation result including a recognition rate for each trial together with a parameter set corresponding to the trial module; and a grouping module for grouping the operation results for each trial. It is characterized by that.

  In order to achieve the above object, the present invention provides a storage medium in which the above program is stored so as to be readable by a computer.

  According to the present invention, even when at least one of a plurality of operating conditions is determined to be optimal, the parameter set is determined so that the other conditions are within an allowable range for the user. be able to.

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

(First embodiment)
FIG. 1 is a block diagram showing a hardware configuration of a parameter setting apparatus for speech recognition according to the first embodiment of the present invention. In the present embodiment, a case where a parameter setting device is realized using a general personal computer will be described. However, the present invention may be a dedicated parameter setting device or another type of device. Needless to say, it is good.

  The parameter setting device is a device for appropriately setting parameters for speech recognition. As shown in FIG. 1, the parameter setting device is a control memory 101 composed of a ROM, a central processing unit (CPU) 102, a memory composed of a RAM, and the like. 103, an external storage device 104, an input device 105, a display device 106, a bus 107, and the like. The control memory 101 stores a control program for realizing the parameter setting device and data used in the control program. These control programs and data are appropriately fetched into the memory 103 through the bus 107 under the control of the central processing unit 102 and executed by the central processing unit 102.

  Next, the module configuration of the parameter setting apparatus will be described with reference to FIG. FIG. 2 is a block diagram showing a module configuration of the parameter setting apparatus of FIG.

  As shown in FIG. 2, the module configuration of the parameter setting apparatus includes a parameter search condition holding unit 201, a parameter selection unit 202, an evaluation data holding unit 203, a speech recognition unit 204, an operation result holding unit 205, and a vector quantization unit. 206, a quantization result holding unit 207, a display processing unit 208, and an input processing unit 209. This module configuration is configured by the central processing unit 102 executing a control program stored in the control memory 101.

  The parameter search condition holding unit 201 holds search conditions such as a maximum value, a minimum value, and a step size when various parameters are changed variously. The parameter selection unit 202 selects one parameter set from the parameter sets that have not been tried so far from the search conditions held by the parameter search condition holding unit 201. The evaluation data holding unit 203 holds data for performing speech recognition evaluation. Specifically, this data is speech data for evaluation, recognition grammar, language model, acoustic model, and the like. The voice recognition unit 204 performs voice recognition using the parameter set selected by the parameter selection unit 202. The operation result holding unit 205 holds an operation result obtained as a result of executing speech recognition in each trial. Here, the operation result indicates a recognition rate, a recognition time, and a memory usage.

  The vector quantization unit 206 performs vector quantization by regarding the operation result as a vector. The quantization result holding unit 207 holds the result of vector quantization by the vector quantization unit 206. The display processing unit 208 displays the result of vector quantization by the vector quantization unit 206. The input processing unit 209 has an operation unit for the user to perform an operation for selecting and inputting a desired one from the quantization results, and the quantization result selected and input by an operation using the operation unit Is accepted. The parameter set determination unit 210 obtains a parameter set corresponding to the user input.

  Next, processing of the parameter setting device will be described with reference to FIG. FIG. 3 is a flowchart showing a processing procedure of the parameter setting apparatus of FIG. The procedure shown in the flowchart of FIG. 3 is executed by the module configuration of FIG.

  In this parameter setting apparatus, as shown in FIG. 3, first, in step S301, the parameter selection unit 202 determines whether there is an untrial parameter set in the search conditions held in the parameter search condition holding unit 201. Determine whether or not. Here, the search condition is defined by a minimum value, a maximum value, a step size, and the like that each parameter can take in order to try speech recognition with parameters of various values. Needless to say, the search condition may be a list of parameter values to be tried. If an untrial parameter set exists, the parameter selection unit 202 selects one of the untrial parameter sets from the search conditions held by the parameter search condition holding unit 201, and the process proceeds to step S302. It is done. On the other hand, if there is no untrial parameter set, the process proceeds to step S304.

