JP7501054B2 - Voice recognition device and voice recognition program - Google Patents

Description

The present invention relates to a voice recognition device, etc.

The speech of customers and visitors collected at exhibitions, showrooms, stores, etc. is converted to text using speech recognition. The converted speech of customers and visitors is analyzed for content and used as marketing data. However, speech recognition has the following issues. At exhibitions, a wide variety of sounds are mixed together, such as the voice of the presenter, environmental sounds, noise, and announcements, making it difficult to create an environment in which the speech of visitors can be collected in good quality. Furthermore, when recognizing the collected speech after it is recorded at the venue, appropriate recognition results will not be obtained unless the speech is suitable for speech recognition. If the quality of the collected speech is below a certain level, it is difficult for people to understand the content of the speech, even if attempts are made to improve the quality in post-processing, and the issue is that high accuracy cannot be expected from computer-based speech recognition.

In response to such a situation, Patent Document 1 proposes a speech recognition device that records background sound near the location of the client device performing the sound collection, generates a noise model based on the recorded background sound, performs noise reduction processing on the audio file from the client device based on the generated noise model, and performs speech recognition on the audio file after the noise reduction processing to obtain the recognized text.

JP 2015-135494 A

However, when the sound collection environment changes significantly, such as in an exhibition hall where the number of people changes, it is difficult with conventional technology to always collect sound under optimal conditions from the start to the end of sound collection. The present invention has been made in consideration of this situation. The purpose is to provide a voice recognition device etc. that changes the operating parameters of the microphone and filter processing so that the collected sound is suitable for voice recognition, even if the sound collection environment changes significantly.

A speech recognition device according to one aspect of the present application includes an acquisition unit having a noise reduction filter and acquiring a test speech and a speech including noise corresponding to a test text, a conversion unit recognizing the acquired test speech and converting it into a speech text, a calculation unit comparing the speech text with correct answer data of the test text and calculating a recognition rate, a modification unit modifying operation parameters of the acquisition unit including parameters of the noise reduction filter, a repetitive control unit repeatedly operating the acquisition unit, the conversion unit, the calculation unit and the modification unit until the recognition rate satisfies a predetermined convergence condition, and a repetitive control unit determining the operation parameters with a high recognition rate based on a plurality of recognition rate results calculated using the modified operation parameters. and a determination unit , wherein the repetitive control unit instructs the change unit to change the operating parameter by a first step size, and the repetitive control unit repeatedly operates the acquisition unit, the conversion unit, the calculation unit, and the change unit to search for a first optimal value of the operating parameter at which the recognition rate is maximized, and after the search is completed, the repetitive control unit instructs the change unit to change the operating parameter by a second step size smaller than the first step size, and the repetitive control unit repeatedly operates the acquisition unit, the conversion unit, the calculation unit, and the change unit to search for a second optimal value of the operating parameter at which the recognition rate is maximized and is a value equal to or greater than the maximum value .

In one aspect of the present application, even if the sound collection environment changes significantly, by changing the operating parameters of the microphone and filter processing, it is possible to collect sound suitable for voice recognition.

1 is a block diagram showing an example of a hardware configuration of a voice recognition device; FIG. 2 is an explanatory diagram illustrating an example of a test audio DB. FIG. 2 is an explanatory diagram illustrating an example of a test text DB. FIG. 2 is an explanatory diagram illustrating an example of a parameter DB. FIG. 2 is an explanatory diagram illustrating an example of a voice DB. FIG. 2 is an explanatory diagram illustrating an example of a recognized text DB; 13 is a flowchart showing an example of a procedure for a sound pickup recognition process. 13 is a flowchart showing an example of a procedure for setting an operation parameter; 13 is a flowchart illustrating an example of an evaluation process procedure. FIG. 11 is an explanatory diagram showing an operation example. FIG. 11 is an explanatory diagram showing an operation example. FIG. 2 is an explanatory diagram illustrating an example of an environment DB. 2 is a block diagram showing an example of functional units included in the voice recognition device; FIG.

