JP2000089780A - Voice recognition method and voice recognition device - Google Patents

Abstract

(57) [Summary] [Problem] To apply speech recognition technology for unspecified speakers to inexpensive products, the memory capacity and CPU
It is difficult to use large-scale standard speaker voice model data from the viewpoint of the processing capability of. SOLUTION: A specific speaker voice model data 92 is created from voice data of an unspecified speaker group belonging to the predetermined range based on a speaker's age, gender, or the like. Then, using the specific speaker voice model data 92, the voice input by the speaker to be subjected to voice recognition is recognized.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method for recognizing speech uttered by a speaker, and more particularly to a speech recognition method for performing speech recognition by narrowing a speaker to be recognized to a certain range. The present invention relates to a voice recognition device.

[0002]

2. Description of the Related Art In recent years, electronic devices using voice recognition technology have been used in various fields, and one of them is a clock called a voice clock. The audio clock can set the current time and the alarm time by voice, and the apparatus can notify the current time by voice.

[0003] Such an audio timepiece is often used not only as a practical product but also as a toy for children, and it is desired to keep the cost of the device itself as low as possible. For this reason, there are great restrictions on the processing capacity of the CPU to be used, the capacity of the memory, and the like, and one of the important issues is how to provide high-performance functions under the restrictions.

[0004]

Conventionally, many devices using this type of speech recognition can perform speech recognition for an unspecified speaker. To enable speech recognition, large-scale standard speaker speech model data is required, which requires a large-capacity ROM or a CPU with a high processing capability, which results in high costs. Was the fact.

[0005] In addition, even though it is intended for unspecified speakers, the age of use and the gender of men and women are often limited to some extent depending on the type of device. Means that standard speaker model data limited to a certain range is sufficient. For this reason, even if large-scale standard speaker voice model data is held, there is a lot of waste, and it is possible to cope with a wide range of unspecified speakers. This means that they can respond on average, and there was also a problem with the recognition rate.

Accordingly, the present invention makes it possible to reduce the size of standard speech model data and to perform recognition at a high recognition rate when the speech recognition technology is applied to a device having a large restriction on the memory capacity and CPU processing capacity. It is an object of the present invention to provide a voice recognition method and a voice recognition device.

[0007]

In order to achieve the above-mentioned object, a speech recognition method according to the present invention is intended for speech recognition in which a plurality of unspecified speakers are recognized and a plurality of predetermined words can be recognized. In the method, among the unspecified plural speakers, the recognition target speaker is narrowed down to a predetermined range based on age, gender, etc., and specific speaker group voice model data is obtained from the voice data of the unspecified plural speakers belonging to the range. Is created, and the voice input by the recognition target speaker is recognized using the specific speaker group voice model data.

The speaker to be recognized is composed of a plurality of speaker groups based on voice characteristics, and speech data corresponding to the plurality of speaker groups is obtained from voice data of unspecified plural speakers belonging to each speaker group. Specific speaker group voice model data is created.

Further, the speech recognition apparatus of the present invention is intended for an unspecified plurality of speakers and is capable of recognizing a plurality of predetermined words. Voice analysis means for analyzing voice, and voice data of unspecified plural speakers belonging to the range by narrowing down the recognition target speaker among the unspecified plural speakers to a preset range based on age, gender, etc. And a control means for recognizing the voice input by the recognition target speaker using the specific speaker group voice model data created from the above.

Further, the recognition target speaker is composed of a plurality of speaker groups based on voice characteristics, and the specific speaker group voice model data includes voices of unspecified plural speakers belonging to each speaker group. It is composed of specific speaker group voice model data corresponding to the plurality of speaker groups created from the data.

As described above, according to the present invention, the memory capacity and C
The present invention exerts a great effect when applied to an apparatus having a great restriction on the processing capacity of the PU. In particular, the present invention is effective when the user of the device is limited to a certain range, for example, mainly a child or mainly a woman.

In the case where the number of speakers to be recognized can be narrowed down to a certain range, even if standard speaker voice model data that can support speakers in all ranges is used, there is much waste and the recognition rate cannot be expected much. . Therefore, among the unspecified multiple speakers, the recognition target speaker is narrowed down to a preset range based on age, gender, etc., and specific speaker group voice model data is created from voice data of a plurality of speakers belonging to the range. The specific speaker group voice model data is used to recognize the voice input by the target speaker for voice recognition.

[0013] Thus, since it is sufficient to have the specific speaker group voice model data corresponding to a certain range of speaker groups, the data amount of the voice model data can be significantly reduced. As a result, the storage capacity of the storage means for storing the data can be small, and the load of the recognition processing on the CPU can be reduced. In addition, since the data is specific speaker group voice model data corresponding to a specific speaker group, recognition performance is greatly improved.

Further, some specific speaker group voice model data corresponding to such a certain range of speaker group can be prepared. For example, it is also possible to prepare corresponding to several speaker groups such as for adult men, adult women, and children. According to this, it is possible to cope with a case where one device is used by a family. In this way, even if we have specific speaker group voice model data corresponding to several speaker groups, rather than having standard speaker voice model data created to support all ranges of speakers, Since the scale of the model data can be reduced, and since it is the specific speaker group voice model data corresponding to each speaker group, the recognition performance is greatly improved.

[0015]

Embodiments of the present invention will be described below with reference to the drawings. In this embodiment,
As one example of the voice recognition device, a voice clock device will be described as an example. This audio clock device (hereinafter sometimes simply referred to as a device) can set a current time and an alarm time by a voice spoken by a speaker, and can also notify the time from the device side by a voice. It is assumed that the device has a strong toy-like element that enables simple content conversation with a speaker.

FIG. 1 is a block diagram showing components necessary for constituting an audio clock device according to an embodiment of the present invention. First, the configuration shown in FIG. 1 will be schematically described, and thereafter, portions particularly necessary for describing the embodiment of the present invention will be described in detail.

This audio clock device can be roughly divided into a microphone 1 as an audio input unit, an audio input amplifier (referred to as a microphone amplifier) 2, an A / D conversion unit 3, an audio analysis unit 4, and a control unit for performing various processes. (CPU) 5, a D / A converter 6, an audio output amplifier 7 and a speaker 8 as an audio output unit, a ROM 9 as a first storage unit, a RAM 10 as a second storage unit, It comprises an operation unit 11 for performing various settings, a clock unit 12 for generating a time signal, and the like.

The voice analysis unit 4 analyzes the A / D-converted voice signal at short intervals by using an arithmetic unit to obtain a pitch cycle and the like, and also obtains feature data (LP) representing voice characteristics.
C cepstrum coefficient is generally output as time-series data.

The control unit 5 performs various processes described below, such as voice recognition, learning, registration, and voice output. These processes will be sequentially described in the description of operations to be performed later.

The ROM 9 stores output voice data 91, specific speaker group voice model data 92, a codebook (referred to as a specific speaker group codebook) 93 created based on the specific speaker group voice model data 92, and further. Stores a universal codebook 94 and the like.

The output voice data 91 is voice data for utterance contents for making a response or inquiry to the speaker from the apparatus side. The specific contents will be described in an operation example described later.

The specific speaker group voice model data 92
Is standard speaker voice model data for the recognition target speaker narrowed down to a certain range. For example, when the speaker to be recognized is narrowed down to children, it is standard speaker voice model data created from voice data of an unspecified number of children.

However, in this embodiment, an adult male
Three speaker groups of adult women and children are set, and adult men are classified as speaker group A, adult women as speaker group B, and children as speaker group C. The first specific speaker group voice model data MD1 created from the voice data of an unspecified number of adult men belonging to the speaker group A, and the speech data of an unspecified number of adult women belonging to the speaker group B are generated. It is assumed that there are second specific speaker group voice model data MD2 and third specific speaker group voice model data MD3 created from voice data of an unspecified number of children belonging to the speaker group C.