  In step S <b> 302, the voice recognition unit 204 performs voice recognition using the parameter set selected by the parameter selection unit 202 and the evaluation data held by the evaluation data holding unit 203. The voice recognition unit 204 calculates a recognition rate, a recognition time, and a memory usage amount. In step S303, the operation result holding unit 205 holds the recognition rate, the recognition time, the memory usage, and the parameter set used for the trial obtained as a result of the speech recognition as a set. Next, the process returns to step S301.

  In step S304, the vector quantization unit 206 regards the operation results (recognition rate, recognition time, memory usage) obtained in each trial so far as a three-dimensional vector, and performs vector quantization. Is held in the quantization result holding unit 207. For vector quantization, a general method may be used. The number of quantization vectors obtained as a result of quantization is preferably as small as possible so that the user can select from among them. This number may be set to an appropriate value in advance or may be set by the user.

  Next, in step S305, the display processing unit 208 presents the quantization result held in the quantization result holding unit 207 to the user, and in the subsequent step S306, the input processing unit 209 accepts the user input. The information input here is one desired by the user from among the quantization results. When the input is accepted, in step S307, the parameter set determination unit 210 selects a parameter set that matches the quantization result held in the quantization result holding unit 207 based on the user selection result received by the input processing unit 209. Find and use it as a set of parameters. Then, this process ends.

  Next, input / output data of vector quantization by the vector quantization unit 206 in the present embodiment will be described with reference to FIG. FIG. 4 is a diagram showing an example of input / output data of vector quantization by the vector quantization unit 206 of FIG.

  As an input of vector quantization input to the vector quantization unit 206, that is, an operation result of all trials, for example, there is an input 401 as shown in FIG. A parameter set corresponding to this input 401 is a parameter set 403. The input 401 and parameter set 403 are held in the operation result holding unit 205 in FIG. In response to the input 401, the vector quantization unit 206 obtains an output 402 of vector quantization. This output 402 is held in the quantization result holding unit 207.

  Thus, according to the present embodiment, since vector quantization is performed for all trials, it is possible to determine a parameter set that is close to the operating condition desired by the user. In other words, even when at least one of the operating conditions of the recognition rate, the recognition time, and the memory usage is determined to be optimal, the other conditions are within the user-acceptable range. A parameter set can be determined.

  In the present embodiment, a case has been described in which the result of vector quantization is presented to the user, and the user is allowed to select a desired result. However, the present invention is not limited to this, and as a result of vector quantization Among the obtained representative vectors, for each element (recognition rate, recognition time, memory usage), the vector containing the element with the maximum value is assigned the recognition rate priority, recognition time priority, and memory saving priority. It may be assigned to each mode, and the parameter set at that time may be provided to the user as a parameter of each mode. As a result, even a user with little experience in parameter recognition for speech recognition can select a parameter set.

  In the present embodiment, the case where vector quantization is performed for all trials has been described. However, the present invention is not limited to this, and after clearly unsuitable speech recognition parameters are removed. Alternatively, vector quantization may be performed. As a method for discriminating that the parameters are obviously inappropriate as a speech recognition parameter, paying attention to the operation result, for example, regarding the recognition rate, the recognition rate is 80% or less, and the maximum value of the recognition rate is 80% or less. And those that are 20% or more lower than the maximum recognition rate may be considered inappropriate. Needless to say, it is effective to perform the same pre-processing on the operation result related to the recognition time and the memory usage, not limited to the recognition rate. Thereby, in the case of assigning the vector of each element (recognition rate, recognition time, memory usage) to each mode of recognition rate priority, recognition time priority, and memory saving priority, the parameter set of the recognition rate priority mode is set. Is selected, it is guaranteed that the recognition rate and the memory usage are within the allowable range.