The embodiment will be described below with reference to the drawings. FIG. 1 is a block diagram showing an example of the hardware configuration of a voice recognition device. The voice recognition device 1 is configured with a notebook PC (Personal Computer), a tablet computer, a smartphone, a smart speaker, or the like. The voice recognition device 1 includes a control unit 11, a main memory unit 12, an auxiliary memory unit 13, a voice input unit 14, a voice output unit 15, a communication unit 16, and a reading unit 17. The control unit 11, the main memory unit 12, the auxiliary memory unit 13, the voice input unit 14, the voice output unit 15, the communication unit 16, and the reading unit 17 are connected by a bus B. The voice recognition device 1 may be configured using a multicomputer consisting of multiple computers, a virtual machine virtually constructed by software, or a quantum computer. Furthermore, all or part of the functions of the voice recognition device 1 may be realized by a cloud service.

The control unit 11 has one or more arithmetic processing devices such as a CPU (Central Processing Unit), an MPU (Micro-Processing Unit), or a GPU (Graphics Processing Unit). The control unit 11 reads and executes a control program 1P (voice recognition program) stored in the auxiliary storage unit 13, thereby performing various information processing, control processing, and the like related to the voice recognition device 1, and realizing functional units such as an acquisition unit, a conversion unit, a calculation unit, a change unit, or a repetition control unit.

The main memory unit 12 is a static random access memory (SRAM), a dynamic random access memory (DRAM), a flash memory, etc. The main memory unit 12 mainly temporarily stores data required for the control unit 11 to execute arithmetic processing.

The auxiliary storage unit 13 is a hard disk or SSD (Solid State Drive) or the like, and stores the control program 1P and various DBs (Databases) necessary for the control unit 11 to execute processing. The auxiliary storage unit 13 stores a test voice DB 131, a test text DB 132, a parameter DB 133, a voice DB 134, and a recognition text DB 135. The auxiliary storage unit 13 may be an external storage device connected to the voice recognition device 1. The various DBs and the like stored in the auxiliary storage unit 13 may be stored in a database server or cloud storage different from the voice recognition device 1.

The audio input unit 14 is a microphone device that picks up audio. The audio input unit 14 includes an audio processing circuit that realizes a noise reduction function and the like. The audio output unit 15 is a speaker device. The audio output unit 15 outputs audio under the control of the control unit 11. There may be multiple audio input units 14 and multiple audio output units 15.

The communication unit 16 communicates with other computers via the network N. The communication unit 16 may download the control program 1P from other computers via the network N, etc., under the control of the control unit 11. The reading unit 17 reads portable storage medium 1a, including CD (Compact Disc)-ROM and DVD (Digital Versatile Disc)-ROM. The control unit 11 may read the control program 1P from the portable storage medium 1a via the reading unit 17 and store it in the auxiliary storage unit 13. The control unit 11 may also read the control program 1P from the semiconductor memory 1b.

The voice recognition device 1 recognizes the voice picked up by the voice input unit 14 using an external voice recognition engine provided by another computer or cloud service. When an external voice recognition engine is not used, the voice recognition device 1 performs feature extraction, phonemic analysis, analysis, and syntactic analysis on the picked up voice, and converts the input voice into a string of kana. Furthermore, the voice recognition device 1 converts the string of kana into a string of mixed kanji and kana. The auxiliary storage unit 13 stores a phonetic dictionary, a syntactic dictionary, a word dictionary, a kana-kanji conversion dictionary (none of which are shown in the figure), and the like, which are used when performing these analyses.

Next, the database used by the voice recognition device 1 will be described. FIG. 2 is an explanatory diagram showing an example of a test voice DB. The test voice DB 131 stores test voices used when adjusting the operating parameters of the voice input unit 14. The test voice DB 131 includes a VoID string, a test voice string, and a TxtID string. The VoID string stores a VoID that can uniquely identify the test voice. The test voice string stores the test voice in a binary format. Note that the test voice may be stored as a voice file in the auxiliary storage unit 13, and the file name of the voice file may be stored in the test voice string. The data format of the test voice may be MP3 format, WAV format, WMA format, etc. The TxtID string stores the main key TxtID of the test text DB 132 that stores text corresponding to the test voice.

Figure 3 is an explanatory diagram showing an example of a test text DB. Test text DB 132 stores text that is a conversion of test audio. Test text DB 132 includes a TxtID string and a text string. The TxtID string stores a TxtID that uniquely identifies the text. The text string stores the text. In the example of Figure 3, the text string is written in hiragana, with a separator / between each word.