The first to third specific speaker group voice model data MD1, MD2, MD3 are composed of several word sets w1, w composed of predetermined recognizable words.
The specific speaker group voice model data stored in the word set w1 is considered to be particularly frequently used in conversations with the device. word (in this case, "good morning", "I'm home", "Hello", "Good night", "when", and is a word such as "alarm") that is a specific speaker group sound model data for. The specific speaker group voice model data for each word in this word set w1 is:
It is also used for performing a speaker learning process described later.

The specific speaker group voice model data stored in the word set w2 is "1 o'clock", "2 o'clock",
Specific speaker group voice model data for words representing time units such as "hour", "3 o'clock", ..., "12:00", and a specific speaker group voice model stored in word set w3. The data here is "1 minute", "2 minutes",
It is assumed that the specific speaker group voice model data is for a word representing a minute unit such as “3 minutes”,..., “59 minutes”.

In addition, words indicating affirmative / negative such as "yes" and "no", and specific speaker group voice model data for words necessary for having a conversation between the speaker and the device are also illustrated here. It is provided as an unset word set.

The specific speaker group codebook 93 is composed of a first specific speaker group codebook CB1 and a second specific speaker group voice created based on the first specific speaker group voice model data MD1. It has a second specific speaker group codebook CB2 created based on the model data MD2 and a third specific speaker group codebook CB3 created based on the third specific speaker group voice model data MD3. ing.

[0028] By the way, belongs to the word set w1 of the above-mentioned "good morning", "I'm home", "Hello", "Good night", "when", the words such as "alarm", as used herein, is also used as a registered word, recognition The target speakers, that is, the respective speakers (here, three speakers, speaker # 1, speaker # 2, and speaker # 3) using this device,
By uttering the above-mentioned registered words, each feature data can be stored in the RAM 10 as the registered word data 101.

The registered word data 101 uses the registered word data RD1 for speaker # 1, the registered word data RD2 for speaker # 2, and the registered word data RD3 for speaker # 3. It is stored for each speaker who is likely to do so.

As described above, the function of creating registered word data for some words for each speaker who is likely to use the apparatus can be arbitrarily performed by the user. Therefore, it is possible to create and store the registered word data using the function, but it is not always necessary to use the function. To execute this function, a predetermined mode setting is performed by the operation unit 11 for registration. The registration method and the like will be described later.

The RAM 10 additionally stores the quantized input speaker codebooks VQ1 and VQ1 created for speaker # 1.
The input speaker codebook VQ2 after quantization created for speaker # 2 and the input speaker codebook VQ3 after quantization created for speaker # 3 are stored. Will be described later.

One of the features of the apparatus according to the embodiment of the present invention is that a user who uses the apparatus is limited to a certain range,
A specific speaker group voice model data created from voice data of an unspecified number of speakers belonging to the range is created, and the voice of a user of the apparatus is recognized and processed using the specific speaker group voice model data. It is mentioned.

In order to realize this, in this embodiment, the users are divided into a speaker group A (adult male), a speaker group B (adult female), and a speaker group C (child). 1st to 1st data created from voice data of an unspecified number of speakers belonging to a speaker group
Third specific speaker group voice model data MD1, MD2, M
D3.

As described above, the number of users who use the apparatus is narrowed down to a certain range, and the standard speaker voice model data (first data in this embodiment) is created from voice data of an unspecified number of speakers belonging to the range. To the third specific speaker group voice model data MD1, MD2, MD3), the standard speaker voice is compared with the case where the standard speaker voice model data for an unspecified speaker is assumed for all speakers. The size of the model data can be significantly reduced, the memory size of the ROM 9 can be reduced, and the control unit (CPU)
5 can reduce the processing load.

In addition, since speech recognition is performed using the specific speaker group voice model data (first to third specific speaker group voice model data MD1, MD2, MD3) corresponding to each speaker group, A higher recognition rate can be obtained as compared with the case of using standard speaker voice model data for an unspecified speaker assuming all speakers.

The specific speaker group voice model data to be selected for the speaker speaking at that time can be determined by, for example, preparing a button for designating the speaker group on the operation unit 11 and using the device. If the user is a child, for example, the user may input a voice after operating a button for designating the speaker group C.

As described above, when the operation of designating the speaker group is performed, the control unit 5 judges the operation and performs speech recognition using the corresponding specific speaker group speech model data. For example, when a button for designating the speaker group C is operated, the control unit 5 determines the operation and performs speech recognition using the specific speaker group speech model data MD3.

Separately from this, the pitch cycle information of the input voice (the pitch cycle information obtained when the voice analysis unit 4 analyzes the voice) without the user of the apparatus performing the setting operation of the speaker group. ), The control unit 5 may determine the speaker group of the input voice, and perform the voice recognition using the corresponding specific speaker group voice model data based on the determination result. .

FIG. 2 shows a flowchart for performing the voice recognition process using the voice model data corresponding to the speaker as described above. The voice uttered by the speaker is input (step s1), and there is a voice input. Is determined (step s
2) If there is a voice input, the voice analysis unit 4 performs voice analysis on the input voice (step s3). Then, based on the analysis result, the control unit 5 determines to which speaker group the voice belongs (step s4), and selects voice model data corresponding to the determined speaker group (step s5). , Perform voice recognition processing (step s
6). Then, a predetermined process based on the recognition result is performed (step s7). The predetermined process is various processes to be performed by the device, such as time setting and alarm setting.

Also, in this embodiment, for a specific word, a user using the device speaks the specific word and registers the voice data, and the registered data and the specific speaker described above are registered. It is also possible to perform speaker learning processing using group voice model data. Hereinafter, this will be described.

[0041] In order to perform this speaker learning processing, in this embodiment, as a specific word, "Good morning", "I'm home", "Hello", "Good night", "when",
A user (speaker) utters words belonging to the above word set w1 such as "alarm" (these words are referred to as registered words), and voice data for each registered word obtained by uttering the speaker is obtained. To create registered word data and register it in the RAM 10. At this time, for example,
In the case where the device is used by one family, the family (for example, father, mother, and child) utters the registered word data by voice data obtained by sequentially uttering a plurality of registered words one by one. Create and register each.

Specifically, for example, the registered word data corresponding to each registered word created by the respective voice data for each registered word obtained by uttering the father (supposed to be speaker # 1) is the speaker The registered word data corresponding to each registered word, which is stored as registered word data RD1 for # 1 and is created by voice data for each registered word obtained by uttering the mother (supposed to be speaker # 2), is the speaker The registered word data corresponding to each registered word, which is stored as the registered word data RD2 for # 2 and created by the respective voice data for each registered word obtained by uttering a child (speaker # 3), It is stored as registered word data RD3 for speaker # 3. This registration is performed as follows.

First, a mode setting for registering a registered word is performed by the operation unit 11. Then, the speaker # 1 is designated, and the speaker # 1 utters the above-described registered words sequentially. Then, the registered word data RD1 corresponding to each registered word is created by voice data for each registered word obtained by uttering the speaker # 1. Then, a mapping function for speaker # 1 is created based on the registered word data RD1 and the specific speaker group voice model data (in this case, the specific speaker group voice model data MD1).

Subsequently, the speaker # 2 is designated and the speaker # 2 is designated.
Utters each of the above-mentioned registered words sequentially. Then, the registered word data RD2 corresponding to each registered word is created from the voice data for each registered word obtained by uttering the speaker # 2. Then, this registered word data RD
2 and a specific speaker group voice model data (in this case, specific speaker group voice model data MD2), a mapping function for speaker # 2 is created.

Subsequently, the speaker # 3 is designated.
Speaker # 3 sequentially utters each of the registered words described above. Then, the registered word data RD1 corresponding to each registered word is created by the voice data for each registered word obtained by uttering the speaker # 3. Then, the registered word data RD3 and the specific speaker group voice model data (in this case,
A mapping function for speaker # 3 is created based on the specific speaker group voice model data MD3).

