JPH02126299A - Speech recognition device - Google Patents

Abstract

PURPOSE:To improve a recognition by registering a reproduced speech for confirmation at the time of registration and recognition by utilizing a speech recording and reproducing device. CONSTITUTION:A speech recognition device 10 fetches the registered speech signal of a user through a BPF 3 and allows the speech recording and reproducing device to start recording operation on detecting the start of a word. Then the recognition device 10 extracts features from the input speech data detects the end of the word while storing it in a dictionary, thereby recording words which do not exceed specific length. Then, a user voices a word to be recognized in front of a microphone 1 and then the recognition device 10 starts fetching the speech signal through the BPF3 to detect the start of the word and detects the start of a signal which does not exceed the specific word length to extract its features. Then, the features are matched with respective word features in the dictionary stored in a speech data RAM 16 to specify an address of the speech data RAM 25 and sends an instruction of reproduction to a device 20, which generates a voice to a speaker 7.

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Field of Application] The present invention relates to a speech recognition device, and is particularly suitable for application to a device in which a specific speaker's Japanese speech is registered in advance.

[Prior Art] It would be convenient for humans to be able to input information into an input target device by voice, and recently various devices have been proposed that include a voice recognition device as an information input device. For example, instead of directly using a button to make a call to close a telephone line, the caller uses voice to specify the destination, and the voice is recognized and the phone number of the other party is entered. A device has been proposed that performs a calling operation by taking out a call from a storage unit.

As a speech recognition device used as an information input device of such a device, speech to be recognized (including speech of one or more words, hereinafter referred to as word speech) is registered in advance,
There is a system for specific speakers that recognizes input word sounds by comparing them with registered single 1181 sounds jh.

In such an 11-voice 32 recognition device, when registering a voice, the user is made to recognize the registered word voice, and when recognizing a voice, the user is made to confirm the recognition result. Confirmation at the time of voice registration and confirmation of recognition results are performed visually using a display device, liquid crystal display, light emitting diode, etc., or by auditory methods such as a synthesized voice or a buzzer. JP-A-63-8799, JP-A-63-14200, JP-A-63-15294, JP-A-62-27
9399, etc.).

FIG. 2 shows an example of a conventional system having a confirmation configuration using synthesized speech. In Figure 2, microphone 1
The system captures the word speech generated by the user during registration or recognition and converts it into an electrical signal. After this captured speech signal is amplified via the amplifier circuit 2,
After unnecessary components are removed through the bandpass filter circuit 3, the signal is provided to the speech recognizer 4.

The speech recognizer 4 includes an analog/digital conversion circuit 41°,
Microprocessor 42, instruction ROM 4
3, a working R, an AM 44, a data ROM 45, a data RAM 46, etc., and executes voice registration processing, voice recognition processing, and confirmation processing thereof. In such processing, when it becomes necessary to output sound jG,
The speech recognizer 4 instructs the synthesized speech reproducer 5 to output via the address bus A and the control line B.

The synthesized speech reproducer 5 extracts a message or a phoneme from the area of the designated address, converts it into an analog signal, and supplies it to the amplifier circuit 6. The amplifier circuit 6 amplifies the incoming composite P signal and sends it to the speaker 7! Toss and generate.

The speech recognizer 4 and the synthesized speech reproducer 5 are connected to the controller 1-roller 8.
controlled by The controller 8 recognizes the sound P via the control bus C so as to execute processing at the time of registration or recognition based on a command from an external device (not shown) via a serial bus D such as a so-called R8232C bus. It also controls the synthesized speech reproducer 5 via the control line E.

Next, the registration operation in such a conventional example will be explained. When the user instructs the registration process using an operation panel or the like of an external device, a registration address and a registration command are given to the controller 8 from the external device. At this time, the controller 8 provides the registered address and the registered command to the voice recognizer 4 as they are, or more precisely, with the data format changed. After instructing the registration process, the user pronounces the registered word into the microphone 1.

When the registered address and registered command are received, the voice recognition unit 1 takes in the combined signal from the bandpass filter circuit 3 and converts it into digital data.1. After that, extract the characteristics, create a dictionary of registered words, and store it in the data RAM 46.
Store in.

