JP2000305589A - Adaptive type voice recognition device, voice processing device and pet toy - Google Patents

Abstract

PROBLEM TO BE SOLVED: To restrain generation of a bad influence on a renewed voice model caused by meaningless or wrong intonation or pronounciation by controlling the renewing level of voice model by the use of renewal parameter in renewing the voice model based on the characteristic quantity of input voice. SOLUTION: In renewing the voice model based on the characteristic quantity of input voice, the degree of renewing the voice model is controlled by a renewal parameter. In a voice recognition device A1, the similarity degrees of the characteristic quantity and voice model are computed by an evaluation value computing part 3 when the characteristics quanty is extracted by a characteristic extraction part 1. Also, in the device A1, a renewal parameter determination part 6 determines the renewal parameter for the input voice according to the conditions considered to be favorable in learning so as to avoid unfaborable learning. The part 6 determines the renewal parameter based on the similarity degrees, for example.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an adaptive speech recognition device, a speech processing device, and a pet toy.
An adaptive speech recognizer that updates the speech model used for speech recognition according to the input speech in order to adapt to a specific speaker,
The present invention relates to a voice processing device and a pet toy.

[0002]

2. Description of the Related Art In recent years, the importance of speech recognition technology has been increasing with the miniaturization and multifunctionality of devices and the sophistication of processing contents. There are various methods for the above-mentioned speech recognition technology, such as dynamic programming matching (DP matching) and hidden Markov model (H
MM) and the like, and an unspecified speaker system that does not specify a speaker has been actively researched. This is because in the speaker-independent system, since speech recognition is performed using a speech model created in advance based on a large amount of linguistic information, there are advantages such as a reduced burden on the user. However, in an unspecified speaker system, the recognition rate may be considerably reduced depending on the speaker. For this reason, depending on the purpose of use, even in an unspecified speaker system, it is necessary to improve the recognition rate for the specific speaker by performing adaptation to the specific speaker. The speaker adaptation in the unspecified speaker system is performed by updating the previously created speech model according to the utterance of the speaker. A technique relating to speaker adaptation of such an unspecified speaker system is disclosed in, for example, Japanese Patent Laid-Open No. 7-230295 (hereinafter referred to as Reference 1).
And Japanese Patent Application Laid-Open No. 9-81183 (hereinafter referred to as Reference Publication 2). The speech recognition technology described in the above-mentioned Reference Publication 1 is a technology relating to speaker adaptation using a mixed continuous distribution type HMM. In the speech recognition technology described in Reference 1, the word HMM of each recognition candidate word is held, and when an input pattern corresponding to an input speech is created, recognition of the input pattern is performed using the word HMM. Is performed, and the recognition result is output. Then, regarding the adaptation, the notation of the recognition result word is referred to, and an adaptation initial word HMM is prepared. This adapted initial word HM
A likelihood calculation based on M and the input pattern is performed for one or a plurality of input patterns, and an average vector after the adaptation is obtained based on the calculation. An adapted HMM is generated from the adapted average vector, and the original HM
By being replaced with M, speaker adaptation is performed. The technique related to speech recognition described in Reference Publication 2 uses a continuous distribution type HMM defined by three parameters of a state transition probability, an average vector, and a variance. , And an HMM approximating the input learning speech is selected from the registered dictionary as an initial model, and the average vector in the selected HMM is calculated as the average vector for the calculated learning speech. Thus, the speaker adaptation is performed only by the user uttering the learning voice a very small number of times.

[0003]

In the prior arts described in the above-mentioned Reference Publications 1 and 2, the input speech is not taken into account, and utterances are erroneously made. Even when irrelevant speech is mixed, speaker adaptation is performed accordingly. For this reason, there is a possibility that the speaker model may be deteriorated even though the speaker adaptation is intended. In recent years, pet toys that simulate pets such as dogs and cats that respond to the speech-recognized linguistic information, such as making a bark, have been attracting attention. It is preferable that the degree of adaptation can be adjusted in order to mimic the phenomenon in which the pet gradually adapts to the owner. The present invention has improved an adaptive speech recognition device, a speech processing device, and a pet toy in order to solve such problems related to the conventional technology.
The degree of updating of the voice model is changed according to the input voice,
For example, when the similarity between the feature amount of the input speech and the speech model is low, by setting the update degree of the speech model to be low, the adaptive speech recognition device and the speech processing device can perform appropriate speaker adaptation. , And pet toys.

[0004]

According to a first aspect of the present invention, there is provided a feature extracting means for extracting a feature of an input speech, and a speech model corresponding to a recognition target word. Voice model storage means, similarity calculation means for calculating the similarity between the feature quantity extracted by the feature quantity extraction means and the voice model stored in the voice model storage means, and similarity calculation means Based on the similarity calculated by (1), a recognition target word selection determining unit that performs selection determination of a recognition target word corresponding to the input voice, and based on the feature amount extracted by the feature amount extracting unit, Voice model updating means for updating the voice model stored in the voice model storage means, and update parameters for controlling the degree of updating of the voice model by the voice model updating means are determined. Is configured as an adaptive speech recognition apparatus comprising comprises an update parameter determining means that.
According to the adaptive speech recognition device of the first aspect, when updating the speech model based on the feature amount of the input speech, the update degree of the speech model is controlled by the update parameter, so that it is meaningless. It is possible to suppress an adverse effect on the updated speech model due to an inappropriate or incorrect utterance.

According to a second aspect of the present invention, in the adaptive speech recognition apparatus according to the first aspect, the updating parameter determining means determines the update parameter based on the similarity calculated by the similarity calculating means. The gist is that update parameters are determined. According to the adaptive speech recognition device of the second aspect, since the update parameter is determined based on the similarity, the update parameter is set such that, for example, the update degree increases as the similarity increases. Is determined, the adverse effect on the updated speech model can be suppressed.
According to a third aspect of the present invention, in the adaptive speech recognition apparatus according to the first or second aspect, there is provided an utterance request determining means for determining and outputting a message prompting utterance of a specific recognition target word, When the input parameter is detected within a predetermined time after the message determined by the utterance request determination module is output, the update parameter determining means converts the update parameter corresponding to the input voice into the voice model. The point is that the determination is made so as to increase the update degree of the. According to the adaptive speech recognition apparatus of the third aspect, a message prompting the utterance of a specific recognition target word is determined, and when the input speech is detected within a predetermined time after the message is output. Since the update parameter corresponding to the input speech is determined so that the degree of update of the speech model is increased, the possibility that the update is performed for a meaningless utterance or an erroneous utterance can be reduced. .

According to a fourth aspect of the present invention, there is provided an adaptive speech recognition apparatus according to any one of the first to third aspects, wherein a normal mode and a learning mode for notifying a user of utterance contents in advance. And a mode selecting means for selecting one of the modes, wherein the update parameter determining means comprises:
The gist is that the update parameter is determined according to the mode selected by the mode selection means. According to a fifth aspect of the present invention, in the adaptive speech recognition apparatus according to the fourth aspect, the update parameter determining means is configured to execute the update mode more than in the case where the normal mode is selected by the mode selection means. The gist is that the update parameter is determined so that the degree of update of the speech model is increased when the learning mode is selected by the selection unit. According to the adaptive speech recognition apparatus of the fourth or fifth aspect, the determination of the update parameter is changed between the learning mode for notifying the user of the utterance content in advance and the normal mode. In the normal mode, it is possible to reduce the burden on the user while giving priority to the degree, and to suppress the adverse effects of meaningless utterances and the like, and as a result, it is possible to perform efficient and effective learning. Further, the invention according to claim 6 is the invention according to claims 1 to 5 above.
In the adaptive speech recognition device according to any one of the above,
A recognition result history storage unit that stores a history of the recognition target words already selected and determined by the recognition target word selection determination unit and the similarity thereof; and the update parameter determination unit is stored in the recognition result history storage unit. The gist is that the update parameters corresponding to the recognition target words having the low similarity among the recognition target words are determined so that the degree of updating of the speech model is increased. According to the adaptive speech recognition apparatus of the sixth aspect, the degree of update of the speech model of the update parameter corresponding to the recognition target word having a low similarity among the recognition target words that have already been selected for recognition is increased. Thus, the bias of the adaptation degree of the recognition target word can be reduced. According to a seventh aspect of the present invention, there is provided the adaptive speech recognition apparatus according to the third aspect, wherein the recognition target word which has already been selected and determined by the recognition target word selection determining means and a history of the similarity thereof are stored. A result history storage unit, wherein the utterance content requesting unit preferentially determines a message prompting the utterance of the recognition target word according to the similarity among the recognition target words stored in the recognition result history storage unit The main point is to do. According to the adaptive speech recognition apparatus of the seventh aspect, a message prompting the utterance of the recognition target word is preferentially determined in accordance with the similarity among the recognition target words that have already been selected for recognition. The bias of the adaptation degree of the recognition target word can be reduced.

