JPH0580794A - Speech recognition device - Google Patents

Abstract

PURPOSE:To provide the speech recognition device having a function which starts an exceptional process by interrupting a normal recognizing process if misrecognition occurs during speech input recognition and returns to the normal recognizing process only when a vocabulary for the exceptional process is recognized. CONSTITUTION:This speech recognition device is provided with an input means 1 for accepting a speech input, a recognizing means 2 which recognizes the speech input, a next process determining means 3 which determines a next process according to the recognition result, an exceptional process recognizing means 4 which performs the exceptional process when misrecognition occurs during the speech recognition, an output means 5 which outputs the recognition result, a normal information storage means 7 which is referenced by the recognizing means 2, and an exceptional information storage means 8 which is referenced by the exceptional process recognizing means 4. If the misrecognition occurs, the exceptional information storage means 8 is referred to for the recognition, so a user need not be questioned from the beginning and the speech recognition device suitable for high-speed and continuous speech input operation is provided.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition apparatus having means for controlling voice recognition by performing voice recognition according to an input voice signal when a voice input is accepted.

[0002]

2. Description of the Related Art In recent years, a voice response recognition device by voice input using a telephone line has been introduced in the financial industry and the distribution industry. Services such as balance inquiry and order entry are provided to users using these devices.
Recently, a voice recognition device has been developed for home electric appliances and in-vehicle products, and has been applied to car audio, VTR reservation, telephone and the like.

FIG. 6 shows the configuration of a conventional voice recognition device. In FIG. 6, reference numeral 51 is input means for receiving voice input, 52 is recognition means for performing voice recognition using the input voice signal, 53 is recognition information storage means referred to when performing recognition, and 54 is recognition result. Next processing determining means for determining the next processing, and 55 is an output means for displaying the processing result.

The operation of the speech recognition apparatus configured as above will be described below. First, the user
Input prompt message from the output means 55 of the voice recognition device (guidance by voice or guidance display by display device)
Follow the steps below to input the voice. The input means 51 accepts this input and performs the processing necessary for the recognition processing by the next recognition means 52. First, the input analog signal is converted into a digital signal. Normally, a sampling frequency of 8 to 10 kHz is used, and quantization is performed with a value of 8 to 16 bits. In addition, high frequency components and noise components are removed as needed. The recognition means 52 selects the most similar information at a certain threshold value or more by pattern matching with the input signal and the voice information in the recognition information storage means 53, and returns the result as a uttered word. The next process determining means 54 determines the input content to be requested next or the content to be output from the output means 55 by looking at the result.

When this voice recognition device is applied to a telephone, it is necessary to utter a number of about 10 digits in order to input the area code. At least a number from 0 (pronounced zero) to 9 (pronounced kyu) is uttered in order, the content is recognized, converted into a dial pulse or tone signal, and a call is made. The recognition result can be confirmed by the user by having a display device attached to the telephone and means for indicating the recognition result using a voicing output device, if necessary (Japanese Patent Laid-Open No. 63-33796). No. The same applies to the input of a password other than the telephone number.

In order to shorten the time for inputting a 10-digit number as much as possible, one number is uttered, recognition is performed, the result is displayed, and at the same time, the next input is awaited. If it is recognized, it is converted into the corresponding numerical information and waits for the next input. When the utterance is small, for example, when the recognition cannot be made, a request for rewording is made to wait for the next input. As you become accustomed to voice input, you will be able to input without looking at information on the display device. Therefore, it is possible to input the telephone number by speaking in order without removing the line of sight from the corresponding number string in the telephone directory. The input result may be confirmed after the completion of all utterances, and it is not necessary to look at the display device or the like every time one by one is input.

[0007]

However, in the conventional configuration, the next input is continuously requested even when the 10-digit number is uttered in order and the recognition is not possible in the middle (in this case, In other words, if the user does not pay attention to the recognition result, there is a problem that an input different from the actual intention is made.

In the specific example, the utterance data is "zero",
If it is a ten-digit number such as "ichi", "ni", "san", "yon", "go", "roku", "nana", "hachi", "kyu", then "yon" By the way, when recognition is not possible due to a small amount of utterance, etc., the recognition result is "Zero", without recognition at "Yon".
The result is "ichi,""ni,""san,""go,""roku,""nana,""hachi," and "kyu," resulting in missing information about "yon."

After all the input is completed, the user looks at the result output to the display device or the like and finds that the information of "Yon" is missing, and the user must input again from the beginning. There were some problems that didn't happen.