Thus, the registration of the voice data of the speaker to be registered is completed. Then, the respective mapping functions and the first to third specific speaker group codebooks CB1, CB2,
Using CB3 (codebook size is 256 size), input speaker codebooks for speaker # 1, speaker # 2, and speaker # 3 are created. The method of creating the input speaker codebook in this way is called codebook mapping.

In this case, since the speaker # 1 is the father, the speaker # 2 is the mother, and the speaker # 3 is the child, the input speaker codebook for the father, the input speaker codebook for the mother, the child An input speaker codebook is created. In this embodiment, in order to further reduce the data amount of the input speaker codebook, vector quantization is performed using a universal codebook 94 prepared in advance.

The universal code book 94 is a code book obtained from a wide range of speakers including all speakers from adults to children.
Or have a large size such as 2048,
Vector quantization is performed by the universal codebook 94 to create a codebook (256 size) using only the code number of the universal codebook 94.

As described above, the universal code book 9
By using a code book with only the code number of 4, the data amount can be greatly reduced.

By the way, assuming that a RAM of about 32 kilobytes is used as the RAM 10 used in the apparatus, in the above-mentioned input speaker codebook, 256 × 10 (10 is the number of dimensions of the LPC cepstrum coefficient) for one codebook. Therefore, it is necessary to store 2560 coefficients, and when one coefficient is stored in a short type, 2 bytes × 256
A total of about 0 kilobytes is required.

Therefore, this code book is assumed to be 4
If it is made, it alone requires a capacity of about 20 kilobytes, occupies much of the RAM 10, and interferes with other processes.

To cope with this, as described above,
Vector quantization by the universal codebook 94,
An input speaker codebook that is vector-quantized using only the code number of the universal codebook 94 is created. In the following, this vector-quantized input speaker codebook is called a quantization codebook, and the speaker #
1 is a quantized codebook VQ1, a quantized input speaker codebook corresponding to speaker # 2 is a quantized codebook VQ2, and speaker # 3.
Is referred to as a quantized codebook VQ3.

Thus, the quantized input speaker codebook (for example, the quantized input speaker codebook VQ1) for a certain speaker only needs to store 256 code numbers, and each code number is stored in a short type. If it is saved in, the capacity of about 0.5 kilobytes in total of 2 bytes × 256 is enough, so the memory capacity required as a codebook is
1 / compared to the input speaker codebook before quantization described above.
It becomes 10.

The universal code book 94
Is the registered word data RD1 for each input speaker described above.
It can also be used for quantization of RD2 and RD3.

FIG. 3 shows the external configuration of the audio clock device according to this embodiment. As described above, the audio timepiece device according to the present embodiment has a strong toy-like element. Actually, the appearance of the audio timepiece device is a character of a popular animation, but the appearance is directly related to the present invention. Here, the description will be made with a simple shape.

In FIG. 3, the housing 50 of the audio clock device is operated each time to make various settings for the microphone 1, the speaker 8, and the device shown in FIG. An operation unit 11 having various operation buttons (details will be described later) is provided.

The operation unit 11 includes, for example, a power switch SW, speaker group designation buttons BT1, BT2, BT3, speaker designation buttons BT11, BT12, BT13, and a mode changeover switch (for function) for switching between a recognition mode and a registration mode. The affirmation button BT21 that allows the speaker to input affirmative (“yes”) or negative (“no”) by operating the button in response to an inquiry from the MSW or the device side.
And a negative button BT22 (a specific operation example of which will be described later), a volume control switch VSW, and the like.

The speaker group designation buttons BT1, BT2,
BT3 designates a speaker group of a user who uses the apparatus, a speaker group designation button BT1 designates an adult male (speaker group A) as a speaker group, and a speaker group designation button BT2 designates An adult woman (speaker group B) is designated as a speaker group, and a speaker group designation button BT3 is for designating a child (speaker group C) as a speaker group.

Also, speaker designation buttons BT11, BT1
2, the BT 13 designates a user who uses the device as a speaker. Here, three users (speaker #)
1, speaker # 2, and speaker # 3).

The operation unit 11 shown in FIG. 3 is provided with various switches and various operation buttons as described above. These are switches and operation buttons necessary for describing this embodiment. However, it is not necessary to provide all of them, and only necessary switches and operation buttons may be provided as appropriate according to the functions of the device. Further, switches such as the power switch sw may be provided separately from the operation buttons instead of inside the operation unit 11.

The configuration shown in FIG. 3 is an example for explanation, and the microphone 1, the speaker 8,
The arrangement of the operation unit 11 and the arrangement of various switches and operation buttons in the operation unit 11 are not limited thereto.

Here, the registration mode and the recognition mode will be described. Any one of the registration mode and the recognition mode is set by the mode change switch MSW.

In the registration mode, each speaker (in this case, speakers # 1, # 2, and # 3) using the device speaks a plurality of predetermined registration words, and is created from each voice data. Register registered word data Registered word data RD1, RD
2, RD3, and a speaker learning process is performed using the registered word data RD1, RD2, RD3. Finally, a quantization codebook VQ1,
A quantization codebook VQ2 for speaker # 2 and a quantization codebook VQ3 for speaker # 3 are created.

The specific processing in the case of the registration mode will be described. Here, as described above, the registered word data obtained by the father (speaker # 1), the mother (speaker # 2), and the child (speaker # 3) speaking each registered word are registered. Shall be.

First, the speaker designation button BT1 of the operation unit 11
1 for the registration mode of the speaker # 1, the speaker # 1
(Father) utters the above-mentioned registered words sequentially. As a result, registered data RD1 of the registered word for speaker # 1 is created, a mapping function is created, and further, vector quantization is performed by universal codebook 94, and speaker # is determined by the code number of universal codebook 94. One quantized input speaker codebook VQ1 is created.

Subsequently, the speaker designation button BT of the operation unit 11
12 to set the speaker # 2 registration mode to the speaker # 2.
(Mother) utters the above-mentioned registered words sequentially. As a result, registered data RD2 of the registered word for speaker # 2 is created, a mapping function is created, and further, vector quantization is performed by universal codebook 94, and speaker # is determined by the code number of universal codebook 94. A second quantized input speaker codebook VQ2 is created.

Subsequently, the speaker designation button BT13 is operated, and in the registration mode of the speaker # 3, the speaker # 3 (child) sequentially utters the above-described registered words. This allows
Registered data RD3 of the registered word for speaker # 3 is created, a mapping function is created, and vector quantization is performed by universal codebook 94, and speaker # 3 based on the code number of universal codebook 94 is generated.
Of the input speaker codebook VQ3 is generated.

At this time, the registered word data RD1,
RD2 and RD3 are also vector-quantized using this universal codebook 94, and hereinafter, it is assumed that registered word data RD1, RD2 and RD3 are also quantized data.

Next, the recognition mode will be described. In this case, the recognition mode is the first to third specific speaker group voice model data MD1, MD2, MD3, speaker # 1 to speaker #
3 registered word data RD1, RD2, RD3, speaker #
Quantization codebooks VQ1, VQ2 for speakers 1 to # 3
This is a mode in which any one of VQ3 is used for voice recognition depending on the situation.

For example, when the mode changeover switch MSW is set to the recognition mode, first, a speaker group to be used is set. For example, if the speaker to be used is, for example, speaker # 3 (child), a voice is input after operating a speaker group designation button BT3 for designating the speaker group C. As a result, the control unit 5 performs voice recognition using the third specific speaker group voice model data MD3.

As described above, when the speaker group is designated by the speaker to be used, the control unit 5 determines the operation, and
Speech recognition can be performed using the corresponding specific speaker group speech model data.

