JP2016143050A - Voice recognition device and voice recognition method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a technology that can recognize a driver's voice with high accuracy.SOLUTION: The voice recognition method includes storing recognition dictionary data that correspond to driver's condition in advance and performing voice recognition by detecting the driver's condition and using the recognition dictionary data that correspond to the driver's condition. Concerning driver's voice recognition, the driver's condition is thought to affect the recognition accuracy more greatly than an individual difference between the drivers. Therefore, the improvement of the voice recognition accuracy is made possible by performing voice recognition through by using the recognition dictionary data that correspond to the driver's condition.SELECTED DRAWING: Figure 2

Description

  The present invention relates to a technique for recognizing a voice generated by a driver when applied to a vehicle.

Today's vehicles are equipped with various functions for the purpose of reducing the driving load on the driver, enabling comfortable driving, or ensuring safe driving. Along with this, there is a tendency that the driver needs to operate or set the vehicle, and in order to reduce the burden on the driver in this aspect, the driver can easily operate or perform various matters. A technology that enables setting is also proposed.
As one such technique, a technique for recognizing a driver's voice has been developed. If the driver's voice can be recognized on the vehicle side, the driver can operate or set the vehicle by speaking.

  In speech recognition, speech is recognized based on the following principle. First, an audio signal detected by a microphone is divided into audio signals having a short predetermined time called “phonemes”, and a plurality of feature amounts are extracted by analyzing each phoneme by a predetermined method. Based on the plurality of feature quantities obtained in this way, the sound represented by the phoneme can be specified. For example, in Japanese, there are 5 types of vowels and 9 types of consonants, and it is said that there are about 40 types of sounds including their variations. , The combination of feature amounts is different. Therefore, a combination of these feature amounts is stored as identification dictionary data. When a combination of feature amounts is obtained from a certain phoneme, the vowel or consonant represented by the phoneme can be specified by selecting the closest combination of feature amounts from the identification dictionary data. By performing such processing on the entire audio signal detected by the microphone, the audio can be recognized.

  However, in order for the driver to operate or set various vehicle functions using voice, it is important that the voice can be recognized with sufficient accuracy. Therefore, it has been proposed to improve recognition accuracy by causing a speaker to read a preset content (such as a word or a sentence) and learning the voice of the speaker (Patent Document 1, etc.). Since the combination of feature values (identification dictionary data) obtained by analyzing vowels and consonants varies little by little depending on the speaker, the story to be recognized by speech rather than using the identification dictionary data obtained from a standard speaker It is considered that recognition accuracy can be improved by using identification dictionary data obtained from a person.

JP 2003-162292 A

  However, for the purpose of the driver operating or setting various functions using voice to the vehicle, the recognition accuracy may be insufficient even if the driver's voice is learned, and the recognition accuracy is further improved. There was a demand for the development of technology that could

  An object of the present invention is to provide a technology capable of recognizing a driver's voice with high recognition accuracy.

In order to solve the above-described problems, the speech recognition apparatus and the speech recognition method of the present invention preliminarily store identification dictionary data corresponding to the driver state. In voice recognition, the driver state is detected, and voice recognition is performed using identification dictionary data corresponding to the driver state.
Although the detailed reason will be described later, regarding the driver's voice recognition, it is considered that the driver state has a greater influence on the recognition accuracy than the influence of the individual difference of the driver. Therefore, the recognition accuracy of voice recognition can be improved by performing voice recognition using the identification dictionary data corresponding to the driver state.

It is explanatory drawing which shows the vehicle 1 carrying the speech recognition apparatus 100 of a present Example. 2 is a block diagram showing a rough internal structure of the speech recognition apparatus 100. FIG. It is explanatory drawing about the outline | summary of voice recognition. It is explanatory drawing which showed notionally the mode that speech recognition was performed using identification dictionary data. It is explanatory drawing which illustrated the result of voice recognition. It is a flowchart of the speech recognition process of a present Example. It is explanatory drawing which illustrated the driver | operator state. It is explanatory drawing which illustrated the identification dictionary data set according to the driver | operator state. It is a block diagram which shows the rough internal structure of the speech recognition apparatus 200 of a modification. It is a flowchart of the process which the voice recognition apparatus 200 of a modification produces | generates the identification dictionary data according to a driver | operator. It is explanatory drawing which showed a mode that identification dictionary data were produced | generated according to a driver | operator.