Figure 4 is an explanatory diagram showing an example of a parameter DB. The parameter DB 133 stores the actual values of the operation parameters in the voice input unit 14. The parameter DB 133 includes a PID column, a Mic No. column, a Power (voice/environment) dB column, a noise reduction (dB/sensitivity/band) column, a Mic (angle/range) column, a filter 1 column, and a recognition rate column. The PID column stores a PID that can uniquely identify the operation parameters. The Mic No. column stores a Mic No. that uniquely identifies the voice input unit 14. The Power (voice/environment) dB column stores the power of the voice to be picked up (human speech) and the power of the environmental sound. The noise reduction (dB/sensitivity/band) column stores the amount of noise attenuation (dB), the power value (sensitivity) of the sound considered to be noise, and the frequency width (band) to be smoothed before and after the noise frequency. The Mic (angle/range) column stores the directivity setting, the central angle (angle) of the directivity, and the angle range (range) for collecting sound. The Filter 1 column stores the coefficients and weights of the first filter. The filter (noise reduction filter) can be a low-pass filter, a high-pass filter, a band-pass filter, a band elimination filter (BEF: Band elimination filter, BRF: Band reject filter), a finite impulse response filter (FIR: Finite Impulse Response), a radio impulse response digital filter (IIR: Infinite Impulse Response), etc. A plurality of filters may be used. In that case, a column of the parameter DB is provided according to the number of filters. The recognition rate column stores the recognition rate when the operation parameters of the record are set, a test voice is output from the voice output unit 15, the voice is collected by the voice input unit 14, and the voice recognition is performed.

Figure 5 is an explanatory diagram showing an example of an audio DB. Audio DB 134 stores collected audio. Audio DB 134 includes a collection ID string, a date string, a start/end time string, a PID string, and an audio string. The collection ID string stores a collection ID that uniquely identifies the collected audio. The date string stores the date of collection. The start/end time string stores the time when collection started and the time when collection ended. The PID string stores the operating parameters used when collecting the audio. The audio string stores the collected audio in binary format. The audio data format is the same as the test audio described above.

Figure 6 is an explanatory diagram showing an example of a recognized text DB. The recognized text DB 135 stores text obtained by speech recognition of collected voice. The recognized text DB 135 includes a collection ID string and a text string. The collection ID string stores the collection ID of the voice data corresponding to the text. The text string stores the text obtained by speech recognition.

Next, the process performed by the voice recognition device 1 will be described. FIG. 7 is a flowchart showing an example of the procedure of the voice collection recognition process. The voice collection recognition process is a process of recognizing the voice collected by the voice input unit 14 and storing the kanji and kana mixed character string (text) obtained as the recognition result in the recognition text DB 135. The control unit 11 of the voice recognition device 1 sets the operation parameters of the voice input unit 14 (step S1). The control unit 11 collects voice via the voice input unit 14 (step S2). The control unit 11 stores the collected voice in the voice DB 134 (step S3). The control unit 11 performs voice recognition using an external voice recognition engine (step S4). The control unit 11 stores the recognition text returned from the voice recognition engine in the recognition text DB 135 (step S5). The control unit 11 determines whether or not to end the process (step S6). If the control unit 11 determines not to end the process (NO in step S6), it determines whether or not to readjust the operation parameters (step S7). For example, the control unit 11 uses an interval timer or the like to determine that readjustment should be performed every time a predetermined time has elapsed. The control unit 11 also determines that readjustment should be performed when an instruction input from the user is received. If the control unit 11 determines that readjustment should be performed (YES in step S7), the process returns to step S1. If the control unit 11 determines that the operating parameters should not be readjusted (NO in step S7), the process returns to step S2. If the control unit 11 determines that the process should be terminated (YES in step S6), the process ends. Note that if the voice recognition engine supports streaming, steps S2 to S5 may be executed in parallel.

