JPH06337695A - Voice recognition device - Google Patents

Abstract

PURPOSE:To properly switch a recognition objective vocabulary without being affected by the deviation in time due to delay, etc., at the time of transmitting a recognition objective vocabulary switching request. CONSTITUTION:This device is constituted so that at the time of inputting the recognition objective vocabulary switching command from a command input part 5, the time when the command is inputted is read from a timer part 3 and stored in a command time storage part 6, and at the time of inputting a sound from a sound input part 1, the starting point and the ending point of the sound are detected by a sound detection part 2, and the time is read from the timer part 3 every time to be stored in a sound time storage part 4, and at the time of storing a sound input ending time, a sound input starting time stored in the sound time storage part 4 is compared with the storage contents of the command time storage part 6 by a recognition objective vocabulary decision part 7, and whether the dealing input sound is voice-recognized or not for the vocabulary after the recognition objective vocabulary is switched is decided by whether the sound input starting time is later or not, and the recognition objective vocabulary switching command is issued for a voice recognition part 8.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device for switching a recognition target vocabulary as needed and recognizing an input voice targeting a minimum vocabulary.

[0002]

2. Description of the Related Art In recent years, a voice recognition device has been remarkably popularized and has been used in many fields, but the performance sufficient for practical use has not yet been obtained. Therefore, as one means for improving recognition performance, a means for recognizing a minimum vocabulary by switching the recognition target vocabulary as necessary has been proposed.

[0003]

However, in the conventional means for switching the recognition target vocabulary, the recognition target vocabulary switching request (recognition target vocabulary switching command) inputted from the recognition target vocabulary switching request input means such as a switch is concerned. Since the recognition target vocabulary of the voice recognition means that controls the recognition processing in the voice recognition device is switched at the time point when the recognition target vocabulary is transmitted to the switching means (recognition target vocabulary determination means), the following problems occur.

(1) The time from the input of the recognition target vocabulary switching request from the recognition target vocabulary switching request input means to the actual transmission of the request to the recognition target vocabulary switching means (recognition target vocabulary determination means). If there is a delay, even though the recognition target vocabulary of the voice recognition device is not changed, the user utters the vocabulary after the recognition target vocabulary is changed, resulting in erroneous recognition.

(2) When a recognition target vocabulary switching request is input while voices are being input, is the currently input voice recognized by the recognition target vocabulary before the input of the switching request? The recognition target vocabulary after switching the recognition target vocabulary according to the switching request cannot be discriminated whether or not it is recognized, and the recognition in the recognition target vocabulary intended by the user is not performed.

Therefore, according to the present invention, the vocabulary uttered by the user is judged to determine which vocabulary is to be recognized, and the recognition processing is performed. It is an object of the present invention to provide a voice recognition device that can appropriately switch the vocabulary to be recognized without being affected by the shift of the voice recognition and can perform voice recognition with high accuracy.

[0007]

A voice recognition device of the present invention comprises a voice recognition means for recognizing an input voice, and a first time storage means for storing the input time of the voice.
A recognition target vocabulary switching request input means for inputting a recognition target vocabulary switching request applied by the voice recognition means; a second time storage means for storing the input time of the recognition target vocabulary switching request; The first and second time storage means are compared with each other, and the recognition target vocabulary determination means for determining switching of the recognition target vocabulary applied by the voice recognition means is provided according to the comparison result. It is characterized in that the speech recognition means performs recognition processing using the recognition target vocabulary according to the judgment result of the judgment means.

[0008]

In the above structure, the time when the voice is input is stored in the first time storage means, and the time when the recognition target vocabulary switching request is input is stored in the second time storage means. The contents stored in both time storage means are, for example, the first
Immediately after the voice input time is stored in the storage means, the recognition target vocabulary determination means makes a comparison. By this comparison process,
It is checked whether or not a recognition target vocabulary switching request has already been input at the voice input time, and whether or not which vocabulary the input voice is for recognition, that is, switching of the requested recognition target vocabulary It is determined whether to perform. The voice recognition means performs the voice recognition processing according to the determination result of the recognition target vocabulary determination means, with the current recognition target vocabulary as it is or switching to the requested recognition target vocabulary.