Hereinafter, examples will be described in order to clarify the contents of the present invention described above.
A. Device configuration :
FIG. 1 shows a vehicle 1 equipped with a voice recognition device 100. As shown in the figure, the vehicle 1 is equipped with a microphone 10 that acquires a voice signal of a voice uttered by the driver, and a voice recognition device 100 that recognizes the voice by analyzing the voice signal acquired by the microphone 10. The vehicle 1 is also equipped with an in-vehicle camera 20 that captures a driver's face image, and the voice recognition device 100 analyzes the face image obtained by the in-vehicle camera 20 to thereby detect the driver state (for example, , Tired state, sleepy state, excited state).
In this embodiment, it is assumed that the driver state is detected by taking a face image with the in-vehicle camera 20. This is because the technology for analyzing the driver's face image to detect the driver state has a sufficient track record as a technology capable of detecting various driver states with high accuracy. However, as long as the driver's state can be detected, it is not limited to photographing a face image with the in-vehicle camera 20. For example, the driver's biological information (heart rate, blood pressure, perspiration, etc.) may be detected to detect the driver state, or the driver's driving behavior (handle operation, accelerator operation, brake operation) May be detected to detect the driver state.

FIG. 2 shows a rough internal structure of the speech recognition apparatus 100 of the present embodiment. As illustrated, the speech recognition apparatus 100 includes a driver state detection unit 101, an identification dictionary data selection unit 102, an identification dictionary data storage unit 103, a speech acquisition unit 104, and a speech recognition unit 105.
The voice recognition apparatus 100 is realized by a microcomputer in which a memory, a timer, an input / output peripheral device, and the like are connected to each other via a bus so that data can be communicated with each other, with a CPU at the center. Accordingly, these “parts” are merely classifications of the internal structure of the speech recognition apparatus 100 for the sake of convenience, focusing on the function of the speech recognition apparatus 100 recognizing speech. It does not indicate that it is physically divided into parts. Therefore, these “units” can be realized as a computer program executed by the CPU, can be realized as an electronic circuit including an LSI or a memory, and further realized by combining them. You can also.

The driver state detection unit 101 detects the driver state by analyzing the driver's face image captured by the in-vehicle camera 20. The contents detected by the driver state detection unit 101 as the driver state will be described later.
When the identification dictionary data selection unit 102 receives the driver status from the driver status detection unit 101, the identification dictionary data selection unit 102 selects identification dictionary data corresponding to the driver status from the identification dictionary data stored in the identification dictionary data storage unit 103. select. The identification dictionary data storage unit 103 stores a plurality of types of identification dictionary data described later according to the driver state.
The voice acquisition unit 104 acquires the driver's voice signal detected by the microphone 10.
The voice recognition unit 105 performs voice recognition using the voice signal acquired by the voice acquisition unit 104 while referring to the identification dictionary data selected by the identification dictionary data selection unit 102.
Thus, in the speech recognition apparatus 100 of the present embodiment, the recognition accuracy is improved by referring to the identification dictionary data corresponding to the driver state and performing speech recognition. The reason for this will be described below. As a preparation for this, an outline of speech recognition will be briefly described.

  Roughly speaking, speech recognition is performed through three steps. First, in the first step, a speech signal is divided into a plurality of phonemes, and feature quantities are extracted for each phoneme. In the subsequent second step, the sound is specified for each phoneme by comparing the extracted feature quantity with the identification dictionary data. In the third step, the phonemes whose sounds are specified are combined and recognized as speech.