The present invention solves the above-mentioned conventional problems. When an error occurs in utterance recognition, the normal voice input is interrupted and the next input is made until a word for exception processing is uttered. It is an object of the present invention to provide a voice recognition device that has a function that does not allow it, so that even if a recognition error occurs, it is not necessary to redo all the inputs.

[0011]

In order to achieve this object, the present invention provides an input means for receiving a voice input, a recognition means for recognizing the voice input accepted by the input means, and a result recognized by the recognition means. Next processing determining means for determining the next processing based on the following, an exception processing recognizing means to be executed when a recognition error occurs during voice recognition, an output means for outputting a recognition processing result, and a normal operation referred to by the recognizing means. It has information storage means and exception information storage means referred to by the exception handling recognition means.

[0012]

With this configuration, when the recognition means produces a recognition error, the next processing determination means interrupts the normal recognition processing, activates the exception processing recognition means, and refers to the exception information storage means. By having a means for performing recognition,
It is possible to provide a voice recognition device suitable for high-speed continuous voice input, which does not require the user to reword everything from the beginning when a recognition error occurs.

[0013]

【Example】

(First Embodiment) A first embodiment of the present invention will be described below with reference to the drawings. FIG. 1 shows the configuration of a speech recognition apparatus according to the first embodiment of the present invention. Figure 1
In 1, the input means 1 mainly accepts a voice input, and the reference numeral 2 is a recognition means, which detects a utterance section and performs voice recognition within the utterance section. Reference numeral 3 denotes a next process determining means, which determines the content to be executed next according to the result of voice recognition, and controls each means in the voice recognition device. Reference numeral 4 denotes an exception processing recognition means, which is executed when a recognition error occurs during voice recognition. An output unit 5 outputs the recognition processing result and the like. Reference numeral 6 denotes a voice information storage means for storing voice information to be recognized. Reference numeral 7 is a normal information storage means, which is referred to in the voice information storage means 6 when the recognition means 2 executes a recognition process. Reference numeral 8 denotes an exception information storage means, which is referred to by the exception processing recognition means 4 when performing exception processing.

The operation of the speech recognition apparatus configured as described above will be described with reference to FIG. First, the necessary settings are made to recognize the first vocabulary. For the purpose of recognizing the numerical information, the input means 1 is activated, and the output means 5 outputs an input prompt to the user. Then, it waits for the user to speak the vocabulary. The detection of utterance is performed as follows. It is assumed that the voice input gain has an initial value or an appropriate value set previously. When the user starts voice input to the voice recognition device, voice level detection is performed, a voice level above a certain threshold is detected, and the start of utterance is confirmed. After confirming the start of utterance, the voice level detection allows the end of utterance to be known when the input reaches a level below a certain threshold.

With respect to the ambient noise at the time of input, there is a means for using a microphone having a strong directivity as the input means 1 so that another ambient signal is not mixed in as much as possible.
There is also a means for changing the input gain by detecting the voice level.
In addition, voice level detection measures the ambient noise level,
It is possible to adopt a configuration in which a signal component necessary for voice recognition is obtained by performing voice signal processing that subtracts the noise level.

When a proper vocabulary is uttered in accordance with the instruction of the guidance, the vocabulary information input from the input means 1 is pattern-matched with the voice information stored in the voice information storage means 6 to perform voice recognition. .. The recognizing means 2 performs voice recognition, converts the corresponding information from the voice signal into information such as appropriate numeral and character information, and outputs it to the next information determining means 3. Then, the output means 5 outputs a necessary message to the user. The guidance can be provided only by vocalization information or by display output.

The voice information to be pattern-matched (hereinafter abbreviated as matching) is divided and stored in groups called groups for each usage type. For example, the first group is operation command names, the second group is numbers used for inputting telephone numbers and the like, and the third group is grouped and stored in vocabulary for detailed commands. In this way, by grouping the vocabulary information according to the purpose of use, work is performed to find the one that is closest to the most vocalized voice information in the designated group. Of course, each information is coded into information that facilitates matching in a short time. Further, a threshold value is provided for this matching, and the recognition means 2 has means for outputting a voice recognition result indicating that there is no matching vocabulary unless the threshold value is not less than a certain value.

When the input voice exceeds the threshold of a specific input level, the start of utterance having vocabulary information is detected, and when it falls below the threshold, the end of utterance is confirmed. The input voice in the utterance section is sampled at several kHz (about 8 kHz to 10 kHz) and quantized at about 8 to 16 bits to be digitized. Speech recognition uses this digital value to find the LPC cepstrum coefficient and compares it with the parameter amount of the standard vocabulary using this as a characteristic parameter to find the closest one of the vocabulary given by the group number. Find out from inside. This is matched throughout the vocalization section, and finally the closest one is recognized as the vocalized vocabulary. Although there are various methods other than the LPC cepstrum coefficient, a method is generally used in which the digital value obtained by the above-described sampling is subjected to various signal processings to extract characteristic parameters and matching is performed.