As described above, the user is narrowed down to a specific speaker group, and the voice recognition is performed using the specific speaker group voice model data created from the voice data of the speaker belonging to the speaker group, so that all the users can be recognized. Compared with the case where speech recognition is performed using large-scale standard speaker speech model data assuming a speaker, speech recognition at a higher recognition rate can be performed. Also, since it is sufficient to have specific speaker group voice model data of only some speaker groups, the scale of the voice model data itself can be significantly reduced, and the memory size of the ROM 9 can be reduced. In addition, the processing load on the control unit (CPU) 5 can be reduced.

The determination of which speaker group the speaker belongs to can be automatically made on the device side based on pitch period information obtained by analyzing the input voice.
There is also a method of determining which speaker group a speaker belongs to by using a similarity obtained by matching input voice data with registered word data. Here, a case where the speaker group is determined based only on the pitch cycle information will be described.

That is, when three speakers, father, mother and child, are considered, the pitch cycle of the child is shortest, the mother is shortest, and the pitch cycle of the father is longest among the three. Is normal, and from the pitch period information, it is possible to specify which of the speaker # 1, the speaker # 2, and the speaker # 3 the input speaker is.

As described above, for example, when it is determined that the input speaker is the speaker # 1, the registered word data RD1 corresponding to the speaker # 1 and the speaker data are created by the speaker learning process. Speech recognition is performed using the quantized input speaker codebook VQ1. In other words, if the word spoken by speaker # 1 is a registered word, the word is recognized by registered word data RD1 corresponding to speaker # 1, and the other words are quantized input speaker codebook VQ1. Is used for speech recognition. If the registered word data has not been created and the speaker learning process has not been performed, speech recognition is performed using the specific speaker group speech model data 92.

In the present invention, as a procedure of a conversation performed between the apparatus and the speaker, first, the speaker speaks one of the above-described registered words. That is, on the device side, first, a conversation procedure that recognizes a registered word belonging to the word set w1 is set. As described above, according to the present invention, the apparatus side knows which word set of a word input scene at present, according to the operation scene at that time. Then, recognition processing of the word input in the scene at that time is performed.

Therefore, assuming that, for example, the speaker # 1 first utters “Alarm” as one of the registered words to the device at the beginning of the conversation, the “Alarm”
Is analyzed, and the device side identifies the speaker based on the pitch period information obtained thereby. In this case, it is determined that the input speaker is the speaker # 1, and the input voice is subjected to voice recognition processing using the registered word data RD1 for the speaker # 1, so that the uttered voice is an "alarm".
It is recognized that it was.

As described above, when the input speaker inputs the registered word, the apparatus determines which speaker the input speaker is and recognizes the registered word. Then, when the voice input thereafter is a word other than the registered word (word other than the word set w1), the voice recognition operation is performed without performing the speaker determination.

For example, the current operation scene on the device side is a scene for recognizing a registered word, and after the recognition process of the registered word is completed in this scene, the word set w2 or w3 is set as the next scene. If it is set so as to move to a scene to be recognized, the recognition of the word set w2 is performed as a speech recognition process by speaker adaptation, and the quantization codebook VQ1 for speaker # 1 and the The speech recognition process is performed by using one specific speaker group codebook CB1 and the first specific speaker group speech model data MD1 (speech model data corresponding to the word set w2).

FIG. 4 is a flowchart for explaining the overall processing procedure of the speech recognition processing enabling the above-described speaker learning processing. The flowchart shown in FIG. 4 shows a processing procedure in a case where the number of speakers is limited to one speaker.

In FIG. 4, first, it is determined whether the mode is the recognition mode or the registration mode (step s11). If the mode is the registration mode, voice input for the registered word spoken by the speaker is performed (step s12), and voice input is performed. Is determined (step s13). If there is a voice input, voice analysis of the input voice is performed (step s14), and the input voice data is registered as registered word data (step s1).
5).

Then, it is determined whether or not the voice data input for the word to be registered has been completed (step s1).
6) If completed, it is determined to which speaker group the voice data for the input voice belongs (step s1).
7), speaker learning processing as described above (creating an input speaker codebook and finally creating a quantization codebook)
Is performed (step s18).

On the other hand, if it is determined in step s11 that the recognition mode is set, the voice input by the speaker is performed (step s19), and it is determined whether or not there is a voice input (step s20). If there is, the voice section of the input voice is detected and voice analysis is performed (step s).
21).

Then, it is determined whether or not the speaker learning process has been performed (step s22). If the speaker learning process has not been performed, the unspecified speaker speech recognition (in this embodiment, the specific speaker group) Speech model data 92 (specific speaker group speech model data prepared corresponding to the input speaker at this time, and if the input speaker is speaker # 1, first specific speaker group speech model data) (Step s2)
3) If the speaker learning processing has been performed, the processing after step s24 is performed.

In step s24, it is determined whether or not the voice data for the input voice is registered word data.
If the input word data is registered word data, the registered word data 101 (registered word data created by the input speaker's voice at this time, and the input speaker is speaker # 1) is used as registration type speech recognition processing. Registered word data R for speaker # 1
Perform voice recognition processing using D1) (step s2)
5).

If the input word is not registered word data, a quantization codebook (based on the input speaker # 1 and the input speaker # 1) created based on the input speaker's voice is used as speaker-adaptive speech recognition processing. If so, the quantization codebook VQ1 for speaker # 1) and the specific speaker group codebook (input speaker is speaker #)
If the input speaker is 1, the first specific speaker group codebook CB1) and the specific speaker group voice model data (first specific speaker group voice model data MD1 if the input speaker is speaker # 1) are used. All voice recognition processing is performed (step s26).

By the way, in the case of this device, the recognizable words are composed of a plurality of word sets such as the word sets w1, w2 and w3, as described above. You know whether the word set is input. For example, the word set w1
After the scene for inputting the (registered word) is over, the scene for inputting the word set w2 (the unit of time such as "1 o'clock" or "2 o'clock") is the current scene. I know what it is.

Therefore, in step s24 described above, it is determined whether or not the voice data corresponding to the input voice is registered word data, depending on what kind of scene is currently at which the registered word is input. If there is, it is determined that the input word data is registered word data, and the corresponding processing is performed.

Then, when the speech recognition processing is completed, a predetermined processing based on the recognition result is performed (step s27).

Although the process shown in FIG. 4 is a process for one specific speaker, for example, as described above, for example, speaker # 1, speaker # 2, and speaker # 3 When there are a plurality of speakers, as shown in the flowchart of FIG. 5, in the registration mode, a speaker is specified before the voice input step s1 (for example, speaker # 1, speaker # 2). Step s28 for specifying the speaker # 3) is added. Therefore, in this case, as shown in FIG. 1, registered word data RD1, RD2, and RD3 for speakers # 1 to # 3 are created as registered word data, and quantization code books are also written for speakers # 1 to # 3. # 3 quantization codebook VQ1,
VQ2 and VQ3 are created.

On the other hand, in the recognition mode, after the speech analysis processing in step s21, the input speaker
A speaker determination process (step s29) for determining which of speaker # 1, speaker # 2, and speaker # 3 is a speaker is added.

In the speech recognition process at this time, in the case of a registered word, the speaker # 1, the speaker # 1 created based on the speech data of the speaker # 1, the speaker # 2, and the speaker # 3, and the speaker # 2, speaker # 3
Voice recognition using any of the registered word data RD1, RD2, and RD3 for the
1, the first specific speaker group codebook CB1 and the first specific speaker group voice model data MD1, the quantized codebook VQ2 for the speaker # 2 and the second
Specific speaker group codebook CB2, second specific speaker group speech model data MD2, quantization codebook VQ3 for speaker # 3, third specific speaker group codebook CB3, and third specific speaker group codebook CB3.
Using any set of the specific speaker group voice model data MD3.