As described above, the contents of the first time storage means for storing the voice input time and the second time storage means for storing the recognition target vocabulary switching request input time are compared, and the voice input time is compared. It is possible to determine which vocabulary the input voice is the recognition target by checking whether or not the recognition target vocabulary switching request has already been input to, and therefore, according to this determination result. By performing speech recognition using the recognition target vocabulary, it is possible to switch the recognition target vocabulary appropriately without being affected by the time lag due to the delay in transmitting the recognition target vocabulary switching request, etc. Can be realized.

[0010]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram schematically showing the configuration of a voice recognition device according to an embodiment of the present invention.

The voice recognition apparatus shown in FIG. 1 has a voice input section 1 for amplifying a voice signal input from a microphone or the like.
A voice detection unit 2 for detecting an input voice from the voice signal amplified by the voice input unit 1, a timer unit 3 having a clock function, and a timer unit 3 for setting the time of the input voice detected by the voice detection unit 2. And a voice time storage unit 4 for reading and storing.

The voice recognition apparatus shown in FIG. 1 also includes a command input section 5 for inputting a command (recognition target vocabulary switching command) requesting switching of the recognition target vocabulary,
A command time storage unit 6 for reading and storing the time at which the recognition target vocabulary switching command is input by the command input unit 5, the voice time storage unit 4, and each time stored in the command time storage unit 6. And the speech recognition unit 8 for comparing the input speech with the speech recognition target vocabulary switching command before and after the recognition target vocabulary switching command is input.
have. The voice recognition unit 8 performs voice recognition on the voice signal amplified by the voice input unit 1 within the range of the recognition target vocabulary determined by the determination result of the recognition target vocabulary determination unit 7.

Next, the operation of the speech recognition apparatus having the above configuration will be described. First, the voice input unit 1 amplifies a voice signal input from a microphone or the like to an appropriate signal level used by the voice detection unit 2 and the voice recognition unit 8.

The voice detection unit 2 receives the voice signal amplified by the voice input unit 1 and detects the voice uttered by the user (input voice). Various means have been conventionally proposed as means for realizing this type of voice detection. In the present invention, the method is not limited, but the power of the input audio signal is obtained,
Since a method is generally used in which a time point that exceeds a preset threshold value is used as a voice input start point, and a time point that is less than the threshold value is used as a voice input end point, the voice is detected here.
It is assumed that the voice detection unit 2 is realized by the method. The voice detection unit 2 corresponds to the vocabulary having the longest voice duration among the vocabularies that can be recognized by this device, when the duration of the input voice (the period when the power of the voice signal continuously exceeds the threshold value). Even when a predetermined time is exceeded, the time point is detected as a voice input end point.

The result of voice detection by the voice detector 2 is sequentially sent to the voice time memory 4. The voice time storage unit 4 sets a time (voice input start time) Ts when the voice input start point is detected by the voice detection unit 2 and a time (voice input end time) Te when the voice input end point is detected as a timer each time. It is read from the unit 3 and the respective times Ts and Te are internally stored. The method of storing the time of the input voice is not limited to the voice input start time and the voice input end time described above,
For example, it is possible to store the time at the center of the input voice by variously modifying it according to the application.

Next, a recognition target vocabulary switching command for requesting switching of (a set of) recognition target vocabularies applied by the speech recognition apparatus (inside the speech recognition unit 8) of FIG. 1 is input from the command input unit 5. . As a means for realizing the command input unit 5, in addition to command input by a user operation from a switch, a keyboard, etc., command input according to the flow of processing according to the flowchart and the like can be applied.
It is assumed that the recognition target vocabulary switching command applied in the present embodiment merely instructs switching of (a set of) the recognition target vocabulary, and does not instruct up to (the set of) recognition target vocabulary after the switching.

When the recognition target vocabulary switching command is input from the command input unit 5, the command time storage unit 6
The time (command input time) Tc at the time of the input is read from the timer unit 3 and internally stored.