According to an eighth aspect of the present invention, in the adaptive speech recognition apparatus according to any one of the first to seventh aspects, a first model registered in advance in the speech model and a user In the case where a second model uniquely registered by the first model is included, the update parameter determining means determines that the degree of update of the second voice model is higher than that of the second voice model belonging to the first model. The point is that update parameters corresponding to the voice model belonging to the above model are determined. According to the adaptive speech recognition apparatus of the eighth aspect, the speaker adaptation can be preferentially performed on the speech model uniquely registered by the user. According to a ninth aspect of the present invention, in the adaptive speech recognition apparatus according to the third or seventh aspect, the first model registered in advance in the speech model and the second model independently registered by the user. And the model of
The utterance content request unit determines, at a higher frequency than the speech model belonging to the first model, a message prompting the utterance of a recognition target word corresponding to the speech model belonging to the second model. Is the gist. According to the adaptive speech recognition apparatus of the ninth aspect, a message prompting the user to utter a recognition target word corresponding to a speech model uniquely registered by the user is preferentially uttered, and as a result, the user Thus, it is possible to preferentially perform speaker adaptation on a uniquely registered voice model. According to a tenth aspect of the present invention, in the adaptive speech recognition apparatus according to any one of the first to ninth aspects, the update parameter determining means determines the update parameter based on a length of the input speech. The main point is that the weighting is performed on. According to the adaptive speech recognition apparatus of the tenth aspect, it is estimated whether the input speech is appropriate for updating according to the length of the input speech. This can suppress the adverse effect on the speech model.

According to an eleventh aspect of the present invention, in the adaptive speech recognition apparatus according to any one of the first to tenth aspects, the update parameter determining means includes a similarity measure corresponding to a continuous input speech. The point is that the update parameter is set to a larger value when the value is equal to or more than the predetermined threshold value than when the value is equal to or less than the predetermined threshold value. According to a twelfth aspect of the present invention, in the adaptive speech recognition apparatus according to the eleventh aspect, the continuous input speech is a recognition target word corresponding to a first model registered in advance in the speech model. The gist is that the combination is a combination with a recognition target word corresponding to the second model uniquely registered by the user. According to the adaptive speech recognition apparatus of the present invention, when the similarity corresponding to the continuous input speech is equal to or greater than a predetermined threshold, the speech model is updated, so that it is meaningless. It is possible to reduce the possibility that the voice model is updated for an inappropriate utterance or an incorrect utterance. According to a thirteenth aspect of the present invention, in the adaptive speech recognition apparatus according to any one of the first to twelfth aspects, the speech model storage means is a non-volatile first read-only storage means. And a non-volatile and rewritable second storage means, wherein a pre-registered voice model is stored in the first storage means, and an updated part of the voice model updated by the voice model updating means is provided. , Is stored in the second storage means. Further, the invention according to claim 14 is based on the first aspect.
3. In the adaptive speech recognition device according to item 3, when the speech model updated by the speech model updating means is stored in the volatile storage means, the volatile storage means is turned off when the power is turned off. Subtracting the speech model stored in the first storage means from the updated speech model stored in the storage section to generate an updated part of the speech model, and storing the updated part in the second storage means. Is the gist. The invention according to claim 15 is the invention according to claim 1.
4. In the adaptive speech recognition device according to 4, when the power is turned on, the speech model stored in the first storage means and the speech model stored in the second storage means are updated. The gist of the present invention is that the part is added and transferred to the volatile storage means. According to a sixteenth aspect of the present invention, in the adaptive speech recognition apparatus according to the fifteenth aspect, the first storage means is a ROM, and the second storage means performs erasing for each block. The gist is that it is a memory. According to the adaptive speech recognition apparatus of any one of claims 13 to 16, even when the power supply is unexpectedly turned off, at least the initial speech model is stored in the nonvolatile, read-only first storage means. Is stored, the complete disappearance of the speech model can be prevented.

According to a seventeenth aspect of the present invention, in the adaptive speech recognition apparatus according to the sixteenth aspect, when two or more blocks of the flash memory are used, the two or more blocks are erased. The updated block of the voice model, error check data, and the number of times of writing after initialization of the flash memory are stored in the block in the state, and the error check is performed from the block having the largest number of times of writing after initialization. It is determined whether or not the updated part of the audio model is normal based on the data for use. Of the updated parts of the audio model determined to be normal, the updated part of the audio model in the block having the largest number of write operations after the initialization is determined. The updated part is transferred to the volatile storage means, and from the block having the updated part of the speech model determined to be abnormal, To become clear the update part of the voice model and the gist thereof. According to the adaptive speech recognition apparatus of the seventeenth aspect, an updated part of the speech model is added to a block in an erased state among two or more blocks.
The error check data and the number of writes after the initialization of the flash memory are stored, and the updated part of the voice model is normal by the error check data from the block having the largest number of writes after the initialization. The updated part of the audio model in the block having the largest number of times of writing after initialization among the updated parts of the audio model determined to be normal is transferred to the volatile storage means, Since the updated part of the voice model is erased from the block containing the updated part of the voice model determined to be abnormal, the power supply is suddenly turned off and the latest update stored in the volatile storage means is restored. Even if such a voice model is lost, the previous update state or the update state just before the previous one stores the flash memory in the volatile storage. It can be transferred to the stage. Therefore, there is almost no need to repeat the learning of the speech model.

The invention according to claim 18 is directed to an adaptive speech recognition apparatus according to any one of claims 1 to 17, and the recognition target word selected and determined by the recognition target word selection determining means. And a response control means for controlling a response to the input voice by selecting a response content stored in advance corresponding to the recognition target word on the basis of the above. According to a nineteenth aspect of the present invention, in the speech processing apparatus according to the eighteenth aspect, the response content is divided into a plurality of response levels and stored, and the similarity calculated by the similarity calculating means is stored. The gist of the invention is to provide a response level determining means for determining the response level based on the degree. Also,
According to a twentieth aspect of the present invention, in the speech processing apparatus according to the nineteenth aspect, the response level determining means sets a threshold value of the similarity for each response level. The point is that the response level is determined by comparing the threshold values of the above. According to a twenty-first aspect of the present invention, in the speech processing apparatus according to the nineteenth aspect, the response level determining means performs an arithmetic process on the similarity using a different coefficient for each response level. The gist is that the response level is determined by comparing the similarity after the arithmetic processing is performed using the coefficient with a predetermined unique threshold value. According to the speech processing device of any one of claims 18 to 21, the user is provided to the user in correspondence with the recognition target word recognized based on the speech model that has been suitably updated as described above. Since a response is made, an appropriate response can be selected. According to a twenty-second aspect of the present invention, in the audio processing device according to the twentieth aspect, it is preferable that a part or all of the plurality of thresholds is changed each time the input voice is present. Make a summary. According to a twenty-third aspect of the present invention, in the audio processing device according to the twenty-first aspect, a part or all of the coefficients are changed each time the input voice is present. According to the voice processing device of the present invention, the utterance of the response is changed each time the input voice is present, so that it is possible to ensure a variety of responses. The invention according to claim 24 is the above-mentioned claim 1.
24. An audio processing apparatus according to any one of claims 8 to 23, an audio synthesizing means, and a driving means for driving a movable portion, wherein the response control means controls the audio synthesizing means and the driving means. Thus, the toy is configured to respond to the input voice. Claim 24
According to the pet toy described in (1), by adjusting the degree of updating of the voice model with the updating parameter, the pet can gradually imitate the phenomenon that the pet gradually adapts to the owner, thereby allowing the user to attach to the pet toy. Easy to hold.