When the storage operation is completed, the voice recognizer 4 reports to the external device via the control bus C, the controller 8, and the serious bus D that the registration has been successfully performed.
In addition, data is outputted from the area where the registration completion message of the synthesized voice reproducer 5 is stored via the address bus A and the control line B, and is converted into an analog signal and output, and the registration completion message is output from the speaker 7. let

Next, a speech recognition process using such registered speech information will be explained.

:7 When the controller 8 receives a recognition command from an external device via the serial bus D, it converts the communication format and then sends it to the speech recognizer 4 via the control line C. After instructing the recognition process, the user generates the voice desired to be recognized into the microphone 1.

When a recognition command is given, the speech recognizer 4 takes in the speech signal captured by the microphone 1 via the bandpass filter circuit 3, extracts its features, and stores the extracted speech features and data in the RAM 46. The system calculates the matching with multiple acoustic features of the dictionary, and recognizes the speech as corresponding to the audio feature of the dictionary that matches the most. The speech recognizer 4 sequentially instructs the address of the synthesized voice If generator 5 which stores the phoneme fragments of the word corresponding to the recognition result, and also activates the synthesized voice PL reproduction 5 to start its reproduction. The words resulting from the recognition are generated from the speaker 7.

Also, the recognition results are transmitted to external equipment via the control bus C, controller 8, and serial bus D almost simultaneously.

[Problems to be Solved by the Invention] In the two-step method, confirmation is made using synthesized speech immediately after registration and immediately after recognition;
When you want to check the percentage, there are actually cases such as IS after Sahachi, and when checking at the timing mentioned above,
Further, confirmation by synthetic RJ' may be insufficient.

That is, as the number of registered words increases, the user may not be able to memorize all the registered words, and may want to check the contents of the registered words. In this case, it may be desired to confirm the registration operation and the recognition operation separately. In the past, in order to avoid such inconveniences, many users used memos or the like in the registered word list, which was a tedious process.

In a speech recognition device, it is necessary that the pronunciation during the registration operation, the pronunciation during recognition, and the pronunciation during the recognition operation are identical enough to be accurately recognized by LF, but in the case of words that are used only occasionally. , During the recognition operation, you may forget how you pronounced it during the registration operation, and in this case, you may want to check the pronunciation during the registration operation. I cannot confirm this and cannot respond. Moreover, even if it is possible to check the registered word r1°J@ in the middle of the process, the confirmation voice is a synthesized voice and the pronunciation trends such as intonation and speed are fixed, so the first word in middle school Although the content can be confirmed, the pronunciation tendency cannot be confirmed.

In addition, the difference between the environmental noise at the time of registration and the environmental noise at the time of recognition operation has a great influence on the recognition result.
Confirmation of such an environment cannot be performed using synthetic sound. Therefore, when recognition is not possible, it has not been possible to determine whether or not environmental noise is the cause. There are cases where it is desired to check such environmental noise at times other than the recognition operation and the registration operation.

By having a dictionary for each of multiple speakers and exchanging the dictionaries, it is possible for multiple speakers to commonly use a speech recognizer. In the past, the only way to confirm which speaker the dictionary loaded in the recognizer belonged to was to actually perform the recognition operation.

In response to such a request, it is possible to satisfy the request in terms of timing by providing a confirmation process at a timing other than the recognition operation and registration operation, but (ii) the output voice of recognition is a synthesized a Therefore, the speaker cannot be recognized.

Also, depending on the device, synthesized speech may be difficult to hear, and similar words may be mistakenly recognized.

The present invention has been made in consideration of the following points. First, the pronunciation of the confirmation voice of the registered word is equal to the pronunciation r5 of the voice input at the time of registration, including environmental noise. The second purpose is to provide a voice recognition device that allows the user to freely select the timing at which a registered single death can be confirmed.
Another object of the present invention is to provide a speech recognition device that can increase the number of contents that can be confirmed even when confirmation is performed at an arbitrary timing.

[Step f for solving the problem] In both the first to third aspects of the present invention, predetermined word contents inputted and registered in the dictionary are processed into words at least immediately after middle registration or immediately after word recognition. The present invention relates to a speech recognition device in which a content output means causes a pronunciation means to pronounce a registered word or a recognized word.

The first aspect of the present invention provides a voice recording method in which the word content output means records the word sounds input for word registration as they are, and outputs the recorded word sounds to the pronunciation means based on a playback command for confirmation. It is characterized by being configured with a playback device.