Figure 8 is a flowchart showing an example of the procedure of the operation parameter setting process. The operation parameter setting process corresponds to step S1 in Figure 7. The control unit 11 of the voice recognition device 1 sets an initial value to the operation parameter of the voice input unit 14 (step S21). The initial value may be a predetermined value or a value input by the user each time. In addition, when the operation parameter is readjusted, the previous value may be the initial value. The control unit 11 sets the width (step width) for varying the operation parameter to large (step S22). The step width (first step width) is predetermined for each operation parameter and stored in the auxiliary storage unit 13. The control unit 11 evaluates the operation parameter (step S23). The control unit 11 judges whether or not the operation parameter has converged (step S24). Each time the control unit 11 changes the operation parameter, the evaluation value obtained in step S23 is stored in a temporary storage area, and judges whether or not the operation parameter has converged based on the change in the evaluation value. The temporary storage area is provided in the main storage unit 12 or the auxiliary storage unit 13. The convergence condition for determining whether or not the convergence has occurred is stored in the auxiliary storage unit 13 in advance. For example, the convergence condition is set to a decrease in the recognition rate five consecutive times. When the control unit 11 determines that the convergence has not occurred (NO in step S24), it changes the operation parameter (step S25). The control unit 11 increases or decreases the value of the operation parameter by the increment width. The control unit 11 returns the process to step S23. When the control unit 11 determines that the convergence has occurred (YES in step S24), it sets the increment width to small (step S26). The value of the operation parameter when the control unit 11 determines that the convergence has occurred is the first optimal value. As in the case where the increment width is large, the increment width (second increment width) is determined in advance for each operation parameter and stored in the auxiliary storage unit 13. The control unit 11 evaluates the operation parameter (step S27). The control unit 11 determines whether or not the convergence has occurred (step S28). The determination method is the same as in step S24. When the control unit 11 determines that the convergence has not occurred (NO in step S28), it changes the operation parameter (step S29). The control unit 11 increases or decreases the value of the operation parameter by the step size. The control unit 11 returns the process to step S27. If the control unit 11 determines that convergence has occurred (YES in step S28), it sets the operation parameter to an optimal value (second optimal value) (step S30). The evaluation value obtained in step S27 is stored in association with the operation parameter. Among the stored evaluation values, the value of the operation parameter with the highest evaluation, for example, the maximum value, is set as the optimal value. The control unit 11 returns the process to the caller. Note that if a high recognition rate (for example, 95%) is obtained in the search with a large step size (steps S23 to S25), the search with a small step size (steps S26 to S29) may be omitted.

Figure 9 is a flowchart showing an example of the procedure of the evaluation process. The evaluation process corresponds to steps S23 and S27 in Figure 8. The control unit 11 of the voice recognition device 1 collects the test voice (step S41). The control unit 11 outputs the test voice stored in the test voice DB 131 from the voice output unit 15. The control unit 11 collects the output test voice by the voice input unit 14. The test voice may be one that is spoken by the user on the spot. The control unit 11 performs voice recognition (step S42). The control unit 11 compares the text obtained as a result of the voice recognition with the correct text stored in the test text DB 132 (step S43). The control unit 11 performs the comparison on a word-by-word basis. The control unit 11 calculates the number of correctly recognized words (number of correct words), the number of incorrectly recognized words (number of incorrectly recognized words), and the number of words that were included in the test voice but whose existence was not recognized (number of dropped words). The control unit 11 calculates the recognition rate from the number of correct words, the number of misrecognized words, and the number of omitted words obtained as the comparison result (step S44). The recognition rate is calculated using the following formula (1).

Recognition rate = number of correct words / (number of correct words + number of misrecognized words + number of omitted words) … (1)

The control unit 11 stores the recognition rate as an evaluation value in a temporary storage area (step S45), and returns the process to the caller. The operation parameter setting process shown in FIG. 8 is a process for solving an optimization problem. In this embodiment, various known methods for solving optimization problems can be used. For example, the Bayesian optimization method is used in this embodiment.

Next, an example of the operation of the operation parameter setting process is shown. Figs. 10 and 11 are explanatory diagrams showing the example of the operation. Figs. 10 and 11 show the operation parameters in correspondence with the recognition rate obtained as the evaluation result of the operation parameters. The contents shown in Figs. 10 and 11 are stored in a temporary storage area in the evaluation process. Fig. 10 shows the progress of the process when the step size is large, and Fig. 11 shows the progress of the process when the step size is small. In Figs. 10 and 11, the number of records increases from top to bottom as the process progresses.

In the example of Figure 10, the recognition rate is 87.1% with the initial values of the operation parameters. In the next evaluation after changing the operation parameters, the recognition rate is 87.9%. Since the recognition rate tends to decrease in subsequent evaluations, the process is terminated and the value of the operation parameters corresponding to a recognition rate of 87.9% becomes the provisional optimal value. The process is terminated, for example, when the recognition rate decreases a predetermined number of times in succession.

Referring to Figure 11, the processing when the step size is small will be explained. In the example of Figure 11, the value of the operation parameter is changed from around the provisional optimal value and evaluation is performed. The maximum recognition rate is 93.2%, and since the recognition rate thereafter has a downward trend, the processing is terminated and the operation parameter corresponding to 93.2% is determined to be the optimal value.