[0011]

Embodiments of the present invention will be described below with reference to the accompanying drawings to provide an understanding of the present invention.
The following embodiment is a specific example of the present invention and does not limit the technical scope of the present invention. FIG. 1 is a functional block diagram showing a schematic configuration of an adaptive speech recognition apparatus according to one embodiment of the present invention.
FIG. 2 is a diagram showing an example of a hardware configuration required for the adaptive speech recognition device. First, as shown in FIG. 1, an adaptive speech recognition apparatus A1 according to an embodiment of the present invention includes a feature extraction unit (corresponding to feature extraction means) 1 for extracting a feature of an input speech, and a recognition target word. Voice model storage unit (corresponding to voice model storage means) 2 for storing a voice model corresponding to.
And the feature amount extracted by the feature extraction unit 1;
An evaluation value calculation unit (corresponding to similarity calculation means) 3 for calculating a similarity with the voice model stored in the voice model storage unit 2 and a similarity calculated by the evaluation value calculation unit 3 A result determination unit (corresponding to a recognition target word selection determining unit) 4 for performing selection determination of a recognition target word corresponding to the input voice, and the speech A speech model updating unit (corresponding to a speech model updating unit) 5 for updating the speech model stored in the model storage unit 2 and an update parameter for controlling the degree of updating of the speech model by the speech model updating unit 5 are determined. And an update parameter determination unit (corresponding to an update parameter determination unit) 6. In order to implement the adaptive speech recognition apparatus, hardware including, for example, a microphone 101, an A / D converter 102, a processor 103, a ROM 104, a flash memory 105, and a work memory 106 as shown in FIG. . Among the above components, a specific example of the voice model storage unit 2 is the ROM 104 and the flash memory 105, and the other feature extraction unit 1, similarity calculation unit 3, recognition word determination unit 4, voice model update The unit 5 and the update parameter determination unit 6 can be realized by causing the processor 103 to execute a program in which processing corresponding to each component is described, using the work memory 106.

Next, the basic operation of the adaptive speech recognition apparatus A1 will be described. The above adaptive speech recognition device A
In 1, the input voice input from the microphone 101 is
It is supplied to the feature extraction unit 1. (If this is expressed in FIG. 2, the input voice input from the microphone 101 is supplied to the work memory 106 via the A / D converter 102 and the like, and the function corresponding to the feature extracting unit 1 is executed by executing the software. Although the processor 103 can be referred to by the realized processor 103, the relationship between the processor 103 and the work memory 106 in executing the program is the same as that of the well-known one. Functional expression according to the above). In the feature extraction unit 1,
The input speech is divided into predetermined units of time called frames, and feature quantities such as well-known LPC cepstrum and LPC mel cepstrum are extracted. Note that the types of the feature amounts are not limited to the two described above, and the order and the like are not limited. The speech model storage unit 2 stores in advance an HMM model corresponding to the recognition target word and a speech model such as a speech feature vector sequence in DP matching. Once extracted, the evaluation value calculation unit 3 calculates the similarity between the feature quantity and the speech model. For example, when the speech model storage unit 1 stores an HMM model constructed for each phoneme, syllable, or word, the similarity is calculated for each H with respect to the input speech.
It is represented by the likelihood of the MM word. In the case of the HMM model,
The greater the likelihood, the greater the similarity. At this time, if the HMM model is constructed for each phoneme or syllable, the HMM models are linked to create the HMM word. Each H calculated by the evaluation value calculator 3
The similarity for each MM word is supplied to the result determination unit 4.
The result determination unit 4 refers to the similarity calculated for each HMM word, and selects and determines, for example, a word having the highest similarity among the HMM words as a recognition word. Further, when the feature amount is a speech feature amount vector sequence in DP matching, the smaller the distance of the feature amount, the higher the similarity. It is selected and determined as a recognition word.

[0013] Speech recognition for input speech is performed in this manner. However, when an unspecified speaker is targeted, it is difficult to ensure a high recognition rate for all speakers. It may be preferable to adapt to the user by learning. This adaptation means updating the voice model stored in the voice model storage unit 2 according to the feature amount of the input voice. The speech model updating unit 5 in the adaptive speech recognition device A1 is for this purpose. Although it differs depending on the speech recognition method (HMM, DP matching, or how the data is held in the HMM), when the above-mentioned speech model is, for example, a continuous mixture distribution HMM, the feature amount average vector of the mixture normal distribution in the speech model is calculated as follows. The update that adapts the speech model to the speaker is performed by approaching that of the input speech. More specifically, when the similarity is calculated (or when the similarity is calculated again after the recognition word is determined).
As shown in FIG. 3, information on the state transition path for calculating the optimum likelihood is stored, and the corresponding state is obtained by backtracking in order from the last frame to the first frame of each state. Note that a method of equally allocating the number of states from the first frame to the last frame may be used.

When the speech feature vector of each frame is assigned to a state in this way, as shown in FIG. 4, the average vector νi () of the speech feature vector assigned to each state of the speech model HMM is next obtained. n) (i is the state number, n is the vector order). This is accommodated by spectral interpolation. That is, H before adaptation
The average vector of MM is expressed as μij (n) (i is the state number, j
Is the mixing number, and n is the vector order), the movement vector Δij (n) is calculated according to the following equation (1). Δij (n) = νi (n) −μij (n) (1) Using the above movement vector Δij (n), the average vector after updating the HMM is expressed by the following equation. μ′ij (n) = μij (n) + kΔij (n) (2) where k is an update parameter in adaptation, and the range of k is [0, 1]. The update parameter k is for controlling the degree of update of the speech model stored in the speech model storage unit 2 with respect to the input speech. If the update parameter k is not set and all input voices are updated equally, there is a possibility that the voice model may be deteriorated due to utterance of a meaningless word or erroneous utterance. Therefore, in the speech recognition apparatus A1 according to the present invention, the update parameter k is determined for the input speech so as to avoid undesired learning by determining the update parameter k according to a predetermined condition that is considered preferable for learning. Update parameter determination unit 6
Is provided. In the update parameter determination unit 6,
For example, the update parameter k is determined based on the similarity. When the similarity is high, the update parameter k is set to a large value, and when the similarity is low, the update parameter k is set to a low value.

In the example shown in FIG. 5, the similarity is divided into four levels, and three thresholds 1 are used for this division.
t, 2t, and 3t are set. The value of the update parameter k is set for each of the categories. When the similarity is smaller than the threshold value 1t, the determined value of the update parameter k is 0, and between the threshold value t1 and the threshold value t2, k = 1. From 2t to threshold 3t, k = 2, and above threshold 3t, k = 0.4. That is, when the similarity is case 1 smaller than 1t, the result is determined as a result of utterance of a meaningless word or erroneous recognition, k = 0, and no adaptive learning is performed. If the similarity increases, k = 0.1, 0.2, 0.4, which is larger, is given in accordance with the magnitude. Further, the setting of the update parameter k is not limited to the magnitude of the similarity, and may be set to a large value when the user intentionally indicates that the learning is required. If there is a switch 7 (corresponding to the mode selection means) for selecting one of a normal mode for performing a normal recognition operation and a learning mode for notifying the user of the utterance content in advance, When the switch 7 is switched to the learning mode, the switch is detected by the update parameter determination unit 6, and the update parameter k is set to a high value of 0.8, for example. In this case, the value of the update parameter k does not necessarily have to follow the magnitude of the similarity, and may be a constant value. This is because, as described above, in the learning mode, the device user is basically notified in advance of the content to be uttered, and the user is supposed to utter according to the notification content.