Once the recognition is done and the appropriate vocabulary is selected,
Control corresponding to this vocabulary information is performed. Designated dial pulses, tone signals are emitted, and various control signals are generated. Then, it outputs a message prompting the next input and waits for the next voice input.

The recognizing means 2 is not always able to correctly recognize the uttered vocabulary. If the vocabulary is incorrect, or if the utterance is interrupted in the middle, the input level is too low for voice recognition (the utterance is confirmed but the input level is too low for recognition), too high, or When the utterance is too long, or when a utterance other than the recognizable vocabulary is received, the output means 5 is provided to display or output as an error message output.

An example of the contents of the result information when a recognition error occurs is shown in FIG. As shown in FIG. 3, when a recognition error occurs, the cause can be known by obtaining the error number. Therefore, when the input is too small, a message indicating that the user wants to speak a little louder is output and the voice input gain is changed. When the input is too large, as in the case where the input is too small, a message indicating that a smaller voice is desired is output and the voice input gain is changed. In addition, although the utterance was confirmed, if the utterance was not recognized, a message prompting the user to utter again is output. In this way, the process when the vocabulary cannot be recognized at the time of recognition is defined as an exception process. In addition, when the voice input necessary for voice recognition is obtained and matching with the corresponding vocabulary is performed,
Even when it is determined that there is no corresponding vocabulary, the result that the vocabulary cannot be recognized is returned in the same manner.

FIG. 4 shows the procedure of speech recognition during exception processing.
Shown in. When the next process determining means 3 receives the error output from the recognizing means 2, the recognizing means 2 is deactivated, the exception information storing means 8 is set, and the exception processing recognizing means 4 is activated.
Then wait for the next voice input. And the exception information storage means 8
Matching is performed between the input voice information and the voice information for exception processing in the above. The recognition result is returned to the next process determining means 3. When the next processing determination means 3 is recognized as a vocabulary for exception processing, the next processing determination means 3 deactivates the exception processing recognition means 8 and activates the recognition means 2 again. Otherwise, exception handling continues. If a recognition error occurs during exception handling recognition, correct input is required until the correct vocabulary can be recognized. Therefore, it also has a function to correct input by changing the input level. Unless the vocabulary for exception handling is uttered, it does not return to normal processing, so you will not lose track of the input in the middle (it stops when an error occurs).

The voice information storage means 6 stores the normal case,
Two types of audio information in the case of exception processing are stored.
This configuration is shown in FIG. The recognition means 2 refers to the normal information storage means 7, and the exception processing recognition means 4 refers to the exception information storage means 8. FIG. 5 shows an example in which numeric information is stored in the normal information storage means 7 and vocabulary information for input restart and correction is stored in the exception information storage means 8. Inside the normal information storage means 7, as described above, a plurality of pieces of information are grouped and stored in addition to the numerical information. On the other hand, the exception information storage means 8 stores vocabulary information for restarting voice input.

In the exception information storage means 8, a plurality of vocabularies that are not used for normal input are prepared. The vocabulary of exception information is
If the vocabulary for normal input is similar, the user may be confused. In the exception processing that operates only when the recognition cannot be performed, there is an effect of notifying the user that a recognition error has occurred and consciousness of returning to the normal processing. If the vocabulary for exception processing and the vocabulary for normal processing are similar, it is difficult to understand whether the recognition error occurred and a specific vocabulary was uttered, and then it returned to normal processing or is still in exception processing. .. Therefore, to distinguish the two vocabularies,
Care must be taken when selecting the vocabulary of two types of information.

As the information stored in the exception information storage means 8, a plurality of similar vocabularies are prepared. In FIG. 5, information such as “saikai”, “saiyonryuku”, “again”, “more” is shown. By preparing a plurality of vocabularies that are similar to each other, flexibility is provided in the vocabulary when returning from exception handling. There is a way to prepare one or several specific vocabulary, but if the number is small, if the user forgets the vocabulary, it can not return to normal processing unless the vocabulary is uttered, which is inconvenient. It may cause you to feel. Also, adopting all standard words in the vocabulary spoken by the user gives the impression that spoken words are different from the words that control the device. By increasing the vocabulary to the extent that it is as comfortable as possible, it is possible to eliminate regional differences and supply equipment that can be used in a wide range.