When the voice recognition device 1 includes multiple voice input units 14, the control unit 11 executes the operation parameter setting process (Fig. 8) and evaluation process (Fig. 9) for each voice input unit 14 individually. In addition, when the voice input unit 14 includes a microphone array that causes multiple physical microphones to function as one logical microphone, the control unit 11 sets the operation parameters of the individual microphones that make up the microphone array so that the recognition rate of the test voice picked up by the microphone array is maximized.

The present embodiment has the following effects. When the voice recognition device 1 determines that a predetermined condition is satisfied, it updates the operating parameters of the voice input unit 14. The operating parameters are updated in the operating environment of the voice recognition device 1, so that even if the sound collection environment changes significantly, it is possible to change the operating parameters to appropriate values. This makes it possible to collect voice suitable for voice recognition. In addition, since the collected voice is stored, even if there is an error in the recognized text, the user can correct the text by referring to the stored voice. In addition, since the operating parameters are set in the actual environment before starting sound collection, it is possible to obtain appropriate voice recognition results even if a different voice recognition engine is used, or even if the same voice recognition engine has changed its operating characteristics due to a version upgrade.

(Modification)
In the operation parameter setting process, in order to converge the process quickly, the initial value of the operation parameter is set to a value estimated to be close to the optimal value. Therefore, information (environment information) on the environment in which the operation parameter with a proven track record was used is stored. As the initial value of the operation parameter setting process, the value of the operation parameter associated with an environment similar to the environment in which it is used is set. FIG. 12 is an explanatory diagram showing an example of the environment information DB. The environment information DB 136 is stored in the auxiliary storage unit 13. The environment information DB 136 includes a PID string, a scene string, a name string, a venue type string, a venue name string, and a noise level string. The PID string stores a PID that identifies the corresponding operation parameter. The PID is the main key of the parameter DB 133. The scene string stores a scene in which the operation parameter was used. The scene is, for example, a large-scale exhibition, a small-scale exhibition, and a permanent exhibition. A large-scale exhibition is an exhibition held at an exhibition hall where dozens of exhibitors gather. A small-scale exhibition is an exhibition held in a rented hall where several exhibitors gather. A permanent exhibition indicates an exhibition in a showroom or the like. The name column stores the name of an exhibition, etc. The venue type column stores the type of venue. The venue type is, for example, an exhibition hall, a rental hall, or a showroom. The venue name column stores the name of the venue. The noise level column stores the sound pressure (dBSPL) of the ambient noise. The unit is dB.

When setting the initial values of the operation parameters, the user searches the environment information DB 136 and selects an environment similar to the environment to be used. The control unit 11 searches the parameter DB 133 using the PID included in the selected environment information as a key. For example, if recording again at an exhibition where sound recording was performed in the past, the environment information DB 136 is searched for the exhibition name. If recording again at a venue where sound recording has been performed in the past but not at a scene where sound recording has been performed in the past, the environment information DB 136 is searched for the venue name. If recording has been performed at a similar venue where sound recording has not been performed in the past but has been performed in a similar venue, the search is performed by venue type. If the scene or venue is new, records with similar sound pressure values of ambient noise are searched for. The control unit 11 performs the operation parameter setting process using the operation parameters included in the records found in the search as the initial values. The operation parameters to be used as initial values may be selected by other methods rather than by search. The attributes of the environment where sound is collected (scene, name, venue type, venue name, noise level, etc.) are compared with the name string, venue type string, venue name string, and noise level string of the environment information DB136 to calculate the similarity, identify the record with the value most similar to the attributes of the environment where sound is collected, and set the operating parameters of the identified record as the initial values.

In this modified example, the initial values of the operating parameters are set to values that are estimated to be close to the optimal values based on past performance, and the operating parameter setting process is then performed. It is expected that the process will converge quickly and the optimal values of the operating parameters will be determined.

Figure 13 is a block diagram showing an example of functional units of a voice recognition device. The voice recognition device 1 includes, as functional units, an acquisition unit 11a, a conversion unit 11b, a calculation unit 11c, a change unit 11d, a determination unit 11e, and a repetition control unit 11f. Each of these functional units is realized by the control unit 11 operating based on a control program 1P.