Of course, as shown in FIG. 6, even in the learning mode, the update parameter k is set according to the similarity.
May be changed. In the example of FIG. 6, in the normal mode, as in the case of FIG.
Are given values of 0, 0.1, 0.2, and 0.4, whereas in learning mode they are larger than those values.
The values 0.1, 0.2, 0.4, 0.8 are given. As described above, in the learning mode, a large value is given to the update parameter k to enhance the adaptation degree.
Learning efficiency can be improved. Furthermore, by performing weighting based on the voice section length, more reliable speaker adaptation is possible. In general, when a word is uttered, the utterance time tends to be proportional to the number of characters. For this reason, by considering the utterance time, it is possible to evaluate to a certain extent whether the utterance is meaningful. Therefore, a weighting l according to the utterance time is introduced into the above equation (2) as in the following equation (3). μ ′ ij (n) = μ ij (n) + 1 × k × Δ ij (n) (3) The weighting l is obtained from a large amount of speech of an unspecified speaker by HMM
When learning in advance, the speech length of the word HMM is statistically measured and stored together with the construction of the speech model, so that it can be determined according to the distribution. That is,
If many input voices are biased to a predetermined utterance time, a large weight is assigned to the utterance time.
For example, in FIG. 7, the maximum value of 1.0 is given when the voice section length is in the range from 2.0 seconds to 3.0 seconds, and the value is reduced on both sides of the range. When the voice section length is shorter than 1.0 second or 4.0
If it is greater than seconds, a value of 0 is given to the weight. This is because when the voice section length is shorter than 1.0 second or longer than 4.0 seconds, there is a high possibility that the voice is meaningless. By determining the values of the update parameter k and the weight 1 in this manner, the adaptive speech recognition apparatus according to the present embodiment can provide a speech model with no meaningful utterance or speech when adapting the speech model to a specific speaker. The adverse effects on the speech model due to incorrect recognition are suppressed.

Next, a speech processing device A2 and a pet toy A3 to which the adaptive speech recognition device A1 is applied will be described. The voice processing device A2 has a configuration in which, in addition to the configuration of the adaptive voice recognition device A1, a voice corresponding to the recognition target word recognized thereby is produced, and for example, a pet simulating a pet such as a dog or a cat It is embodied as being incorporated in the toy A3. Here, FIG. 8 shows a schematic configuration of the voice processing device A2 and a pet toy A3 incorporating the same. In FIG. 8, the update parameter determination unit 6
Are included in the voice model updating unit 5 and are represented. As shown in FIG. 8, in addition to the configuration corresponding to the adaptive speech recognition device A1, the speech processing device A2 is adapted to correspond to the recognition target word in advance based on the recognition target word determined by the result determination unit 4. And a response control unit (response control means) 8 for selecting a response content stored as a response and controlling a response to the input voice. The response control unit 8 includes a response level determination unit (response level determination means). ) 81 and a response determination unit 82. The response level determination unit 81
Is used to select the response level based on the similarity when the selected response content is divided into a plurality of response levels. When the response is included, it is for selecting one of them. The pet toy A3 incorporating the audio processing device A2 has a movable part 201 for operating a part 200 corresponding to the tail of a dog and the like, a motor 202, a driving part 203 for driving the motor, a dog A speech synthesizer 204 for synthesizing a voice simulating the sound of a cat or the like;
And a speaker 205 that outputs a synthesized voice synthesized by the P.4 as an audio signal.
The speech synthesizer 204 is controlled according to the response determined by the response determiner 82.

The pet toy A3 responds to an input voice from the user by, for example, a part 20 corresponding to the tail or the like.
0 responds to the user by operating 0 or uttering a cry such as "one-one". The voice processing device A2 for making such a response and the pet toy A incorporating the same.
The operation of No. 3 is as follows. The above audio processing device A2
In the toy A3 and the pet toy A3, when the result determination unit 4 of the adaptive speech recognition apparatus A1 determines the recognition target word selected and determined for the input speech, the recognition target word and the recognition word are determined. The similarity used at this time is supplied to the response level determining unit 81. The corresponding response content is stored in advance for the recognition target word to be recognized. The above-mentioned response contents include utterances of various kinds of calls, selection of specific parts or simultaneous movement, and the like, and a response level according to the similarity is further set in each response content. For example, the response level is as shown in FIG. In the example of FIG. 9, four response levels 11, 21, 31, and 41 are defined by three thresholds 1 t ′, 2 t ′, and 3 t ′, and the level 11 1 having the lowest similarity, that is, the threshold 1 t If the similarity is smaller than ', no response is made even if a voice is input from the microphone 101, and the similarity is between the response level 2l, ie, the threshold 1t' and the threshold 2t '. In some cases, the utterance “Wan Wan” is set, and as the similarity increases, the utterances “Hello” and “Hello, how are you?” Are set. The response level basically becomes more aggressive as the hierarchy becomes higher. The response content and the response level may be set for each recognition target word, may be set for a specific recognition target word, or may be set commonly for all recognition target words. May be.

Also, a threshold value 1 for the above similarity
t ', 2t', and 3t 'need not be fixed. In the case of the above-mentioned pet toy A3, it is preferable to change the way of responding according to the occasion so as not to bore the user. The setting of the thresholds 1t ', 2t', and 3t 'is changed so as not to change the vertical relationship of the response levels based on, for example, random variables. For generating the random variable, for example, an internal clock or a voltage value of a device can be used, but it is not particularly limited.
The setting of the threshold is changed by adding or subtracting this random variable to or from the original threshold. At this time, in order not to change the vertical relation of the response level, for example, another threshold 1 above and below the threshold 2t '
t 'and 3t' are set as a minimum value and a maximum value, and when the random variable is less than the minimum value or greater than the maximum value, the random variable is set in the range from the minimum value to the maximum value. The generation of the random variable may be repeated until the threshold value 2t 'falls. Thus, even when the recognition target word having a small number of response levels is uttered by the user, the response of the pet toy A3 can have a variety of responses. Furthermore, even if the device user utters the same word as the adaptation of the speech model progresses, the similarity increases, and accordingly the threshold value is also increased overall. By
The rejection effect on the utterance of another person can be increased. As a result, it is possible to imitate a phenomenon in which the pet becomes familiar with the owner and becomes inactive to utterances other than the owner, thereby increasing the user's attachment to the pet toy. The rejection effect can be effectively enhanced by gradually increasing the lower threshold value.

The threshold values 1t ', 2t', 3
As an index for increasing t ′, for example, the total use time acquired from the internal clock, the number of adaptations for each recognition target word, and the like can be used. Further, in the example of FIG. 9 described above, the similarity is associated with the response level by using three thresholds, but the present invention is not limited to this.
As shown in (1), only one threshold value is set, and the similarity is calculated based on the coefficients 1s, 2s, 3s, 4s, etc. corresponding to the respective response levels.
The response level may be selected based on whether or not the similarity calculated based on s, 3s, and 4s exceeds the one threshold. In the example of FIG.
Although the similarity exceeds the threshold value when the arithmetic processing is performed using s, for example, the one having the lowest value, that is, the level 3l may be selected. Even when one threshold value and a coefficient corresponding to each response level are used, as in the case of using a plurality of threshold values, the coefficient is randomly changed or recognized every time speech is input. It is possible to increase the coefficient in accordance with the degree of adaptation of the target word to ensure a variety of responses.
As described above, the response level determination unit 81 determines the response level based on the recognition target word and the similarity used in the selection determination of the recognition target word, and the response level is determined by the response determination. It is supplied to the section 82.

The response deciding section 82 determines a response finally made to the user and generates a control signal according to the determined response. The response level determination unit 81
When there is only one response corresponding to the response level supplied from, the control signal is sent to the driving unit 203 and the voice synthesizing unit 204 without any change. For example, as shown in FIG. 11, "responsible" or "what is it?"
This is a case in which other utterance contents such as are set. Further, as another example, when the response content for operating the part 200 corresponding to the tail is divided into a plurality of response levels according to the speed, the operation is performed in the horizontal direction for each response level. There are cases where a plurality of responses are assigned, such as operating as if drawing an image or operating in the front-back direction. The final response may be selected, for example, at random from the plurality of responses, or the number of responses may be counted for each response, and a response with a small number of responses may be preferentially selected. It may be. Then, a response is determined by the response determination unit 82, and the speech synthesis unit 2
When the control signal is sent to the drive unit 04 and the drive unit 203, the drive unit 203 and the voice synthesis unit 204 are operated according to the control signal, and an actual response is made. As described above, in the pet toy A3 incorporating the voice processing device, it is possible to secure a variety of responses and to imitate a phenomenon familiar to the owner, thereby improving the interactivity with the user.