Further, the second invention is characterized in that an arbitrary point confirmation command means is provided for giving a command to output a predetermined word to the word content output means in response to an input operation.

The third aspect of the present invention is to provide an arbitrary time confirmation command means for changing the output command to the word content output means in accordance with an input operation, and to output the word sound inputted for word registration to the word content output means as it is. The present invention is characterized in that it is constituted by a voice recording and reproducing device that records the recorded word voice and also outputs the recorded word voice to a pronunciation means based on a reproduction command for confirmation.

[Operation] The first aspect of the present invention is to record the academic language sounds input for registration by the audio recording and reproducing device as they are at the time of word registration, and to register the academic language sounds from the audio recording and reproducing device immediately after registering the word or immediately after recognizing the word. The voice of the word or the voice of the recognized word is given to the pronunciation means, and the word r1 voice is pronounced for confirmation. Here, the pronounced word sound includes not only the word content but also information such as intonation and environmental noise at the time of registration.

The second aspect of the present invention provides an arbitrary time point confirmation command means, and when the pregnancy time point confirmation command means detects an input operation instructing activation of confirmation pronunciation, the word content output means instructs the pronunciation means. The registered words are output and the sounds of the words are pronounced. Therefore, in use, the registered word -R and voice can be pronounced not only immediately after word registration or immediately after word recognition, but also at any time.

The third invention provides word sounds pronounced at any time,
This is the word sound itself that was registered using a sound recording and reproducing device.

[Example] Hereinafter, an example of the present invention will be described in detail with reference to the drawings.

Here, FIG. 1 is a block diagram showing the configuration of this embodiment,
3 to 5 respectively show the registration process of this embodiment,
It is a flowchart showing the operation procedure of recognition processing and confirmation processing at an arbitrary point in time.

In FIG. 1, in which parts corresponding to those in FIG.
and a speech recognizer 10 via a bandpass filter circuit 3
I can't believe it. This acoustic recognizer 10 has an analog/digital conversion circuit 11, a microprocessor 12, an instruction ROM 13, a working RAM 14, a data ROM 15, and data R, AM 16, etc., and performs sound registration processing, speech recognition processing, and confirmation processing thereof. It is something to be carried out. This speech recognizer 10 includes a speech recording and reproducing device 2
0, address bus A, control line B, word start detection signal line F
and is connected via a word end detection signal line G.

The audio recording and reproducing device 20 includes an analog/digital conversion circuit 21, a microprocessor 22, and an instruction R.
OM23, working R, AM24, audio data RAM
25 and a digital/analog conversion circuit 26.

This audio recording/playback device 20 includes a microprocessor 22
In the control F, the audio signal given from the bandpass filter circuit 3 is converted into digital data (a゛ voice data) by the analog/digital conversion circuit 21 and recorded in the audio data RAM 25. A word start detection signal applied from the word start detection signal line F is used as the recording start command, and a word end detection signal applied from the word end detection signal line G is used as the recording path r command. .

The playback from this voice recording and playback device 20 is performed by the voice recognition 31.
Playback command and playback address from 0, or control 1:1
This is done by the reproduction command and reproduction address via the control line E from the controller 30. The audio recording/reproducing device 20 extracts the audio data from the area of the audio data RAM 25 specified by the reproduction address, and then converts it into an analog signal via the digital/analog conversion circuit 26 and sends it to the amplifier circuit 6 for driving the speaker. It is designed to output.

The controller 30 in this embodiment not only controls data exchange between an external device and the voice recognizer 10, but also commands a playback command to the voice recording and playback device 20 when a confirmation command is given from the external device at any time. It also functions to give a playback address. Note that a clock signal used by the audio recording and reproducing apparatus 20 during recording and reproducing operations is supplied from the controller 30 via a control line E.

Next, the third section regarding the audio signal registration process in the embodiment will be explained.
This will be explained with reference to the figures.

Upon receiving the registration address and registration command from the external device, the controller 30 specifies the recording start address of the audio data RAM 25 to the audio recording/reproducing device 20, and also specifies the dictionary address and registration start to the speech recognizer 10. Send command (step 100.10)
1).

After the user executes the command output operation from the external device, the user generates the voice to be registered into the microphone 1.