The recognition target vocabulary determination unit 7 has a command input time Tc stored in the command time storage unit 6 and a voice input start time Ts stored in the voice time storage unit 4.
Based on the voice input end time Te, it is determined whether or not to switch the recognition target vocabulary applied by the voice recognition unit 8. Details of this determination will be described later. The recognition target word vocabulary determination unit 7 has a current recognition target word vocabulary identifier holding unit 71 that holds an identifier of (a set of) the current recognition target vocabulary.

The voice recognition unit 8 performs a recognition process for an input voice based on the voice signal amplified by the voice input unit 1 within the range of (a set of) the recognition target vocabulary according to the determination result of the recognition target vocabulary determination unit 7. To do. Various means have been conventionally proposed as means for realizing the voice recognition unit 8, but any method may be used in the present invention. Here, as an example, regarding the case of the voice recognition unit 8 to which an unspecified speaker voice recognition method in which the user does not need to register voice in advance,
This will be described with reference to FIGS. 2 and 3.

First, the voice signal amplified by the voice input unit 1 is input to the acoustic analysis unit 81 in the voice recognition unit 8 shown in FIG. The acoustic analysis unit 81 is for acoustically analyzing an input voice signal to obtain a characteristic parameter. Here, the input voice signal is LPC (Linear Predictive Co).
ding) Mel cepstrum analysis. The acoustic analysis unit 8
The acoustic analysis in 1 is not limited to the LPC mel cepstrum analysis, but may be a BPF (Band Pass Filter) analysis or the like.

The acoustic analysis unit 81 has an A / D (analog / digital) converter 81, as shown in detail in FIG.
1, a power calculation unit 812 and an LPC analysis unit 813.

The audio signal input to the acoustic analysis unit 81 is
In the A / D converter 811, for example, the sampling frequency 1
After being quantized with 12 bits at 2 kHz, it is input to the power calculation unit 812, the voice power thereof is calculated, and further input to the LPC analysis unit 813, and LPC mel cepstrum analysis (LPC analysis) is performed. This LPC analysis is performed using a 16th-order LPC mel cepstrum as an analysis parameter with a frame length of 16 msec and a frame period of 8 msec, for example.

Now, in the speech recognition unit 8 shown in FIG. 2, in addition to the acoustic analysis unit 81 described above, a speech quantization unit for obtaining a label for each frame based on the characteristic parameters obtained by the acoustic analysis unit 81. 82, a PS dictionary storage unit 8 in which a recognition dictionary in units of a predetermined PS (speech segment) is stored
3, the label sequence obtained by the voice quantizer 82 is set to H
HMM recognized using MM (Hidden Markov Model)
A recognition unit 84, an HMM parameter storage unit 85 in which the parameters of the HMM model M are stored, and a recognition result output unit 86 for outputting the recognition result are provided.

The voice quantizer 82 continuously performs matching processing of the characteristic parameters analyzed by the acoustic analyzer 81 with the recognition dictionary of a predetermined PS unit registered in the PS dictionary memory 83 in the time axis direction, The PS having the maximum similarity for each frame is output to the HMM recognition unit 84 as the quantization result. Here, the continuous matching process by PS in the voice quantizing unit 82 is performed using the composite LPC mel cepstrum similarity measure shown in the following expression (1).

[0025]

[Equation 1] In the equation (1), C is the LPC mel cepstrum, W _m ^(Ki) and φ _m ^(ki) are weights and eigenvectors obtained from the eigenvalues of the PS name Ki, respectively. Also,(
・) Indicates dot product, and ‖ ‖ indicates norm.

The PS used in this embodiment is, for example, as follows. (1) Persistence segment: (1-1) Vowel stationary part (1-2) Friction consonant part (2) Consonant segment: Part including transition (transition part) to vowel [semi-syllable] (3) Syllable boundary segment : (3-1) Vowel boundary (3-2) Vowel, consonant boundary (3-3) Vowel, silence boundary (4) Other segment: unvoiced vowel, etc. Among these, (1), (2) and (4) ) Is often used when the recognition segment is a syllable. However, the advantage of PS in this embodiment is that the syllable boundary segment of (3) above is adopted in addition to the segments of (1), (2), and (4) above.