When the adaptive speech recognition device A1 is applied to the pet toy A3 or the like as described above, a speech is input, and the speech model updated by the speech model updating unit 5 is used as a volatile model. There is a risk that the power may be unexpectedly turned off despite being in the work memory 106. Therefore, it is necessary to protect the updated voice model as much as possible when the power is suddenly turned off, and it is necessary to keep some voice model even if it disappears. For this reason, in the adaptive speech recognition apparatus A1, the initial speech model is stored in the non-volatile, read-only ROM 104, and only the data corresponding to the displacement of the speech model by the update is stored in the non-volatile, rewritable flash memory. It is stored in the memory 105. When the voice model and the data corresponding to the displacement of the voice model are used, the data is transferred to the work memory 106 used by the processor 103 in a state where both are added. On the other hand, when evacuation is performed from the work memory 106, data of the displacement obtained by subtracting the initial voice model from the updated voice model held in the work memory 106 is transferred to the flash memory 105. Is done. The displacement data transferred to the flash memory 105 is basically written in one block of the flash memory 105, and when erasing, all data in the block is collectively erased.

The recording process of the voice model when the power is turned off and the data corresponding to the displacement will be described more specifically. Here, FIG. 12 is a diagram for explaining a state in which the data of the displacement is normally transferred to one block of the flash memory, and FIG. 13 is a diagram for explaining the displacement of the data to one block of the flash memory. FIG. 9 is a diagram for explaining a state in which data for a minute has not been transferred normally. When the user instructs to turn off the power by turning off the power switch or the like, first, as shown in FIGS. 12C and 13C, the volatile and rewritable work memory 106 is used. Only the data corresponding to the displacement from which the value of the initial voice model is subtracted from the voice models stored in the memory 106 is saved in one block of the flash memory 105.
When the displacement data is completely saved, error check data such as a checksum for determining whether the displacement data has been normally saved is written to the end of the block. And the power supply circuit is shut down. When turning on the power again,
First, the processor determines whether or not the data of the displacement written in the flash memory is normal by using the error check data. At this time, if it is determined that the sound model is normal, as shown in FIG.
After adding the displacement data from the block of the flash memory 105, the work memory 1
06. When the data for the displacement is transferred to the work memory 106, the block holding the data for the displacement in the flash memory 105 is erased as shown in FIG. 12B. On the other hand, if it is determined that it is not normal, only the initial voice model stored in the ROM 104 is transferred to the work memory 106 as shown in FIG. Is set. At this time, as shown in FIG. 13 (b), the block holding the data content of the displacement is erased. In this manner, the initial voice model is stored in the nonvolatile read-only ROM 104, the displacement data is stored in the nonvolatile rewritable flash memory 105, and the data is transferred to the work memory 106 when necessary. Thus, it is possible to prevent the entire voice model from being accidentally lost when the power is turned off.

In the above example, the displacement data and the like are stored using only one block of the flash memory 105. However, if the flash memory 105 has a sufficient capacity, two or more blocks may be used. It is also possible to perform the processing by using. Here, FIG. 14 is a diagram for explaining a state in which the latest data of the displacement is normally transferred when two blocks of the flash memory are used, and FIG. 15 is a diagram of the flash memory. FIG. 10 is a diagram for explaining a state in which the latest data of the displacement is not transferred normally when two blocks are used. When the power is turned off, the displacement data related to the latest update is stored in one of the two blocks of the flash memory 105 that is not used, for example, the block B1. At this time, error check data such as a checksum and write count data (N-1) after initialization of the flash memory 105 are stored at a predetermined position of the block B1 in the flash memory 105, for example, at the end of the block B1. Is done. Next, when the power is turned on, as shown in FIGS. 14A and 15A, first, each block B1,
Regarding B2, the number-of-times-of-writing data (N-1) and (N-2) at the above-mentioned predetermined position are referred to, and the block B1 having the largest number of times, that is, the block B1 that has been updated most recently is set as a read target. Then, the error checking data of the block B1 set as the read target is referred to, and the processor determines whether or not the displacement data stored in the block B1 is normal.
As shown in FIG. 14A, if it is determined that the displacement data stored in the block B1 is normal, the displacement data is added to the initial voice model from the ROM 104. The data is transferred to the work memory 106. Further, when the data corresponding to the displacement is transferred to the work memory 106, as shown in FIG. 14 (b), the block B2 having the smaller number of times out of the write number data (N-1) and (N-2). The data corresponding to the displacement (corresponding to the update two times before) stored in the memory is deleted. Then, when the power is turned off again, the data of the displacement related to the latest update is stored in the block B2 which has been erased (0 × FFFF) by the block erasure.
At this time, error check data such as a checksum and the number-of-writes data N after initialization of the flash memory 105 are stored at a predetermined position of the block B2 in the flash memory 105, for example, at the end of the block B2.

On the other hand, as shown in FIG. 15A, when it is determined that the data corresponding to the displacement stored in the block B1 is not normal, the number of times of writing after initialization is one more. A small number of blocks B2 are set as read targets. Since it is already confirmed that the data of the displacement before last stored in the block B2 is normal, the data is added to the initial voice model from the ROM 104 and transferred to the work memory 106. Further, since the previous data that has not been normally stored in the block B1 is unnecessary, as shown in FIG. 15B, the block B1 is erased and the erased state is set. When the power is turned off again, each block B of the flash memory 105 is turned off.
For 1 and B2, the number-of-writes data at a predetermined position is confirmed, and the value is 0 × FFFF (erase state).
The data for the above displacement is written for the block B1 in the block B1, and the error check data and the number of times of writing N are stored at the end of the block B1. As described above, the flash memory 105 is used to store the displacement data.
When two or more blocks are used, the data of the displacement one immediately before the latest one is used. However, in the above adaptive speech recognition device, the updating degree is not set so high, so the speech recognition processing is not performed. Has a small effect. As a result, the power is suddenly turned off, and the work memory 10 is turned off.
Even if the voice model related to the latest update in 6 disappears,
There is almost no need to re-learn.

In addition to the above-mentioned power-off operation, in order to improve the applicability of the adaptive speech recognition device and the speech processing device to the pet toy and the like,
It is necessary to update the above speech model in a manner that makes it as transparent as possible to the user. Therefore, a modification A3 'of the pet toy including the adaptive speech recognition device A1' and the modification A2 'of the speech processing device will be described next.
Here, FIG. 16 shows a schematic configuration of a pet toy A3 'according to the modified example, and FIG. 17 shows functional blocks of the adaptive speech recognition device A1', the voice processing device A2 ', and the pet toy A3'. As shown in FIGS. 16 and 17, the adaptive speech recognition device A1 'included in the pet toy A3' and the speech processing device A2 'is the adaptive speech recognition device A1'.
In addition to the configuration described above, an utterance request determination unit 9 and a recognition result history storage unit 10 are further provided. The switch (mode selection means) 7 in the adaptive speech recognition apparatus A1 may be a switch 71 for a part corresponding to the mouth of the pet toy A3 ', whose owner tends to touch naturally for expressing affection, or a part corresponding to the forefoot. , 72 are provided. still,
The adaptive speech recognition device A1 ', speech processing device A2',
The control device 300 including the driving unit 203 and the voice synthesizing unit 204 is operated by being supplied with power from the power supply circuit 301 or the battery 302 as shown in FIG. The utterance request determination unit (corresponding to an utterance request determination unit) 9 determines a message for prompting the user to utter a specific recognition target word, and the message determined by the utterance request determination unit 9 is The voice is synthesized by the voice synthesizing unit 204 at an appropriate timing, and output from the speaker 205. For example, the message determined by the utterance request determination unit 9 is a message such as "Please say Konnichiwa" or "Please say Oide". When this message is uttered from the pet toy A3 ', the possibility that the owner utters "Connichiwa" or "Oide" in a short time is increased. Therefore, the utterance request determination unit 9
If the voice is detected by the voice detection unit 107 within a predetermined time after the message determined by the update parameter k is output, the update parameter k determined by the update parameter determination unit 6 is applied to the input voice. Set a large value for. This makes it possible to update the model effectively and appropriately while avoiding updating of the voice model due to meaningless voices and the like.