When the speech recognizer 10 receives the registration command and dictionary address, it filters the speech signal generated by the user via the bandpass filter circuit 3 (step 110).
．． 111). When such import is started, it is determined whether a voiced part has been reached, that is, whether the beginning of a word has been detected, and the detection is awaited (step 112).
. Eventually, when the beginning of a word is detected, the speech recognizer 1o
Voice recording and playback bag: Word start detection signal line F for 20
A detection signal is sent out via the controller to start a recording operation (step 114).

At this time, the audio recording and reproducing device 20 controls the controller 30
The audio data RAM 25 instructed by the audio data RAM 25 starts storing the audio data that is given via the bandpass filter circuit 3 and converted into digital data by the internal analog/digital conversion circuit 21.

The speech recognizer 10 performs feature extraction on the input speech data, stores it in a dictionary (data RAM 16 ), detects whether the speech data is in a silent state, that is, at the end of a word, and identifies the word. until the end is detected.
Repeat import and feature extraction operations (step), 14.1
15). When the end of a word is detected in this way, it is determined whether the word length is shorter than a predetermined length or not, and if it is longer, error processing is executed (steps 116 and 117).

On the other hand, if the word length is shorter than the predetermined length,
The R voice recognizer 10 via the word end detection signal line G,
Audio recording reproduction bag? fY 20 is instructed to end the recording operation (step 118).

Thereby, the audio recording/reproducing device 20 ends the recording operation of the audio signal from the bandpass filter circuit 3.

After instructing the audio recording/reproducing device 20 to end the recording operation described above, the audio recognizer 10 sends a registration completion signal to the controller.
Output to the roller 30 and finish the registration operation (Step 1
19).

After instructing the voice recognizer 10 to perform a registration operation, the controller 30 waits to receive a registration completion signal from the voice recognizer 10, and upon receiving the registration completion signal, the controller 30 instructs the voice recording/reproducing device 20 to perform the registration operation. The area that has just been recorded is designated by the playback address, a playback command is given to start the playback, and then the completion of registration is reported to the two external devices (steps 102 to 104).

The audio recording/reproducing device 20 outputs the just-registered audio/voice data according to the command, and causes the speaker 7 to output the sound/voice data.

Next, the recognition process will be explained with reference to FIG. 4.

When the user inputs voice and causes the acoustic recognizer 10 to recognize it, the user instructs the recognition process using an external device, and then
Pronounce the word to be recognized into the microphone 1.

At this time, a recognition command is output from the external device to the =7nl-roller 30, and the controller 30 converts the communication format of the recognition command to the voice recognizer 1.
Output to 0.

When the speech recognizer 10 receives this command, it starts capturing the position signal captured by the microphone 1 via the bandpass filter circuit 3, and waits to detect the start of a word (steps 130 to 132). When the beginning of a word is detected in this way, the input audio signal is stored in the working RAM 1.6 and feature extraction processing is performed until the end of the word is detected. It is determined whether the word length is shorter than a predetermined length (steps 133 to 135). As a result of this determination, if it is determined that the length is long, error handling is executed (step 136).

In this way, when a word with a predetermined long side is input and its features are extracted, it is matched with the features of each word in the dictionary stored in the audio data RAM L6 and recognized as one of the words. 137.138, which specifies the address of the voice data R and AM25 of the voice recording and reproducing device 20 related to the recognized word, and instructs the voice recording and reproducing device 20 to reproduce it via the control line A. ).

As a result, the audio recording/reproducing device 20 performs a reproducing operation, and supplies the instructed #j self signal to the speaker 7 via the amplifier circuit 6 to produce the voice of the recognized word.

The sound 71 recognizer 1.0 instructs the voice recording and reproducing device 20 to reproduce the recognized voice, then reports the recognition result to the external device via the controller 30, and ends the series of recognition processes (step 139). ).

In this way, the user can confirm the recognition result using the same pronunciation that was pronounced this evening.

Next, a confirmation process that allows the user to confirm the registered voice content at any time, regardless of the registration or recognition operation, will be explained with reference to FIG. 5.

In this case, the user instructs the external device to perform confirmation processing at an arbitrary time, and also inputs, for example, the word number of the word to be confirmed. At this time, the external device sends the 1st Moment Point Confirmation command and the voice recording/playback device M corresponding to the word number.
The address of the area of the audio data RAM 25 in 20 is output to the controller 30.