Although the processing in the recognition mode described above is a processing in the case where a registered word is registered, the recognition processing can be similarly performed even when the registered word is not registered. In this case, since the registered word data RD101 is not created, speech recognition is performed using the specific speaker group speech model data 92.

Next, an example of a conversation between a device and a speaker when setting an alarm time will be described with reference to a state transition diagram of FIG. In FIG. 6, the utterance content in parentheses is the content uttered by the speaker using the device. First, the apparatus urges the speaker to utter an example of the uttered content, such as "Please utter the alarm time at 7:00 am, 30 minutes, etc." (step s31). Thus, it is assumed that the speaker (here, speaker # 1) has uttered “AM”, “1:00”, and “20 minutes” as the times to be set (step s32).

The utterance contents here are "morning", "1"
As a set of three words, such as "hour" and "20 minutes", continuous utterances are made with a slight interval between words. Then, the device performs speech recognition continuously for each word. As a result of the recognition,
The device responds as follows: "Is it 1 am, 20 minutes? Answer yes or no?" (Step s33).

If the recognition result is correct, the speaker # 1 responds with "yes" (step s34), and from the apparatus side,
An utterance content such as "alarm time set to 1:20 am" is issued (step s35). On the other hand, if the recognition result is wrong, the speaker # 1 is “No”.
(Step s36), and the apparatus utters a content such as "I will listen in order" (step s37).

Then, the user first utters a content such as "Are you in the morning?" To confirm whether it is "AM" or "PM" (step s38).

On the other hand, the speaker # 1 replies “yes” if it is all right in the morning (step s39), and replies “no” if it is wrong (if not in the morning) (step s40). . In the case of "morning" and "afternoon", since the choice is an alternative, if the response of the speaker # 1 is "No"
If it is, it can be determined with the other word. In other words, the first question asked by the device is "morning"
And if the speaker responds “No”,
"Afternoon" is determined.

[0099] In addition to "am" and "pm", for example, when "noon" is added to determine one of three "am", "pm", and "noon", for example, , The device side is "AM", the speaker side is "NO", the device side is "PM", and the speaker side is "NO" and "noon"
Is determined. As described above, if the number of words is about three, by repeating the "yes" or "no" response from the speaker two or three times, a desired word can be determined in a short time.

After the determination of “am” or “pm” is performed in this way, the unit of “hour” is determined next.
In this case, the word input by speaker # 1 is "1 o'clock".
On the other hand, from the apparatus side, for example, among the first to n-th recognition candidates for "1 o'clock", step s
Assuming that "1 o'clock?" Is uttered as the first candidate at 41, speaker # 1 responds with "Yes" (step s4).
2) "1 o'clock" is determined, and the process enters a minute unit confirmation process (step s51).

However, if, for example, "7 o'clock" is output as the first candidate from the device side in step s41, speaker # 1 replies "no" (step s4).
3). Thereby, the second candidate is output from the device side (step s44). For example, assuming that "8:00" is output as the second candidate, if it is correct, the speaker # 1 determines "Yes". (Step s45),
If wrong, reply "No" (step s4
6).

Here, since it is “No”, the device side further selects “1” as a third candidate, for example, “1”.
Is output (step s47). If it is correct, the speaker # 1 answers "yes" (step s4).
8) If wrong, reply "No" (step s49). Here, since speaker # 1 responds with "yes", "1 o'clock" is determined at that time, and the process proceeds to the minute unit confirmation process (step s51).

Also, if the third place candidate is “No”
In this case, the apparatus prompts the speaker # 1 to input a voice in the unit of "hour" again, for example, "Please say something again." s50). Thus, speaker # 1 again
The utterance of the unit of time (in this case, "1 o'clock") is performed, and the process returns to step 41.

When the speech is input again for the unit of "hour" and it is recognized, the word already denied by "no" (in the above case, "7 o'clock", "8 o'clock", etc.) Is to be excluded from recognition candidates.

Also, the unit of "minute" can be performed by the same processing as steps s31 to s50 described in the unit of "hour". Then, when finally determined, the apparatus utters an utterance such as "alarm time set to 1:20 am" (step s3).
5).

Such processing can be performed not only for the alarm setting but also for the current time setting in the same manner.

When the first candidate is denied by the speaker in the above-described processing, the apparatus does not utter the second candidate, the third candidate,... For example, if "1 o'clock" is denied,
.., "2 o'clock", "3 o'clock",...

By performing such processing, when performing time setting or the like, the speaker can select “am”, “what time”, “what minute”.
Can be continuously uttered as a group of voices, thereby simplifying a voice input operation for time setting. Moreover, if these series of voices are correctly recognized, the time setting ends at that point,
If it is incorrect, the recognition operation will be performed until it is correctly recognized, while checking the correctness of each word, so that finally the correct time can be set, and convenience and certainty Time setting that combines both can be performed.

The processing of recognizing such a continuous word is not limited to the time, but also recognizes a speech composed of a plurality of words and continuously uttered with a slight interval between the words. Widely applicable when you do.

FIG. 7 shows the first to n-th words as a plurality of words.
Is a flowchart for explaining a processing procedure for recognizing a speech composed of n words up to the first word and continuously uttered with a slight interval between the words (herein referred to as continuous word recognition). It is. The processing procedure shown in FIG. 7 will be briefly described. Here, three words (the first word, the first word,
The second word and the third word). Also, these first words, second words
When recognizing the 3rd word and the 3rd word,
As the recognition candidates, m-th recognition candidates are output. Here, it is assumed that m takes a value from 1 to 3.

In FIG. 7, first, recognition is performed on a continuous word composed of a plurality of words spoken by the speaker (step s61). Then, the recognition result is output (step s6).
2) The speaker recognizes "yes" or "no" for the recognition result (step s63). here,
If the response of the speaker is "Yes" (step s64), the recognition result is determined as the whole continuous word (step s6).
5) At that point, the recognition process ends.

On the other hand, if the response of the speaker is “No”,
The first word recognition process starts. In the recognition processing of the first word, first, an m-th candidate (here, m = 1) is output (step s66), and "yes" or "no" of the speaker corresponding thereto is recognized (step s67). ).
Here, if the response of the speaker is “yes” (step s6)
8), the first-place candidate is determined as the first word (step s69), and the process for the second word is started (step s70).

On the other hand, if the response of the speaker is "no",
It is determined whether or not m is a set value (in this case, the set value is m = 3) (step s71). Here, since it is not the set value yet, the process returns to step s66, and the second candidate is output this time as m = 2.

Then, "YES" or "NO" of the speaker corresponding thereto is recognized (step s67). here,
If the speaker's response is "yes" (step s68), the second candidate is determined as the first word (step s6).
9), the process for the second word is started (step s70). If the response from the speaker is "No", the determination in step s71 is performed.

As described above, if the response of the speaker is "Yes", the first word is determined by the recognition candidate at that time, but if the response of the speaker is "No", the next recognition is performed. A similar process is performed for the candidate. Then, in this case, if the third candidate is not correctly recognized,
It requests the speaker to speak the first word again (step s72). Then, the same operation is repeated for the content in which the speaker uttered the first word again.

As described in the example of FIG. 6 described above,
When the speaker utters the first word again and performs speech input thereof, and recognizes it, words that have already been denied “No” are excluded from recognition candidates.

When the recognition process for the first word is completed in this way, the process then proceeds to the recognition process for the second word (step s70), where the recognition process for the second word is performed. Is completed, the process enters a recognition process for the third word (step s73). Note that the recognition processing for the second and third words is the same as the recognition processing for the first word, and a description thereof will be omitted.

When performing the above-described voice input operation, the distance between the speaker and the voice input unit (microphone) on the device side is an important factor from the viewpoint of recognition performance. Therefore, in this embodiment, control is performed to switch the voice input gain of the voice input unit, that is, the voice input gain of the microphone amplifier 2, according to the status of the voice input operation of the speaker. This control is performed by the control unit 5 together with the control of the various processes described above. Hereinafter, this will be described.