At this time, the value of the update parameter k is
It may be changed according to the similarity to the input voice,
Irrelevant, for example, a fixed value may be determined. Further, the value of the update parameter k may be reduced according to the time elapsed since the message was output. This is because, over time, the recognition target word prompted by the above message is less likely to be uttered,
This is another method for effectively and appropriately updating the model. Such determination of the update parameter k corresponds to the update parameter k in the learning mode described above. However, even with a message having the above content, unnaturalness becomes noticeable depending on the output timing. Therefore, the switch 7 (switches 71 and 72) for outputting the message, that is, for shifting to the learning mode.
The part that touches when petting, such as the back of the head or ears, that the owner touches naturally when expressing affection for the pet,
It should be provided in the mouth and other parts that are touched when feeding and when touched by hands. It is also important what kind of recognition target words are uttered by the above message. Of course, the utterance request determining unit 9 may randomly select a target recognition target word. However, even if the user only urges the utterance of the similar recognition target word, the adaptation proceeds to some extent, and This is because the learning effect on is reduced. For example, the update parameter k is often determined according to the similarity even in the learning mode. This is because unnecessary utterance may be input even after the message prompting the utterance is output, thereby avoiding adverse effects on the speech model. Therefore, if the initial speech model for a certain recognition target word does not match the owner, the similarity decreases and the value of the update parameter k is set to a small value. However, in such a case, it is better to actively update the voice model.

For this purpose, the adaptive speech recognition apparatus A
1 ′ is provided with a recognition result history storage unit 10 for holding the similarity by the evaluation value calculation unit 3 and the past history of the recognition target word by the result determination unit 4, and in this case, the utterance request determination The unit 9 selects a recognition target word having a low degree of similarity from past recognition target words stored in the recognition result history storage unit 10 and determines a message corresponding thereto. In addition, the recognition result history storage unit 10 also stores the number of past selection judgments of the recognition target word, and the utterance request determination unit 9 gives priority to the message corresponding to the recognition target word having the small number of selection judgments. Alternatively, it may be determined. At this time, the learning of the recognition target word may be promoted by determining a large value of the update parameter k in accordance with the low similarity or the small number of selection determinations. As a result, it is possible to perform appropriate and effective learning without bias for a plurality of recognition target words. Also, in the case of the pet toy A3 'as described above, the owner not only recognizes and responds that the owner utters a pre-registered word (a word corresponding to the voice model belonging to the first model) but also the owner. May need to respond to uniquely determined words (words corresponding to speech models belonging to the second model). For example, the name given to the pet by the owner may be unusual or completely creative, but when the owner utters the name to the pet toy, the need to respond is extremely high. However, the recognition rate of a word uniquely registered by the owner becomes extremely low unless learning is advanced. For this reason, it is necessary to set the value of the update parameter k larger for a unique word than for a word registered in advance. Further, it is preferable to increase the frequency at which the utterance request determination unit 9 determines a message corresponding to the unique word. Also,
As a method of avoiding the model from being updated due to unnecessary utterance or erroneous voice detection, it is also preferable to operate the adaptive speech recognition device, the voice processing device, and the pet toy using a combination of a plurality of words. . In other words, if the similarity of multiple words is greater than a certain threshold,
The update parameter k is set large. This gives
Even when an input signal having a high similarity to a registered word is detected, if the similarity of another word that has been input at the same time is low, the update parameter k is set to a small value. Also, for example, a combination of two words, such as the name of a call to a pet toy (independently registered) and an initial registration word such as a command, such as "Shiro Oide," is used within a certain time of the unique word. The model update may be performed only when the initial registration word is recognized.

The above adaptive speech recognition apparatus A as described above
An example of the processing procedure in 1 'will be described below with reference to FIGS. Here, FIG. 18 is a flowchart for explaining the determination of the utterance request in the adaptive speech recognition apparatus A1 ', and FIG.
It is a flowchart explaining the determination of model update in 2 '. As shown in FIG. 18, when the utterance request is determined by the utterance request determination unit 9, one word is detected based on the timing at which certain switches 71 and 72 are pressed and the speech recognition result, and the similarity is calculated. It is done at the timing. It is determined which of the timings is the cause of the determination (S801). In step S801, when it is determined that the switch is pressed down, the utterance request determination unit 9 refers to the recognition result history storage unit 10 and performs learning based on the number of past selection determinations and similarities of each recognition target word. The degree of progress Ki is calculated (S802).
This learning progress degree Ki is a positive value that decreases as the learning of the word i progresses. Next, the request frequency for each registered recognition target word is determined based on the learning progress degree Ki (S803). The request frequency is determined by a predetermined parameter F1 and the learning progress degree Ki.
The parameter may be a constant value, or more generally, a function such that the request frequency increases as the word progresses slowly. Further, the request frequency for each recognition target word uniquely registered by the owner is determined based on the learning progress degree Ki and the parameter F2 having a value different from the parameter F1 (S804). The above parameters F1,
By appropriately adjusting the parameter F2, the request frequency for each uniquely registered recognition target word can be made higher than each previously registered recognition target word. Finally, a recognition target word is selected using a random number or the like according to the determined frequency, and a message prompting the utterance of the selected recognition target word is determined (S805). If the learning progress degree of all the words has progressed to some extent, a determination that the word is not requested to be spoken may be added to the above processing. On the other hand, when the utterance request determination unit 9 operates at the timing when a certain word is recognized, a value K representing the learning progress degree similar to the learning progress degree Ki is obtained (S806), and the value K is determined. , Is compared with a certain threshold value K0 (S807). If the value K is larger than a certain threshold value K0, that is, if the learning progress degree for a certain word is not advanced, the word is selected as a word that prompts utterance (S808). If the value K is smaller than a certain threshold value K0, it means that the learning progress degree for a certain word has progressed to some extent, so that a determination that there is no request is made (S809). Here, for the value K, for simplicity, the distance between the word model and the feature amount of the detected speech, which has the opposite relationship to the similarity, may be used as it is.

Next, in the update parameter determination unit 6,
An example of a process for determining whether to update a model will be described with reference to FIG. As shown in FIG. 19, the update parameter determination unit 6 determines whether or not to update the model, and determines the value of the update parameter k. For this purpose, first, a learning progress degree K is calculated (S901), and further, the update parameter determination unit 6 operates based on a voice uttered in response to the utterance request message, or based on a normal voice recognition result. It is determined whether the operation has been performed, that is, whether it is in the learning mode or the normal mode (S902). If it is determined in step S902 that the learning mode is set, it is determined whether the recognition target word that prompted the utterance has been registered in advance or has been uniquely registered by the owner (S903). Then, the threshold values T1 × K and T1 of the similarity are determined according to whether they are registered in advance or are registered independently by the owner.
2 × K and the values S1 × K, S2 × K of the update parameter k
Is determined (S904, S905). On the other hand, if it is determined in step S902 that the mode is the normal mode, it is determined whether the speech recognition target word has been registered in advance or has been uniquely registered by the owner (S906). ). Then, the threshold values T3 × K and T4 × K of the similarity and the values S3 × K and S4 × K of the update parameter k are determined depending on whether they have been registered in advance or registered by the owner. (S907,
S908). The parameters T1 to T4 and S1 to S4 are used to adjust the similarity and the update parameter, respectively. Then, the above step S9
04, S905, S907, and the threshold value of the similarity calculated in S908 and the similarity to the recognition target word to be updated are compared (S909). If the similarity of the recognition target word to be updated is larger than the similarity threshold, the process S904
The update parameter k calculated in S905, S907, and S908 is adopted (S910). When the similarity of the recognition target word to be updated is smaller than the similarity threshold, it is determined not to update the model, that is, the update parameter k is set to 0. It is set (S911). As described above, in the adaptive speech recognition device, the speech processing device, and the pet toy according to the present embodiment, the adverse effect on the updating of the speech model due to meaningless utterance or erroneous utterance is suppressed. Since the value of the speech parameter is determined according to the progress of the learning, etc., it is possible to prevent the learning from being biased toward the recognition target word, and to perform the learning quickly and appropriately when necessary. In addition, even if the power is suddenly turned off, all voice models are prevented from disappearing. For example, when two or more blocks of the flash memory are used, in most cases, at least, the update amount of the last update is used. Since the data can be used, there is almost no need to repeat the learning even when the updated speech model is lost. Further, by adjusting the voice parameters, it is possible to cause the pet toy to imitate, for example, a phenomenon in which the pet becomes familiar with the owner or actively responds only to the voice of the owner.