(Second Embodiment)
Next, a second embodiment of the present invention will be described with reference to FIG. FIG. 5 is a diagram showing the result of grouping on a two-dimensional plane regarding the recognition time and the memory usage in the parameter setting apparatus according to the second embodiment of the present invention.

  In the first embodiment, it is desirable that the number of quantization vectors obtained as a result of quantization be as small as possible so that the user can select from among them. In contrast, the present embodiment employs a method of grouping and presenting each operation result value.

  In the case of the method according to the present embodiment, the operation results obtained by all trials are classified by combinations of memory usage and recognition time. For example, a method of quantizing in units of 0.2 MB in the case of memory usage and in units of 50 milliseconds in the case of recognition time is used. From this point of view, all operation results are classified. Next, in each group, an operation result that maximizes the recognition rate is selected. The operation result selected in each group corresponds to the vector quantization result in the first embodiment. This is then presented to the user.

  The amount of memory used and the increment of recognition time should be large enough not to cause problems in system design. That is, it can be considered that there is no great difference between those belonging to the same group. Among them, since the operation result that maximizes the recognition rate can be selected, the memory usage amount and the recognition time can be reduced without sacrificing the recognition rate, as is clear when compared with the first embodiment. There is an advantage that the step can be cut into a reasonable size.

  In the present embodiment, grouping results are presented on a two-dimensional plane regarding recognition time and memory usage, for example, as shown in FIG. Here, the number on the plane is the maximum recognition rate in each group. The user can select a parameter set that achieves a desired recognition rate, memory usage, and recognition time from FIG.

  For example, when comparing the group 501 having the maximum recognition rate 93 and the group 502 having the maximum recognition rate 94 in FIG. 5, the group 501 has a higher recognition rate than the group 502, but the difference is only 1%. Moreover, regarding recognition time, each group 501 and 502 becomes the same group of 850 milliseconds-900 milliseconds. On the other hand, focusing on the memory usage, the group 501 needs a memory amount of 2.2 MB to 2.4 MB, but the group 502 has a memory of 1.6 MB to 1.8 MB, which is smaller than the group 501. The amount is. Therefore, the user can select a group 502 that has a slightly low recognition rate but has a very small memory usage.

  In this embodiment, the case where the recognition rate is presented on the memory usage-recognition time plane has been described. However, the present invention is not limited to this, and the memory usage-recognition rate plane and the recognition time-recognition rate plane are You may make it show. Needless to say, the minimum value of the recognition time is displayed in the case of the memory usage amount-recognition rate plane, and the minimum value of the memory usage amount is displayed in the case of the recognition time-recognition rate plane.

  In the first embodiment, the case where the operation result (recognition rate, recognition time, memory usage) is regarded as a vector and vector quantization is described, but the present invention is not limited to this. It is also possible to perform vector quantization by regarding the memory usage as a vector. Further, in combination with the second embodiment, the quantization result may be displayed on the memory usage amount-recognition time plane. The number presented on the plane is the maximum recognition rate in the set of motion results represented by the representative vector. In this case, it is necessary to record which representative vector the operation result belongs to.

  Further, instead of displaying the recognition rate on the memory usage-recognition time plane, a memory usage-recognition rate plane or a recognition time-recognition rate plane may be presented. In the case of the plane, the minimum value of the recognition time is displayed, and in the case of the recognition time-recognition rate plane, the minimum value of the memory usage is displayed.

  In the vector quantization of the first embodiment, the recognition rate, the recognition time, and the memory usage have different properties. Therefore, the distance scale normalized by each variance may be used. The variance in this case can be realized by using all or a part of the trial results to obtain the variances of the recognition rate, recognition time, and memory usage values that are the respective elements.

  The present invention may be applied to a system composed of a plurality of devices or an apparatus composed of a single device. The program code of the software realizing the functions of the above-described embodiments is downloaded and executed via a network or the like, or a recording medium on which the program code is recorded is supplied to the system or apparatus, and the computer ( Needless to say, this can also be achieved by reading and executing the program code stored in the recording medium by the CPU or MPU.