Further, when the exception information storage means 8 is composed of a rewritable memory or an additional memory, the standard vocabulary information is held in the main body memory and the local vocabulary information is supplied by the additional memory or the like. By taking this, it is possible to deal with this without taking a method of collecting all the local information in advance. Having multiple vocabularies for the same information is effective not only in exceptional information but also in normal information. However, in the normal information, it is assumed that the more the number of objects to be controlled by voice increases, the more information corresponding to the local spoken language will increase, and thus the feasibility is low. On the other hand, if it is limited to exception information, the number of vocabulary is limited,
In addition, since the target of use is clear, it is easy to handle.
In addition, "mistake", "sour", "more"
Even if the vocabulary such as "is" is a word that can be easily spoken without being conscious of it, it is very effective to use only these vocabulary as exceptional information and construct voice information considering regional characteristics.

A case where a plurality of pieces of numerical information are continuously input by voice will be described. This includes telephone number input, personal identification number input, program reservations for various audio equipment, etc. Since there are multiple input elements, it is necessary to shorten the time until the next input as much as possible in order to complete these inputs. That is, it is necessary to multiplex the recognition process for starting recognition of the next vocabulary information at the same time that one vocabulary information is recognized. On the other hand, when uttering at high speed continuously, it is easy to be affected by shortness of breath and intonation, unlike when singing one by one separately. Therefore, in high-speed continuous voice input, the possibility that voice information for recognizing in the middle cannot be obtained becomes considerably higher than in the case of single utterance. Even though it is said to be high-speed continuous voice input, it is okay to take a short break on the way, but it may be easier to say numerical information such as a telephone number that is memorized in your head at a stretch.

As described above, according to this embodiment, when a plurality of high-speed voice continuous inputs are made, when a recognition error occurs in the middle, the normal recognition process is interrupted and the exception process is executed. According to the configuration, it is possible to configure an easy-to-use device that does not cause loss of information on the way and only needs to restate only the information after the time when the error occurs.

(Second Embodiment) A second embodiment of the present invention will be described below. The constituent elements are the same as in the first embodiment. However, the next process determining means 3 is the output means 5 and the input means 1.
Both of them are controlled.

The output means 5 is provided with a voice synthesis function to perform voice synthesis, and voice input is prohibited while the voice guidance is being output from the output means 5. The purpose is to eliminate the noise component (which becomes noise for the input for voice recognition) due to the voice guidance output, and to reduce the erroneous input due to the user's mistake.

When the user becomes accustomed to voice input and demands high-speed input, a vocabulary for recognizing a command to reduce the amount of voice synthesis (to shorten a message) and a command to stop voice synthesis are provided, and these commands are provided. When a vocabulary is recognized, it becomes easier to use if it has a function of changing, stopping, and resuming the contents of speech synthesis according to the contents. Similarly, it is possible to change the display content or cancel the display.

As described above, by using the operation command information of the device in the voice recognition device, it is possible to construct a device that provides necessary information in a desired form according to the purpose of use.

[0033]

As described above, according to the present invention, the recognizing means for recognizing a voice input, the next processing determining means for deciding the next processing based on the recognition result, and the following processing when a recognition error occurs during voice recognition By having the exception processing recognition means to execute, the normal information storage means referred to by the recognition means, and the exception information storage means referred to by the exception processing recognition means, when the result of recognition error is produced, the next processing determination means By disabling the normal recognition processing, activating the exception processing recognition means, and recognizing by referring to the exception information storage means, if a recognition error occurs, the user can do everything from the beginning. It is to realize a voice recognition device suitable for high-speed continuous voice input without rewording.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a voice recognition device according to an embodiment of the present invention.

FIG. 2 is a control procedure diagram for explaining the operation of the voice recognition device.

FIG. 3 is a content diagram of recognition error information output from a recognition unit that is a constituent element of the present invention.

FIG. 4 is a control procedure diagram for explaining an exceptional processing operation of the voice recognition device.

FIG. 5 is a configuration diagram of normal information storage means and exception information storage means, which are constituent elements of the present invention.

FIG. 6 is a configuration diagram of a conventional voice recognition device.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Input means 2 Recognition means 3rd process determination means 4 Exception processing recognition means 5 Output means 7 Normal information storage means 8 Exception information storage means

Claims (1)

[Claims] 1. Input means for receiving a voice input, recognition means for recognizing the voice input accepted by the input means, and next processing determining means for determining the next processing based on the result recognized by the recognition means. Exception processing recognition means that is executed when a recognition error occurs during voice recognition, output means that outputs the recognition processing result, normal information storage means that the recognition means refers to, and exception information that the exception processing recognition means refers to A voice recognition device comprising storage means.