[0031]

In the above embodiment, the movement vector Δij (n) is multiplied by the update parameter k in equation (2). However, the present invention is not limited to this, and the movement vector Δij (n) is added to the movement vector Δij (n). Alternatively, multiplication and addition may be performed. Also, when updating the speech model, it is assumed that all of the mixture distributions in the speech model are moved. However, the present invention is not limited to this. For example, νi (n ) And μij
By using the vector distance of (n), the one having the smallest value may be moved. In the above embodiment, the adaptive speech recognition device and speech processing device according to the present invention are applied to a pet toy. However, the present invention is not limited to this. For example, a car navigation system or a portable music playback device may be used. It is also possible to apply to other devices. It is easier to secure the capacity of each of the ROM 104, the flash memory 105, and the work memory 106 in the above-described embodiment if they are externally attached. You may do so.

In the above embodiment, the model is updated only when the similarity is equal to or greater than the threshold value. In this case, the lower the degree of progress of the learning, the higher the adaptation speed for updating the model. We decided to set the parameters, but these fixed rules alone may not work. For example, for words whose learning progress is low, the threshold for judging the model update is set low and the adaptation speed is also fast. In some cases, learning is done with another word instead of a word. In order to avoid this, a functional relationship may be provided between the degree of progress of learning and the set adaptive speed as shown in FIG. 20, for example.
In this case, the learning progress degree and the adaptation speed are not necessarily in a monotonous magnitude relation, and adjustments such as proceeding the adaptation at a cautious speed can be made for words for which the learning has not progressed extremely as shown in FIG. Becomes Furthermore, if such a fixed functional relationship is used, if the user happens to have a sound quality that cannot obtain a very high degree of similarity with the initial value of the model, the model can be updated forever. Since a phenomenon in which learning by does not progress may occur, a threshold or an adaptation speed parameter may be determined by adding a random parameter such as a random number to such a fixed rule. Then, even if the user has any kind of sound quality, the model may be updated by random numbers in some cases, so that a fatal phenomenon that the user does not adapt at all can be avoided. in this way,
There are various variants of the parameter determination rules.
The present invention includes these.

[0033]

As described above, according to the adaptive speech recognition apparatus of the first aspect, when updating the speech model based on the feature amount of the input speech, the speech model of the speech model is updated by the update parameter. Because the degree of update is controlled,
It is possible to suppress an adverse effect on the updated speech model due to a meaningless utterance or an erroneous utterance. Further, according to the adaptive speech recognition apparatus of the second aspect, since the update parameter is determined based on the similarity, for example, the update degree increases as the similarity increases. By determining the update parameter, the adverse effect on the updated speech model can be suppressed. According to the adaptive speech recognition device of the third aspect, a message prompting the utterance of a specific recognition target word is determined, and the input speech is detected within a predetermined time after the message is output. In such a case, the update parameter corresponding to the input speech is determined so as to increase the degree of update of the speech model, so that the possibility of updating the meaningless utterance or erroneous utterance is reduced. Can be. According to the adaptive speech recognition apparatus of the fourth or fifth aspect, the determination of the update parameter is changed between the learning mode for notifying the user of the utterance content in advance and the normal mode. In the normal mode, the burden on the user can be reduced while suppressing an adverse effect due to meaningless utterance or the like in the normal mode. As a result, efficient and effective learning can be performed. According to the adaptive speech recognition apparatus of the sixth aspect, the update parameter corresponding to the recognition target word having a low similarity among the recognition target words that have already been selected for recognition is the update degree of the speech model. Is determined so as to increase, so that the bias of the adaptation degree of the recognition target word can be reduced. According to the adaptive speech recognition apparatus of the seventh aspect, a message prompting the utterance of the recognition target word is preferentially determined according to the similarity among the recognition target words that have already been selected for recognition. Therefore, it is possible to reduce the bias of the adaptation degree of the recognition target word. Further, according to the adaptive speech recognition device of the eighth aspect, the speaker adaptation can be preferentially performed on the speech model uniquely registered by the user. According to the adaptive speech recognition apparatus of the ninth aspect, a message prompting the user to utter a recognition target word corresponding to a speech model uniquely registered by the user is preferentially uttered. The speaker adaptation can be preferentially performed on the voice model uniquely registered by the user. According to the adaptive speech recognition apparatus of the tenth aspect, it is estimated whether or not the input speech is appropriate for updating according to the length of the input speech. , The adverse effect on the updated speech model can be suppressed.

According to the adaptive speech recognition apparatus of the present invention, when the similarity corresponding to the continuous input speech is equal to or greater than a predetermined threshold, the speech model is updated. Therefore, it is possible to reduce the possibility that the speech model is updated for a meaningless utterance or an erroneous utterance. Further, in any one of the above-mentioned claims 13 to 16,
According to the adaptive speech recognition apparatus described in the section, even when the power supply is unexpectedly turned off, at least the initial speech model is stored in the non-volatile first read-only storage means. Complete disappearance can be prevented. According to the adaptive speech recognition apparatus of the seventeenth aspect, an updated part of the speech model, error check data, and a flash memory are added to an erased block of two or more blocks. The number of times of writing after initialization is stored, and from the block having the largest number of times of writing after initialization, it is determined whether the updated part of the voice model is normal by the error check data and is determined to be normal. The updated portion of the voice model in the block having the largest number of times of writing after the initialization among the updated portions of the voice model is transferred to the volatile storage means, and the updated voice model of the voice model determined to be abnormal is transferred. Since the updated part of the voice model is deleted from the block where the updated part is located, the power is suddenly turned off and stored in the volatile storage means. Even speech model according to the latest update had disappeared, it is possible that or previous update state, the update state of the second last and transfers the flash memory in the storage means of the volatility. Therefore, there is almost no need to repeat the learning of the speech model. According to the speech processing device of any one of claims 18 to 21, the speech processing apparatus is used in correspondence with a recognition target word recognized based on the speech model that has been suitably updated as described above. Since a response is made to the person, an appropriate response can be selected. Further, according to the voice processing device of the present invention, since the utterance of the response is changed each time the input voice is present, it is possible to ensure a variety of responses. According to the pet toy of the twenty-fourth aspect, by adjusting the update degree of the voice model by the update parameter, the user can imitate a phenomenon in which the pet gradually adapts to the owner, thereby enabling the user to do so. The attachment to the pet toy can be easily held, and only the user can positively respond.

[Brief description of the drawings]

FIG. 1 is a functional block diagram showing a schematic configuration of an adaptive speech recognition device A1 according to an embodiment of the present invention.

FIG. 2 is a diagram illustrating an example of a hardware configuration required for the adaptive speech recognition apparatus A1.

FIG. 3 is a diagram for explaining a process of obtaining a corresponding state by backtrack processing.

FIG. 4 is a diagram for explaining movement of a mixture distribution.

FIG. 5 is a diagram for explaining an example of determining update parameters.

FIG. 6 is a view for explaining a case where a determined value of an update parameter is changed between a normal mode and a learning mode.

FIG. 7 is a view for explaining how to adjust the update degree of a voice model according to the length of an input voice.

FIG. 8 shows an audio processing device A2 according to an embodiment of the present invention.
The figure which shows the schematic structure of pet toy A3.

FIG. 9 is a view for explaining that a response level is selected in accordance with a degree of similarity when a plurality of threshold values are used.

FIG. 10 is a view for explaining that a response level is selected in accordance with a degree of similarity when only one threshold value is used.

FIG. 11 is a diagram for explaining a case where a plurality of responses are set for one response level.

FIG. 12 is a view for explaining a state where data corresponding to the displacement is normally transferred to one block of the flash memory;

FIG. 13 is a view for explaining a state in which data of the displacement is not normally transferred to one block of the flash memory.

FIG. 14 is a view for explaining a state where the latest data of the displacement is normally transferred when two blocks of the flash memory are used.

FIG. 15 is a view for explaining a state where the latest data of the displacement is not normally transferred when two blocks of the flash memory are used.