The control of the audio input gain (hereinafter simply referred to as gain) of the microphone amplifier 2 is performed by judging what kind of scene the sound input is to be performed, and increasing the gain according to the scene. This is control to make it smaller.

For example, it is determined whether a scene requiring a particularly high recognition rate is input when the apparatus operates, or a scene other than that is input. In particular, in the case where a voice requiring a high recognition rate is input, control is performed such that the voice input gain is reduced as compared with the case where other voices are input.

More specifically, a scene in which a speech is input for a word whose speech feature amount is similar to a plurality of words and is likely to be erroneously recognized, a case where word registration is performed, a case where speaker learning is performed, and the like. In such a situation, control is performed to reduce the voice input gain.

In the case of this apparatus, a voice input for a word that is likely to be erroneously recognized is made by inputting a time (am, what time,
Minutes). For such time setting, for example, there are many voices such as "1 o'clock" and "7 o'clock" which have similar feature amounts of voice and are easily erroneously recognized, and the time is set accurately. Is required.

Therefore, when performing such a voice input, the gain of the microphone amplifier 2 is reduced, and the speaker approaches the microphone and speaks.

On the other hand, in a situation where a call is made to the device, the gain is increased so that a voice can be input from a position far away from the device. This is to make it convenient when a conversation with the device is performed at a distance from the device.

In the scene where the gain is increased, for example, registered words (words belonging to the word set w1) such as “good morning”, “now”, “good night”, “what time”, and “alarm” are input. Such a scene, and such a word input in such a scene is less likely to be erroneously recognized among other words belonging to the same word set w1 as compared with the above-described time information (word set w2 or w1). In this case, the number of words to be recognized is not so large because it is only a few words. In addition, since the content is such that the device can be easily called, it is preferable to use a method in which the gain can be increased and the input can be performed even from a position away from the device.

As described above, an appropriate speech input gain is set depending on what kind of word is inputted.

As described above, the gain of the microphone amplifier 2 is reduced in a case where a voice is input for a word that is easily misrecognized, a case where a word is registered, or a case where speaker learning is performed. , The distance between the mouth of the speaker and the microphone 1 does not change much depending on the speaker, and is almost the same in many cases. As a result, the S / N ratio is improved and the dynamic range is reduced, and the utterance content in this situation is recognized with a high recognition rate.

As described above, in order to control the gain of the microphone amplifier 2 and perform the voice input according to the situation when the voice input is performed, some notification is given from the device side. The notification may be, for example, "speak closer to the microphone", "speak a little louder", or "speak a little lower". Output from the side.

The control unit 5 determines in what situation the current voice input is a voice input (for example, whether the speaker is simply calling the device, whether the registered word is being input, or the time). Setting, etc.) can be grasped, so that the gain of the microphone amplifier 2 is switched and controlled accordingly, and the speaker can speak in accordance with the gain of the microphone amplifier 2. The contents can be output.

For example, in a case where a speech for a word that is likely to be misrecognized is input, a case where a word is registered, a case where speaker learning is performed, etc., the gain of the microphone amplifier 2 is reduced, and the apparatus side outputs “a little more microphone”. Please talk closer to "." As a result, the speaker naturally comes closer to the microphone and speaks, thereby improving the S / N ratio and reducing the dynamic range, thereby improving the recognition performance.

FIG. 8 is a flowchart for explaining an example in which the control for switching the gain of the microphone amplifier 2 is performed in accordance with the situation of the above-described speaker's voice input operation.
An example in which the gain is switched when performing the processing of FIG. 4 will be described.

In FIG. 8, it is determined whether the mode is the recognition mode or the registration mode (step s81). In this determination, if the mode is the registration mode, gain switching is performed so as to reduce the gain of the microphone amplifier 2 (step s82), and voice input is performed using the switched gain (step s83). Step s89 from the voice inputting step s83 to the speaker learning processing is the same as step s18 from the voice inputting step s12 to the speaker learning processing in FIG. 4, so that the description is omitted here.

On the other hand, when the recognition mode is set, the gain of the microphone amplifier 2 is switched according to the scene. For the switching, it is determined whether or not the scene in which the voice input is performed is a scene in which the gain is increased (step s90). If the gain is to be increased, the gain is switched so that the gain of the microphone amplifier 2 is increased. (Step s9
1) When the gain is to be reduced, the gain is switched so that the gain of the microphone amplifier 2 is reduced (step s92).

Note that the scene where the gain is increased in step s90 is a scene where a voice input for making a call is made, as described above.
When the gain is reduced, there is a high possibility that a mistake is made when entering a registered word or when setting an alarm time or current time, such as when entering a time (am, what time, how many minutes). , Especially when trying to obtain a high recognition rate.

As described above, the voice is input by setting the gain according to the scene at that time (step s).
93). The process from the voice input (step s93) to the predetermined process (step s101) is the same as the process from the voice input (step s19) to the predetermined process (step 27) in FIG. 4, and a description thereof will be omitted.

The apparatus receives such a command by voice and performs an operation in accordance with the command. The speaker is recognized as a speaker which is well recognized by the voice quality and the characteristics of the way of speaking. There are often speakers with low rates.

As described above, especially for a speaker having a low recognition rate, the apparatus outputs the utterance content in the form of a question, and the speaker listens to the utterance content. In the case of denial, by operating a button that means negative, it is possible to perform an operation equivalent to recognizing “yes” or “no”.

That is, for the information that the speaker wants to set for the device (called setting information), the device side sequentially outputs candidates as setting information in the form of a question, one by one.
Whenever one of the candidates is output, a button input of either a button indicating affirmation or a button indicating negation is received for the output candidate, and when the button input indicating affirmation is received, the candidate is set to the above-described setting. When the information is determined and a button input meaning negative is received, the next candidate is output.

The reason why the affirmative or negative input can be performed by the button operation is to assist a speaker with a poor recognition rate, and is referred to as a help mode here. Hereinafter, the help mode will be described.

An affirmation button BT21 meaning affirmation and a negation button BT2 meaning negation are provided on the operation unit 11 of the apparatus.
2 is provided. That is, the affirmation button BT21 depends on the speaker.
Is pressed, the control unit 5 determines that the answer is affirmative, and when the negative button BT22 is pressed, the control unit 5 determines that the answer is negative.

To set the help mode, in the ordinary recognition mode described so far, for example,
When the affirmation button BT21 and the negation button BT22 are pressed simultaneously for several seconds (about 2 seconds) or more, the control unit 5 detects this and determines that the help mode has been set, and shifts to the help mode. Note that the transition to the help mode is not limited to the condition that the affirmation button BT21 and the negation button BT22 are simultaneously pressed for several seconds or more, but may be another method.

Thus, when the help mode is entered,
The apparatus does not perform the voice recognition operation, detects that the speaker has pressed either the affirmation button BT21 or the denial button BT22, and proceeds with the processing accordingly.

FIG. 9 is a flowchart for explaining the recognition processing when the help mode is provided. In FIG. 9, M1 is a processing procedure in a normal recognition mode, and M2 is a processing procedure.
Is a processing procedure in the help mode. The transition from the normal recognition mode M1 to the help mode M2 is a button interruption in which the positive button BT21 and the negative button BT22 are simultaneously pressed for several seconds (about two seconds) or more (step s110).
Transition by The processing from the voice input (step s121) to the predetermined processing (step s127) in the normal recognition mode M1 shown in FIG. 9 is the same as the processing from steps s1 to s7 shown in FIG. Then, the description is omitted.

Now, it is assumed that recognition is being performed in the normal recognition mode, and an interrupt by the affirmation button BT21 and the denial button BT22 has occurred in this normal recognition mode (step s110).