The HMM recognizing unit 84 has a PS that maximizes the similarity for each frame output from the voice quantizing unit 82.
That is, the PS series (label series) is received and word matching of the corresponding input voice is performed. The word matching of the HMM recognition unit 84 will be described below.

In the word matching in this embodiment, the PS series is obtained as a label series as described above, and this is passed through the HMM for each word (category). Where H
A general formulation of MM will be described. It is assumed that the HMM has N states S ₁ , S ₂ , ..., _SN , and the initial state is stochastically distributed to these N states. In voice,
A model that transitions a state with a certain probability (transition probability) is used for each fixed frame period. At the time of transition, a label is output with a certain probability (output probability), but a null transition that transitions the state without outputting the label may be introduced. Even if the output label sequence is given, the state transition sequence is not uniquely determined. Since only the label series can be observed, it is called a hidden markov model (HMM). The HMM model M is defined by the following six parameters.

N: number of states (states S ₁ , S ₂ , ..., S _N ) K: number of labels (labels R = 1, 2, ..., K) p _ij : transition probability Probability of transition from S _i to S _j q _ij (k): Probability of outputting label k at the transition from S _i to S _j m _i : Probability of initial state S Probability of initial state S _i F: Set of final states Next, for model M Add restrictions on transitions that reflect the characteristics of voice. In the case of speech, generally, a transition of a loop that returns from a state S _i to a previously passed state (S _i-1 , S _i-2 , ...) Is not allowed because it disturbs the temporal context.

The HMM is evaluated by the probability Pr that the model M outputs the first-ranked label sequence O ₁ = o ₁₁ , o ₂₁ , ..., O _T1.
(O / M). At the time of recognition, the HMM recognition unit 84 searches the HMM parameter storage unit 85 for a model M that maximizes Pr (O / M) assuming each model. Then, the recognition result output unit 86 outputs the category corresponding to the model having the maximum probability Pr (O / M) as the recognition result for the input voice.

Here, in the HMM parameter storage unit 85, the model M (parameters) of the three words of the alphabets "a", b ", and" c "forming the recognition target vocabulary set A and the recognition target vocabulary set. The model M (parameters) of the three words “1”, 2 ”, and“ 3 ”forming B is stored. At the time of recognition by the HMM recognition unit 84, an operation of searching for a model M that maximizes Pr (O / M) is performed for either the recognition target vocabulary set A or the recognition target vocabulary set B. Which recognition target vocabulary set is targeted is specified by the recognition target vocabulary determination unit 7.

For learning the HMM, the label series of the learning data may be given to the HMM in advance, and the parameter of the model M that maximizes Pr (O / M) may be estimated there. Then, the estimated parameter is stored in the HMM parameter storage unit 85.
Register in.

Next, regarding the recognition target vocabulary switching operation directly related to the present invention, the recognition target vocabulary set A will be described.
(“A”, “b”, “c”) and recognition target vocabulary set B
Taking "(1", "2", "3") as an example,
This will be specifically described.

First, at present, an identifier indicating a recognition target vocabulary set A is set in the current recognition target vocabulary identifier holding unit 71 of the recognition target vocabulary determination unit 7, and the recognition target vocabulary in the voice recognition unit 8 is set in the set A. It has been done. In this state, a recognition target vocabulary switching command is input from the command input unit 5 to switch and set the current recognition target vocabulary set A to the recognition target vocabulary set B.

Here, assuming that the input voice is S1, the time relationship between the input voice S1 and the recognition target vocabulary switching command is shown in FIG. In FIG. 4, Ts is the voice input start time of the input voice S1, and Te is the voice input end time of the same input voice S1, which is stored in the voice time storage unit 4 as described above. Next, Tc is a command (recognition target vocabulary switching command) input time, which is stored in the command time storage unit 6 as described above. Also, Td is the command input section 5
Dt is a time delay at which the recognition target vocabulary switching command input from the above is actually transmitted to the recognition target vocabulary determining unit 7 assuming that the conventional method is applied. Here, at each of the above times, Tc <Ts <Td <Te
It is assumed that

Thus, in the example of FIG. 4, the input of the voice S1 is started from the time Ts after the recognition target vocabulary switching command is input at the time Tc.
Shows the time relationship when the command is transmitted to the recognition target vocabulary determination unit 7 at time Td before the input of the voice S1 is completed.