FIG. 16 is a pet toy A according to an embodiment of the present invention.
The figure which shows schematic structure of 3 '.

FIG. 17 shows an adaptive speech recognition device A1 ′, a speech processing device A2 ′, and a pet toy A according to an embodiment of the present invention.
Functional block diagram regarding 3 '.

FIG. 18 is a flowchart illustrating an example of a procedure for determining an utterance request in the adaptive speech recognition apparatus A1 ′.

FIG. 19 is a flowchart illustrating an example of a model update procedure in the adaptive speech recognition apparatus A1 ′.

FIG. 20 is a diagram illustrating an example of a functional relationship between a learning progress degree and an adaptation speed.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Feature extraction part 2 ... Speech model storage part 3 ... Evaluation value calculation part 4 ... Result judgment part 5 ... Speech model update part 6 ... Update parameter determination part 7 ... Switch 8 ... Response control part 9 ... Speech request part 10 ... Recognition Result history storage units 71, 72 switch 103 processor 104 ROM 105 flash memory 106 work memory 201 movable unit 202 motor 203 drive unit 204 voice synthesis unit

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI Theme coat ゛ (Reference) G10L 3/00 551H 571T (72) Inventor Hiroshi Hashimoto 1-5-5 Takatsukadai, Nishi-ku, Kobe-shi, Hyogo Stock Kobe Steel, Ltd. Kobe Research Institute (72) Inventor Yoshiro Nishimoto 1-5-5 Takatsukadai, Nishi-ku, Kobe-shi, Hyogo Prefecture Kobe Steel Works Kobe Research Institute F-term (reference) 2C150 BA11 CA02 DF02 DF04 DF31 ED56 EF30 5D015 GG01 GG04 GG06 KK01 KK02 KK04 LL10

Claims (24)

[Claims] 1. A feature amount extraction unit for extracting a feature amount of an input speech, a speech model storage unit for storing a speech model corresponding to a recognition target word, the feature amount extracted by the feature amount extraction unit, A similarity calculating means for calculating a similarity with the voice model stored in the voice model storing means; and a recognition target word corresponding to the input voice based on the similarity calculated by the similarity calculating means. A speech model updating means for updating the speech model stored in the speech model storage means based on the feature quantity extracted by the feature quantity extracting means; An adaptive speech recognition apparatus comprising: an update parameter determining unit that determines an update parameter for controlling a degree of updating of the voice model by the voice model updating unit. 2. The adaptive speech recognition apparatus according to claim 1, wherein said update parameter determination means determines said update parameter based on said similarity calculated by said similarity calculation means. 3. An utterance request deciding means for deciding and outputting a message prompting utterance of a specific recognition target word, wherein said update parameter deciding means outputs said message decided by said utterance request deciding means. The adaptive type according to claim 1, wherein when the input voice is detected within a predetermined time from, the update parameter corresponding to the input voice is determined so that the degree of updating of the voice model is increased. Voice recognition device. 4. A mode selecting means for selecting one of a normal mode and a learning mode for notifying the user of the utterance content in advance, wherein said update parameter determining means selects said mode by said mode selecting means. And determining the update parameter according to the set mode.
An adaptive speech recognition device according to any one of the preceding claims. 5. The degree of updating of the voice model when the learning mode is selected by the mode selecting means, rather than when the normal mode is selected by the mode selecting means. 5. The adaptive speech recognition device according to claim 4, wherein the update parameter is determined so that the value of the update parameter increases. 6. A recognition result history storage unit for storing a history of recognition target words already selected and determined by the recognition target word selection determining unit and a similarity thereof, and wherein the update parameter determining unit includes the recognition result history. The update parameter corresponding to the recognition target word having a low similarity among the recognition target words stored in the storage unit is determined so that the degree of update of the speech model is increased. An adaptive speech recognition device according to the section. 7. A recognition result history storage unit for storing a history of recognition target words already selected and determined by the recognition target word selection determining unit and a similarity thereof, wherein the utterance content requesting unit stores the recognition result history. 4. The adaptive speech recognition device according to claim 3, wherein a message prompting the utterance of the recognition target word among the recognition target words stored in the storage unit is preferentially determined according to the similarity. 8. When the voice model includes a first model registered in advance and a second model uniquely registered by a user, the update parameter determining means includes:
8. An update parameter corresponding to the speech model belonging to the second model is determined so that the degree of updating of the speech model belonging to the first model is higher than that of the speech model belonging to the first model. The adaptive speech recognition device according to claim 1. 9. When the voice model includes a first model registered in advance and a second model uniquely registered by a user, the utterance content request unit transmits the first model.
4. A message which prompts the user to utter a recognition target word corresponding to the voice model belonging to the second model at a higher frequency than the voice model belonging to the model.
Or the adaptive speech recognition device according to 7. 10. The adaptive speech recognition apparatus according to claim 1, wherein said update parameter determining means weights said update parameters based on a length of said input speech. 11. The update parameter determining means sets the update parameter to a larger value than when the similarity corresponding to the continuous input speech is equal to or greater than a predetermined threshold value. The adaptive speech recognition apparatus according to any one of claims 1 to 10, wherein the apparatus is set. 12. The continuous input voice includes a recognition target word corresponding to a first model registered in advance in the voice model and a recognition target word corresponding to a second model uniquely registered by a user. The adaptive speech recognition device according to claim 11, wherein the combination is a combination with the following. 13. The voice model storage means includes a non-volatile first read-only storage means and a non-volatile rewritable second storage means. 13. The speech model update part stored in the first storage means and updated by the speech model update means is stored in the second storage means.
An adaptive speech recognition device according to any one of the preceding claims. 14. When the power is turned off when the voice model updated by the voice model updating means is stored in the volatile storage means, the update stored in the volatile storage means is performed. 14. The adaptation according to claim 13, wherein an updated part of the speech model is generated by subtracting the speech model stored in the first storage means from the completed speech model, and stored in the second storage means. Type speech recognition device. 15. When the power is turned on, the voice model stored in the first storage means and an updated part of the voice model stored in the second storage means are added. 15. The adaptive speech recognition apparatus according to claim 14, wherein said apparatus is transferred to said volatile storage means. 16. The adaptive speech recognition apparatus according to claim 15, wherein said first storage means is a ROM, and said second storage means is a flash memory for erasing each block. 17. When two or more blocks of the flash memory are used, an updated part of the voice model, error check data, and the flash memory are added to an erased block of the two or more blocks. And the number of times of writing after initialization are stored. From the block having the largest number of times of writing after initialization, it is determined whether or not the updated part of the voice model is normal by the error check data. The updated voice model update portion in the block having the largest number of times of writing after the initialization among the determined update portions of the voice model is transferred to the volatile storage means, and the voice model determined to be abnormal is transferred. 17. The adaptive speech recognition apparatus according to claim 16, wherein said updated part of said speech model is deleted from a block having said updated part. 18. An adaptive speech recognition apparatus according to claim 1 and said recognition target word selected and determined by said recognition target word selection determining means.
A speech processing apparatus comprising: a response control unit that selects a response content stored in advance corresponding to the recognition target word and controls a response to the input voice. 19. A response level determining means for determining the response level based on the similarity calculated by the similarity calculating means, wherein the response content is divided into a plurality of response levels and stored. 19. The voice processing device according to claim 18, comprising: 20. The response level determination means sets the similarity threshold for each response level,
20. The speech processing device according to claim 19, wherein the response level is determined by comparing the similarity with a plurality of thresholds. 21. The response level determining means performs arithmetic processing on the similarity using different coefficients for each response level, and performs the arithmetic processing on the similarity using the coefficients. 20. The voice processing device according to claim 19, wherein the response level is determined by comparing the response level with a predetermined threshold value. 22. The apparatus according to claim 2, wherein a part or all of the plurality of thresholds is changed every time the input voice is present.
0. The audio processing device according to 0. 23. The audio processing apparatus according to claim 21, wherein a part or all of the coefficients are changed every time the input audio is present. 24. An audio processing apparatus according to claim 18, comprising: a voice synthesizing unit; and a driving unit for driving a movable unit, wherein said response synthesizing unit controls said voice synthesizing unit. And a pet toy which responds to the input voice by controlling the driving means.