When the controller 30 receives the arbitrary point confirmation command, it outputs the address of the area to be reproduced to the audio recording and reproducing device 20, and then outputs a reproduction command and ends the series of processing (steps 150 to 152). .

As a result, the audio recording and reproducing device 20 outputs audio data from the designated area of the audio data RAM 25, converts it into an analog signal via the digital/analog conversion circuit 26, sends it to the amplifier circuit 6, and outputs the audio data from the speaker 7. Pronounce a confirmation voice.

This allows the user to confirm the content of the specified word using the same pronunciation as when the word was registered.

Therefore, according to the above-described embodiment, it is possible to check the registered content at times other than the time of registration and recognition.

In this case, an audio signal is recorded at the time of registration, and the audio signal is emitted when checking at any point in time, so unlike confirmation using synthesized speech, it is possible to not only confirm the registered content itself, but also to confirm who the registered speaker is. In addition, it is possible to check the background quality at the time of registration, such as environmental noise, intonation of pronunciation, pronunciation speed, etc.

As a result, when I recognized it, I got an A that was not exactly what I intended! After trying to pronounce the i, the j2 recognition rate is rFq-L. I can do two things.

Moreover, at the time of registration and recognition, the registered pronunciation itself is reproduced for confirmation, rather than confirmation using synthesized speech, so that the operator can be sure of confirmation. In other words, confirmation accuracy is improved.

In the above embodiment, the audio recording and reproducing apparatus 20 is shown using a semiconductor memory as a recording medium, but it is also possible to use other recording media with fast search operations, such as a floppy disk or an optical disk. It's okay.

In addition, in the description above, an input device for a telephone was shown as an example of using the voice recognition device according to the present invention.
It may also be used for other purposes.

Furthermore, in the above description, a clock signal with the same cycle is given to the audio recording and reproducing device 20 during recording and reproducing, so that the same pronunciation as the registered audio is outputted. By making the clock cycle different from that at the time of recording, the pronunciation of the reproduced sound may be arbitrarily varied.

[Effects of the Invention] Hereinafter, as in Ix, according to the first aspect of the present invention, the reproduced voice for confirmation at the time of registration and recognition is made into a registered voice using a voice recording and reproducing device, thereby generating a synthesized voice. Unlike the confirmation by It is possible to check whether there is a problem, and the confirmation accuracy can be improved.

This can be reflected at the time of recognition, and the recognition rate can be improved.

Furthermore, according to the second aspect of the present invention, since the registered words are pronounced by the playback command at times other than the time of registration and recognition, the operator can input the registered contents as necessary other than the time of registration and recognition. It can be confirmed. Therefore, there is no need for the operator to write down the registered contents on memo paper or the like.

Furthermore, according to the third aspect of the present invention, even when checking at any point in time, the registered pronunciation itself is played back for confirmation, instead of checking using a synthesized voice. It is possible to realize a device that is easier to use and has the following effects.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the speech recognition device according to the present invention, FIG. 2 is a block diagram showing a conventional device, and FIGS. 3 to 5 show registration processing, recognition processing, and optional processing of the above embodiment, respectively. This is a first coach yard showing the operating procedure of time confirmation processing. 1...Microphone, 7...Speaker, 10...
・Speech recognizer, 20...Speech recording/playback device, 30...
・Con1~Lola. Fig. 7 Helping diagram showing the configuration of the conventional device Fig. 2

Claims (3)

[Claims] (1) At least immediately after word registration or word recognition, the word content output means gives a predetermined word content that has been input and registered in the dictionary to the pronunciation means to make it pronounce, thereby confirming the registered word or recognized word. In the speech recognition device, the word content output means records the word sound input for word registration as it is, and outputs the recorded word sound to the pronunciation means based on a playback command for confirmation. A speech recognition device comprising a recording/playback device. (2) At least immediately after word registration or word recognition, the word content output means gives a predetermined word content that has been input and registered in the dictionary to the pronunciation means to make it pronounce, thereby confirming the registered word or recognized word. What is claimed is: 1. A speech recognition device comprising: arbitrary time confirmation command means for giving a command to output a predetermined word to the word content output means in response to an input operation. (3) The word content output means records the word sounds input for word registration as they are, and the sound production means outputs the designated word sounds from among the recorded word sounds based on a confirmation output command. 3. The speech recognition device according to claim 2, wherein the speech recognition device comprises a speech recording and reproducing device that outputs a speech signal to a user.