In the present embodiment, the recognition target vocabulary determination unit 7
Is stored in the command time storage unit 6 when the voice input end time Te is stored in the voice time storage unit 4 with the voice input start time Ts already stored in the storage unit 4 as the voice input time. Time (command input time Tc)
In comparison with whether the voice input start time Ts is later, whether the corresponding input voice should be voice-recognized for the vocabulary after switching the recognition target vocabulary, that is, the recognition target vocabulary is determined. Determine whether to switch.

When it is determined that the recognition target vocabulary is switched, the recognition target vocabulary determining unit 7 sets the contents of the current recognition target vocabulary identifier holding unit 71 to an identifier indicating the recognition target vocabulary set currently applied to the speech recognition. Is changed to an identifier indicating the other recognition target vocabulary set, and the fact is notified to the voice recognition unit 8 (the HMM recognition unit 84 therein).

In the example of FIG. 4, when the voice input start time Ts of the input voice S1 is stored in the voice time storage unit 4,
The recognition target vocabulary switching command has already been input and the command input time Tc has been stored in the command time storage unit 6. In this case, since the voice input start time Ts is after the command input time Tc, the recognition target vocabulary determination unit 7 determines the recognition target vocabulary switching, and changes the contents of the current recognition target vocabulary identifier holding unit 71 to the current one. The identifier indicating the recognition target vocabulary set A is changed to the other identifier indicating the recognition target vocabulary set B, and the voice recognition unit 8 (the HMM recognition unit 84 therein) is notified of the switching to the recognition target vocabulary set B. . Then, the recognition target vocabulary determination unit 7 sets the contents of the command time storage unit 6 to the invalid state when the switching notification is made. If the voice input start time Ts is before the command input time Tc, or if the content of the command time storage unit 6 is invalid, the voice recognition unit 8 (H in
There is no notification of switching to the MM recognition unit 84).

The HMM recognizing unit 84 in the voice recognizing unit 8 starts from the input end time of the input voice S1 (when the voice input continues even after the lapse of a predetermined time, when the predetermined time is exceeded). The recognition process of the voice S1 is started. Here, when the above switching notification is received from the recognition target vocabulary determination unit 7, the recognition process of the input voice S1, that is, Pr (O
/ M) is maximized, the process of searching for the model M is shown in FIG.
As shown in, the recognition target vocabulary set B is used instead of the current recognition target vocabulary set A. As a result, when the recognition target vocabulary switching command is input in order to switch the current recognition target vocabulary set A to the recognition target vocabulary set B, for example, when the number "1" is uttered, the input voice of this number "1" is input. (S1) is a recognition target vocabulary set B
Since the recognition processing is performed for ("1", "2", "3"), it is correctly recognized. This recognition result is the voice recognition unit 8
Is output from the recognition result output unit 86.

On the other hand, in the conventional method of switching the recognition target vocabulary at the time Td when the recognition target vocabulary switching command is transmitted to the recognition target vocabulary determination section 7, the input voice S1 is shown in FIG. As a recognition target vocabulary set A
Recognition processing is performed for ("a", "b", "c").
Therefore, even if the recognition target vocabulary switching command is input to switch the current recognition target vocabulary set A to the recognition target vocabulary set B as in the above example, and then the number "1" is uttered, this number The input speech (S1) of “1” is the recognition target vocabulary set A (“a”, “b”,
In the conventional method in which “c”) is recognized, only the alphabet “a”, “b”, or “c” can be obtained as the recognition result, and the recognition result cannot be recognized correctly.