  In this case, the program code itself read from the recording medium realizes the functions of the above-described embodiment, and the recording medium on which the program code is recorded constitutes the present invention. As a recording medium for supplying the program code, for example, a floppy (registered trademark) disk, a hard disk, an optical disk, a magneto-optical disk, a CD-ROM, a CD-R, a magnetic tape, a nonvolatile memory card, a ROM, or the like is used. be able to.

  Further, by executing the program code read by the computer, not only the functions of the above-described embodiments are realized, but also an OS running on the computer performs actual processing based on an instruction of the program code. It goes without saying that a case where the functions of the above-described embodiment are realized by performing part or all of the processing, is also included.

  Furthermore, after the program code read from the recording medium is written in a memory provided in a function expansion board inserted in the computer or a function expansion unit connected to the computer, the function expansion is performed based on the instruction of the program code. It goes without saying that the CPU or the like provided in the board or the function expansion unit performs part or all of the actual processing, and the functions of the above-described embodiments are realized by the processing.

It is a block diagram which shows the hardware constitutions of the parameter setting apparatus for setting appropriately the parameter for speech recognition based on the 1st Embodiment of this invention. It is a block diagram which shows the module structure of the parameter setting apparatus of FIG. It is a flowchart which shows the procedure of the process of the parameter setting apparatus of FIG. It is a figure which shows an example of the input / output data of the vector quantization by the vector quantization part 206 of FIG. It is the figure which represented the result of grouping on the two-dimensional plane regarding the recognition time and memory usage in the parameter setting apparatus which concerns on the 2nd Embodiment of this invention.

Explanation of symbols

101 Control memory 101
102 Central processing unit (CPU)
103 Memory 103
104 External storage device 105 Input device 106 Display device 107 Bus 201 Parameter search condition holding unit 202 Parameter selection unit 203 Evaluation data holding unit 204 Speech recognition unit 205 Operation result holding unit 206 Vector quantization unit 207 Quantization result holding unit 208 Display Processing unit 209 Input processing unit

Claims (10)

A parameter setting device for speech recognition,
Trial means for attempting speech recognition using each of a plurality of parameter sets prepared in advance for the operating conditions;
Operation result holding means for holding an operation result including a recognition rate for each trial together with a parameter set corresponding thereto;
A parameter setting device comprising grouping means for grouping operation results for each trial.   2. The parameter setting apparatus according to claim 1, wherein the grouping unit groups the operation results for each trial using vector quantization.   2. The parameter setting apparatus according to claim 1, further comprising a removing unit that removes an inappropriate operation result as a pre-processing of grouping by the grouping unit. A parameter setting method for speech recognition,
A trial step of trying speech recognition using each of a plurality of parameter sets prepared in advance for the operating conditions;
An operation result holding step of holding an operation result including a recognition rate together with a parameter set corresponding to the recognition rate for each trial;
And a grouping step of grouping the operation results for each trial.   5. The parameter setting method according to claim 4, wherein in the grouping step, the operation results for each trial are grouped using vector quantization.   5. The parameter setting method according to claim 4, further comprising a removal step of removing an inappropriate operation result as a pre-processing of grouping by the grouping step. A program for realizing a parameter setting method for speech recognition,
A trial module that attempts speech recognition using each of a plurality of parameter sets prepared in advance for the operating conditions;
An operation result holding module that holds an operation result including a recognition rate for each trial together with a parameter set corresponding thereto;
A program comprising: a grouping module for grouping operation results for each trial.   The program according to claim 7, wherein the grouping module groups the operation results for each trial using vector quantization.   8. The program according to claim 7, further comprising a removal module that removes an inappropriate operation result as a pre-processing of grouping by the grouping step.   A storage medium storing the program according to claim 7 in a computer-readable manner.

Images