Thus, it is determined whether or not the help mode has been set (step s111). If the help mode has been set, the speaker is notified of the help mode (step s112). Whether or not the help mode is set in step s112 is determined here based on whether or not the affirmation button BT21 and the denial button BT22 are simultaneously pressed for several seconds (about two seconds).

When the mode is shifted to the help mode M2 in this manner, the input button BT21 or the negative button BT22 is in a state of waiting for input (step s113).
It is determined whether either the affirmation button BT21 or the denial button BT22 has been pressed (step s114), and the pressed button is determined to be the affirmation button BT21 or the denial button B
It is determined whether the current state is any of T22 or how long the pressed state is continued (step s).
115), and perform a predetermined process for the process (step s).
116).

For example, each candidate is output from the device side in the form of a question for asking the speaker whether the speaker wants information to be set for the device, and each time one candidate is output. When the button input of either the affirmation button BT21 or the denial button BT22 is received for the output candidate, and when the input of the affirmation button BT21 is received, the candidate is determined as the setting information, and the input of the denial button BT22 is received. , The next candidate is output. The specific processing content will be described later.

Then, it is determined whether or not the help mode is completed (step s117). If not completed, the process returns to step s113. If completed, the process returns to the normal recognition mode M1.

A specific operation example will be described below with reference to the state transition diagram of FIG. In FIG. 10, the operations in parentheses are operations performed by the speaker.

The apparatus first notifies the speaker that the help mode has been entered, such as "Help mode has been entered" (step s131). Next, the speaker utters information that the speaker wants to set for the device, such as "Is the current time set?" (Step s132). When the speaker affirms the utterance content, the affirmation button BT21 is pressed (step s133), whereby the current time setting mode is entered (step s13).
4).

If the current time is not set, the negative button BT22 is pressed (step s135). This allows
The device utters a content such as "Is alarm time set?" (Step s136). When the speaker affirms the utterance content, the affirmative button BT21 is pressed (step s137), whereby the alarm time setting mode is entered (step s138). If the alarm time has not been set, the negation button BT22 is pressed (step s139).

As described above, for each question from the device side, instead of responding “yes” or “no”, the speaker does not respond to the affirmation button BT21 or the denial button BT2.
Press 2. Thus, the device proceeds with the process according to the button operation. The contents of the question are not only the above two, but also, for example, "Do you want to input a registered word?", But these processes are not shown in FIG.

The affirmation button BT21 and the denial button B
When T22 is operated by mistake, the affirmation button BT21 and the negation button BT22 are simultaneously operated so that the user can return to the previous question.

The current time setting mode (step s
134) will be briefly described. Affirmation button BT21
Or the current time setting using the negative button BT22 is
First, the device asks whether it is "morning". On the other hand, when the speaker presses the affirmation button BT21, the morning is determined, and when the negation button BT22 is pressed, “afternoon” is determined.

After the morning or afternoon is determined, the apparatus first asks "Is it one o'clock?". If the time unit of the time that the speaker wants to set is "1:00", the affirmation button is pressed. B
Press T21, thereby confirming 1:00. Also, 1
If not, press the negative button BT22. This allows
The device asks “2 o'clock?”, And if the time unit of the time that the speaker wants to set is “2 o'clock”, the affirmation button BT
Press 21 to confirm 2 o'clock. If it is not 2 o'clock, the negative button BT22 is pressed.

Such a procedure is performed until the time unit is determined. Then, once the time unit is determined,
Start processing in minutes.

The operation in the unit of minute is basically the same as the operation in the unit described above, and the description thereof is omitted here. However, as for the unit of minute, it is troublesome to operate all of the minutes starting from 1 minute up to 59 minutes every minute.
It will take a long time. In particular, if the time is toward the end of the minute unit (for example, 59 minutes) in the minute unit desired by the speaker, much time is required.

Therefore, in the setting mode in units of minutes, if the negative button BT22 is kept pressed, the apparatus will send, for example, "10 minutes", (after one second),
"20 minutes", (after 1 second), "30 minutes", etc., voice is output every 10 seconds in units of 1 minute, and "50 minutes" is spoken from the device side. When the user stops operating the negation button BT22, the operation returns to the normal operation based on 50 minutes. When the device utters "51 minutes", the user presses the negation button BT22 and then utters "52 minutes". By performing the operation of pressing the BT22, “59”
The setting method may be such that the affirmation button BT21 is pressed with "minute" to set 59 minutes.

As described above, the time setting may take a long time in some cases. In order to shorten the time required for the setting even a little, the affirmation button BT21 and the negation button BT21 are output during the sound output from the apparatus side. The input of the BT 22 is received. In addition, in order to improve the usability, the setting is made so that the function corresponding to the pressing method of the affirmation button BT21 and the denial button BT22 (continuous pressing, simultaneous pressing, etc.) is performed, and a description of how to use the sound from the device side. Can be performed.

As described above, when the speaker utters the information (setting information) to be set to the apparatus and the setting information is not correctly recognized on the apparatus side, the help mode can be set. The candidates as the setting information are sequentially output one by one from the device side in the form of a question. When an input of the affirmation button BT21 is received each time one candidate is output, the candidate is determined as the setting information. However, when the input of the negative button BT22 is received, the next candidate is output, so that the information that the speaker wants to set for the device can be set by button operation.

As a result, even if the voice is not well recognized due to the voice quality or the way of speaking of each speaker, it is possible to set some information such as time setting sufficiently. Further, even a person who has difficulty speaking can easily set the time and the like.

By the way, in the audio clock device according to this embodiment described above, when the current time reaches the alarm time set by the speaker, the current time is transmitted from the device until the speaker stops ringing. Keep output.
For example, suppose that the set alarm is at 1:20 am, and when the time reaches that time, the device utters “1:20 am”, and after one minute, “1 am” 21 minutes. "
The current time is spoken in one minute intervals, such as "1:22 am." This is continued until the speaker performs the silencing operation. In addition, it is not necessary to be in 1 minute increments.
The current time may be spoken every second.

Further, when this alarm is issued, it is possible to output not only the time but also a word such as "Please go early" or a combination of the word and music.

[0164] Also, as a sound stop operation from the speaker,
There is a method in which the speaker utters some sound after the utterance at the time as described above from the device, a method in which a button is operated, and the like.

In order for the speaker to stop the sound by emitting some sound, the sound is stopped when the level of the sound input by the speaker is equal to or higher than a certain level and the level equal to or higher than the certain level is a certain length. . In this case, no voice recognition is performed. Alternatively, it is also possible that the speaker speaks a specific word determined in advance and stops the sound only when the device recognizes it. Further, when the sound is stopped by a button operation instead of inputting a voice, a dedicated button can be provided, but the above-described positive button BT21 and negative button BT22 can be used instead.

Further, when a volume control switch VSW is provided, the sound can be stopped by turning off the volume control switch VSW.

The volume adjustment switch VSW is a switch for adjusting the volume of the output audio from the apparatus side. The audio output amplifier 7 is controlled by adjusting the volume adjustment switch VSW by the operation unit 11, and the output audio is output. Size can be set. As a specific example of the volume adjustment, for example, the output sound can be switched and set to “loud”, “medium”, “small”, and furthermore, “off” to make no sound output. You can also be.

When the device having such a voice recognition function is used in a home, the voice clock reacts one by one due to voices emitted from a television, conversation between humans in the home, and the like, and the voice is carelessly output. It plays a role in preventing emission. When the volume control switch VSW is turned off, in order to reduce power consumption, power supply to peripheral circuits such as an amplifier is stopped, and the control unit (CPU) 5 is put into a sleep state. Then, when the volume control switch VSW is turned on, the CPU is operated by a CPU interrupt to supply power to peripheral circuits.