Although the present invention has been described with reference to the embodiments, the present invention is not limited to the above embodiments. That is, the gist of the present invention is to store the input time of the voice and the input time of the recognition target vocabulary switching command (recognition target vocabulary switching request) in order to appropriately switch the recognition target vocabulary and obtain high recognition performance. Then, it is to determine which vocabulary the voice input from these times should be recognized, and perform voice recognition. Therefore, the voice recognition method, the voice detection method, the recognition target vocabulary switching command input method, the recognition target vocabulary determination rule, and the like are not limited to those shown in the above embodiment.

For example, in the above-described embodiment, the recognition target vocabulary determination unit 7 sends a recognition target vocabulary switching notification to the voice recognition unit 8 to the HMM recognition unit 84 in the recognition unit 8 to recognize the notified recognition target vocabulary. Although it has been described that the recognition processing is performed within the range, the recognition result output unit 8 outputs the switching notification.
6, the HMM recognition unit 84 performs recognition processing on all of the recognition target vocabulary set A and the recognition target vocabulary set B, and the recognition result output unit 86 performs the switching specified by the switching notification. It is also possible to output the recognition result within the range of the later recognition target vocabulary set. Therefore, “performing recognition processing using the recognition target vocabulary” described in the claims also includes output processing of this recognition result.

In the above embodiment, the case where the voice input start time Ts is used as the voice input time and compared with the contents of the command time storage unit 6 (command input time Tc) has been described. Instead, the time of the center of the voice input start time Ts and the voice input end time Te (T
s + Tc) / 2, or the voice input end time Te may be used as the voice input time.

Further, in the above embodiment, when the voice input end time Te is stored in the voice time storage unit 4, the voice input start time Ts already stored in the voice time storage unit 4 is obtained.
It has been described that the switching of the recognition target vocabulary is determined by comparing with the content of the command time storage unit 6 (command input time Tc) as the voice input time, but as in this example, the voice input time is used as the voice input time. Start time T
When s is applied, switching of the recognition target vocabulary may be determined when the voice input start time Ts is stored in the voice time storage unit 4.

In the above embodiment, the command input section 5
Although the recognition target vocabulary switching command from (1) is used to instruct only the switching of (the set of) the recognition target vocabulary, the recognition target vocabulary (set) after switching may also be designated. Absent. In this case,
Instead of switching the recognition target vocabulary in a fixed order, it is sufficient to switch to the designated recognition target vocabulary. In addition, the present invention can be variously modified and implemented without departing from the scope of the invention.

[0047]

As described above, according to the present invention, the input time of the voice and the input time of the request for switching the recognition target vocabulary are stored, and the input voice is changed from the stored time. Since it is configured to determine which vocabulary was uttered as the recognition target vocabulary, the recognition target vocabulary is appropriate without being affected by the time lag due to the delay time etc. when transmitting the request for switching the recognition target vocabulary. It is possible to achieve a great effect in practical use, such as switching to, and realizing high-performance voice recognition.

[Brief description of drawings]

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

2 is a block diagram showing a configuration of a voice recognition unit 8 in FIG.

FIG. 3 is a diagram showing details of an acoustic analysis unit 81 in FIG.

FIG. 4 is a diagram showing an example of a front-back relation of time used for recognition target vocabulary determination in the embodiment.

[Explanation of symbols]

1 ... Voice input section, 2 ... Voice detection section, 3 ... Timer section, 4 ...
Voice time storage unit (first time storage unit), 5 ... Command input unit (recognition target vocabulary switching request input unit), 6 ... Command time storage unit (second time storage unit), 7 ... Recognition target vocabulary determination unit , 8 ... voice recognition unit, 71 ... current recognition target vocabulary identifier holding unit, A, B ... recognition target vocabulary set.

Claims (1)

[Claims] 1. A voice recognition means for recognizing an input voice, a first time storage means for storing an input time of the voice, and a request for switching a recognition target vocabulary applied by the voice recognition means. And a second time storage means for storing the input time of the recognition target vocabulary switching request, and a comparison of both stored contents of the first and second time storage means. And a recognition target vocabulary determination unit that determines switching of the recognition target vocabulary applied by the voice recognition unit according to the comparison result, wherein the voice recognition unit recognizes according to the determination result of the recognition target vocabulary determination unit. A speech recognition device characterized by performing recognition processing using a target vocabulary.