The acquisition unit 11a acquires test speech corresponding to the test text and speech including noise. The conversion unit 11b recognizes the test speech acquired by the acquisition unit 11a and converts it into spoken text. The calculation unit 11c compares the spoken text with the correct answer data of the test text and calculates a recognition rate. The modification unit 11d modifies the operation parameters of the acquisition unit, including the parameters of a noise removal filter, based on the calculated recognition rate. The determination unit 11e determines the operation parameters with a high recognition rate based on the results of multiple recognition rates calculated using the modified operation parameters. The repetitive control unit 11f repeatedly operates the acquisition unit, the conversion unit, the calculation unit, and the modification unit until the recognition rate satisfies a predetermined convergence condition.

The technical features (constituent elements) described in each embodiment can be combined with each other, and by combining them, new technical features can be formed.
The embodiments disclosed herein are illustrative in all respects and should not be considered as limiting. The scope of the present invention is defined by the claims, not by the above meaning, and is intended to include all modifications within the scope and meaning equivalent to the claims.

REFERENCE SIGNS LIST 1 Voice recognition device 11 Control unit 11a Acquisition unit 11b Conversion unit 11c Calculation unit 11d Change unit 11e Determination unit 11f Repetition control unit 12 Main memory unit 13 Auxiliary memory unit 131 Test voice DB
132 Test Text DB
133 Parameter DB
134 Audio DB
135 Recognition Text DB
136 Environmental Information DB
14 Voice input unit 15 Voice output unit 16 Communication unit 17 Reading unit 1P Control program 1a Portable storage medium 1b Semiconductor memory

Claims (5)

an acquisition unit having a noise reduction filter and acquiring a test speech corresponding to a test text and a noisy speech;
A conversion unit that recognizes the acquired test voice and converts it into spoken text;
a calculation unit that compares the spoken text with correct answer data of the test text and calculates a recognition rate;
a change unit that changes operating parameters of the acquisition unit, including parameters of the noise reduction filter;
a repetitive control unit that repeatedly operates the acquisition unit, the conversion unit, the calculation unit, and the change unit until the recognition rate satisfies a predetermined convergence condition;
a determination unit that determines an operation parameter having a high recognition rate based on a plurality of results of the recognition rates calculated using the changed operation parameter ,
the repetitive control unit instructs the change unit to change the operation parameter by a first step size;
the repetitive control unit repeatedly operates the acquisition unit, the conversion unit, the calculation unit, and the change unit to search for a first optimum value of the operation parameter that maximizes the recognition rate;
After the search is completed, the repetitive control unit instructs the change unit to change the operation parameter by a second step size smaller than the first step size;
The repetitive control unit repeatedly operates the acquisition unit, the conversion unit, the calculation unit, and the change unit to search for a second optimal value of the operation parameter that maximizes the recognition rate and is equal to or greater than the maximum value. 2. The speech recognition device according to claim 1, wherein the recognition rate is calculated based on a result of comparison between words constituting the spoken text and words constituting correct answer data. 3. The speech recognition device according to claim 1, wherein a Bayesian optimization method is used to search for the first optimum value and the second optimum value. a storage unit that stores environmental information regarding an environment in which the acquisition unit acquired the voice in association with the second optimum value,
4. The speech recognition device according to claim 1, wherein the repetitive control unit acquires the environmental information, reads the second optimum value associated with the acquired environmental information from the storage unit, and controls the change unit to set the read second optimum value as an initial value of the operation parameter. A computer including an acquisition unit having a noise reduction filter and configured to acquire speech,
an acquisition process for acquiring a speech corresponding to a test text and a speech including noise by the acquisition unit;
A conversion process that recognizes the acquired voice and converts it into spoken text;
A calculation process of comparing the spoken text with the correct answer data of the test text and calculating a recognition rate;
A modification process for modifying operating parameters of the acquisition unit, including parameters of the noise reduction filter;
A process of determining an operation parameter having a high recognition rate based on a plurality of recognition rate results calculated using the changed operation parameter;
A speech recognition program for causing the computer to execute the above ,
repeatedly executing the acquisition process, the conversion process, the calculation process, and the change process of changing the operation parameter by a first step size until the recognition rate satisfies a predetermined convergence condition, thereby searching for a first optimum value of the operation parameter that maximizes the recognition rate;
After the search is completed, the acquisition process, the conversion process, the calculation process, and the change process for changing the operation parameter by a second step size smaller than the first step size are repeatedly executed until the recognition rate satisfies a predetermined convergence condition, thereby searching for a second optimum value of the operation parameter at which the recognition rate is maximized at a value equal to or greater than the maximum value.
A speech recognition program comprising :

Images