Further, in the embodiment of the present invention, a plurality of response contents to the voice input by the speaker are prepared, and the response contents can be changed at random and output. This is to prevent the user from getting bored. For example, a plurality of response contents from the device side when the speaker inputs "good morning" are prepared, and among the plurality of response contents to "good morning", Randomly select and output the response content.

As a result, the user is less tired of using the apparatus for a long time.

The present invention is not limited to the embodiments described above, but can be variously modified without departing from the gist of the present invention. For example, the audio timepiece described in the above-described embodiment may have various functions as described in the above-described embodiment, but all of these functions are necessary. is not. Therefore, various switches and operation buttons provided on the operation unit 11 need only be provided with buttons necessary for operating their functions.

Further, the contents of various processes described in the above embodiments are not limited to the contents described in the embodiments. For example, as the speaker learning processing, in the above-described embodiment, the quantization codebooks VQ1, VQ2, and VQ3 are created using the universal codebook 94,
At the time of recognition, the speaker adaptation is performed using the quantization codebook. However, the present invention is not limited to this.
, RD2, RD3 and the first to third specific speaker group speaker model data MD1, MD2. Speaker adaptation can also be performed using a mapping function created from MD3, and this mapping function and first to third specific speaker group codebooks CB1, CB
2, it is also possible to perform speaker adaptation using an input speaker codebook created by CB3 (a codebook before vector quantization using the universal codebook 94).

Further, in the above-described embodiment, since it is assumed that speakers (speakers # 1, # 2, and # 3) belonging to a plurality of speaker groups use one device, The speaker group voice model data also corresponds to the first to third specific speaker group voice model data MD1, MD2. Although the MD3 is provided, depending on the type of the device, the user may be able to target only a speaker (for example, a child) belonging to one speaker group. In such a case, the specific speaker group voice is used. The model data need only have one specific speaker group voice model data corresponding thereto, and according to this, the capacity of the ROM 9 can be further reduced, and the processing capability of the control unit 5 can be small. In addition, the capacity of the RAM 10 may be small, and the cost of the entire apparatus can be significantly reduced.

Further, in the above-described embodiment, an example has been described in which various modes are set by using various buttons provided on the operation unit. However, it is also possible that all buttons can be set without providing any buttons. it can.

The processing program for performing the processing of the present invention described above can be recorded on a recording medium such as a floppy disk, an optical disk, or a hard disk. The present invention also includes the recording medium. Also,
The processing program may be obtained from a network.

[0176]

As described above, according to the present invention, the speaker to be recognized is narrowed down to a predetermined range based on age, gender, etc., and the specific speaker is determined from the voice data of the unspecified plural speakers belonging to the range. Group voice model data is created and stored, and the voice input by the recognition target speaker is recognized using the specific speaker group voice model data. As a result, since only the specific speaker group voice model data corresponding to a certain range of speaker groups needs to be held, compared with the conventional standard speaker voice model data created so as to be able to support all the range of conventional speakers, , The data amount can be greatly reduced. Therefore, the storage capacity of the memory (ROM) for storing the specific speaker group voice model data can be small, and the load of the recognition processing on the CPU can be reduced, and the memory capacity and the processing capacity of the CPU can be reduced. It can be said that this is a speech recognition technology suitable for inexpensive products with great restrictions. In addition, since the data is specific speaker group voice model data corresponding to a specific speaker group, recognition performance is greatly improved.

Further, according to the present invention, the specific speaker group voice model data corresponding to the speaker group in a certain specific range is
Several can also be prepared. For example, for adult men,
It is also possible to prepare corresponding to several speaker groups such as for adult women and children. According to this, it is possible to cope with a case where one device is used by a family. in this way,
Even if we have specific speaker group voice model data corresponding to several speaker groups, the size of the voice model data is larger than having standard speaker voice model data created to support speakers in all ranges. Can be reduced, and the speech model data corresponds to each speaker group, so that the recognition performance is greatly improved.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating an embodiment of a speech recognition device of the present invention.

FIG. 2 is a flowchart illustrating an example in which a speech recognition process is performed using speech model data (specific speaker group speech model data) corresponding to a speaker.

FIG. 3 is a diagram schematically illustrating an appearance of an audio clock device according to an embodiment of the invention.

FIG. 4 is a flowchart illustrating an overall processing procedure of a speech recognition process enabling a speaker learning process, and is a flowchart illustrating a processing procedure for a specific one speaker.

FIG. 5 is a flowchart illustrating an overall processing procedure of a speech recognition process enabling a speaker learning process, and is a flowchart illustrating a processing procedure for a plurality of speakers.

FIG. 6 is a state transition diagram illustrating recognition processing of three consecutive words (“am”, “1:00”, and “20 minutes”) according to the embodiment of the present invention.

FIG. 7 is a flowchart illustrating a processing procedure when the three-word recognition process illustrated in FIG. 7 is applied as a general continuous word recognition process.

FIG. 8 is a flowchart for explaining a processing procedure including gain switching of a microphone amplifier according to a situation of a voice input operation of a speaker.

FIG. 9 is a flowchart illustrating a speech recognition processing procedure including a help mode.

FIG. 10 is a state transition diagram illustrating a help mode (“Yes” or “No” input mode using a positive button and a negative button) according to the embodiment of the present invention with respect to the voice recognition processing.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 Microphone 2 Audio input amplifier (microphone amplifier) 3 A / D conversion part 4 Audio analysis part 5 Control part (CPU) 6 D / A conversion part 7 Audio output amplifier 8 Speaker 9 ROM 10 RAM 11 Operation part 12 Timing means 91 Output speech data 92 Specific speaker group speech model data 93 Specific speaker group codebook 94 Universal codebook 95 Standard speaker speech model data 93 Standard speaker codebook 101 Registered word data MD1, MD2, MD3 1st to 1st 3 specific speaker group voice model data CB1, CB2, CB3 First to third specific speaker group codebooks RD1, RD2, RD3 Registered word data VQ1, VQ2, VQ3 for speaker # 1 to speaker # 3 Quantization codebook for speaker # 1 to speaker # 3

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Hiroshi Hasegawa 3-3-5 Yamato, Suwa City, Nagano Prefecture Inside Seiko Epson Corporation (72) Inventor Masahisa Ikejiri 3-5-5 Yamato, Suwa City, Nagano Prefecture Seiko Epson Co., Ltd. F term (reference) 5D015 AA02 GG01

Claims (4)

[Claims] 1. A speech recognition method in which a plurality of unspecified speakers can be recognized, and a plurality of predetermined words can be recognized. Based on the specified speaker range, the specific speaker group voice model data is created from the voice data of the unspecified multiple speakers belonging to the range, and the specific speaker group voice model data is used for the recognition target speaker. A speech recognition method characterized by recognizing input speech. 2. The speaker to be recognized is composed of a plurality of speaker groups based on the characteristics of speech, and speech data corresponding to the plurality of speaker groups is obtained from speech data of unspecified plural speakers belonging to each speaker group. 2. The speech recognition method according to claim 1, wherein the specific speaker group speech model data is created. 3. A speech recognition apparatus for an unspecified plural speaker and capable of recognizing a plurality of predetermined words, wherein at least a speech analyzing means for analyzing a speech obtained by uttering the speaker. A specific speaker group created from voice data of unspecified multiple speakers belonging to the range by narrowing down the recognition target speakers among the unspecified multiple speakers to a preset range based on age, gender, etc. A voice recognition device comprising: speaker voice model data; and control means for recognizing a voice input by the recognition target speaker using the specific speaker group voice model data. 4. The speaker to be recognized is composed of a plurality of speaker groups based on voice characteristics, and the specific speaker group voice model data is the voices of unspecified multiple speakers belonging to each speaker group. 4. The speech recognition apparatus according to claim 3, wherein the speech recognition apparatus comprises specific speaker group voice model data corresponding to the plurality of speaker groups created from data.