FIG. 3 conceptually shows a state in which a speech signal is divided into a plurality of phonemes and feature amounts are extracted for each phoneme. For example, if an audio signal as shown in FIG. 3A is obtained, the audio signal is divided into a plurality of phonemes having a predetermined time (for example, 25 msec) width. FIG. 3B illustrates a state in which the audio signal is divided into phonemes (1) to (16).
In FIG. 3C, some phonemes are enlarged and displayed. For example, the phoneme of (2) has a clearly different waveform from the phonemes of (8), (10), and (16). The phoneme of (16) has a clearly different waveform from the phonemes of (2), (8), and (10). On the other hand, the phoneme of (8) and the phoneme of (10) have a feature common to waveforms. Actually, the phoneme of (8) and the phoneme of (10) are recognized as the same sound by humans, and the phoneme of (2) and (16) are recognized as different sounds from other phonemes. Therefore, for all phonemes (1) to (16), a feature amount is extracted by analyzing a waveform included in the phoneme. The feature value extracted at this time is not a feature value related to the volume or pitch of the sound, but a feature value related to the type of sound identified by humans (that is, whether it is a vowel or a consonant, what kind of vowel or consonant) And As such a feature quantity, a plurality of feature quantities represented by formant frequencies are known. Generally, about 10 types of feature quantities are extracted from each phoneme.

Subsequently, the type of sound represented by the phoneme is specified by collating a plurality of feature amounts extracted from the phoneme with the identification dictionary data. Here, the identification dictionary data is as follows.
For example, it is assumed that a lot of voices “A” are collected and a plurality of feature amounts are extracted. If the number of collected voices is 100, 100 sets of a plurality of feature amounts are obtained. These 100 sets of feature quantities should be roughly similar values even if they do not exactly match. For example, when five types of feature quantities (feature quantity 1 to feature quantity 5) are extracted from one phoneme, and a set of these feature quantities is expressed as a coordinate point of a five-dimensional coordinate, it is obtained from the voice “a”. Phoneme coordinate points should be collected and distributed.
When the same operation is performed on the voice of “I”, the coordinate points of the phoneme obtained from the voice of “I” should be gathered and distributed at positions different from the coordinate points of the phoneme of “A”. Similarly, for the voice of “U”, the voice of “E”, and the voice of “O”, the coordinate points of each phoneme are gathered and distributed at the respective positions. FIG. 4 conceptually shows how phonemes are collected and distributed at different positions in a coordinate space (referred to as feature amount space) having a plurality of feature amounts as axes.

Next, when a feature amount is extracted from a phoneme, the coordinate point of the phoneme is point A in FIG. In this case, it may be considered that the sound represented by the phoneme is “A”. Similarly, if the coordinate point of the feature amount extracted from another phoneme is point B in FIG. 4, the sound represented by that phoneme is considered “e”.
In this way, the distribution range in the feature amount space is examined and stored in advance for each vowel and consonant. Then, for unknown phonemes, by determining in which distribution range the feature quantity extracted from the phoneme is included, the type of sound (which vowel or consonant corresponds to, or which Is not applicable). The identification dictionary data is a set of data describing the distribution range in the feature amount space for each vowel and consonant.

As described above, a plurality of feature amounts extracted from phonemes are collated with the identification dictionary data. In this way, the type of sound can be specified for each of the plurality of phonemes shown in FIG. 3, as shown in FIG. For example, the phonemes (1) to (3) represent the consonant “J”, the phonemes (4) to (6) represent the vowel “I”, and (7) to (10) represent the consonant. As said to represent “B”, the type of sound can be specified for each phoneme.
Next, a process for collecting a plurality of phonemes is started. This is because a phoneme is obtained by mechanically dividing an audio signal with a certain time width (for example, 25 msec), and the time width is too short to be recognized as speech. In the example shown in FIG. 5, the phonemes (1) to (3) are grouped into a consonant “J”, and the phonemes (4) to (6) are grouped into a vowel “I”. Similarly, the phonemes of (7) to (10) are “B” of consonants, the phonemes of (11) to (13) are “u” of vowels, and the phonemes of (14) to (16) are “ N ”.
By combining these, voice recognition of “ji”, “bu”, and “n” is performed.

As is clear from the above description, since the accuracy of speech recognition greatly depends on the identification dictionary data to be used, it is important to use appropriate identification dictionary data. Further, as is apparent from the fact that there are individual differences in human voices, there are also individual differences in the identification dictionary data. For example, different identification dictionary data is used for a person who speaks clearly with his / her mouth wide open and a person who speaks without opening his / her mouth. For this reason, the identification dictionary data is set so as to obtain a certain degree of recognition accuracy without being erroneously recognized by any person. This means that recognition accuracy is sacrificed in order to avoid misrecognition. Therefore, if a speaker is identified and dedicated identification dictionary data is used, the recognition accuracy of speech recognition can be improved. Conceivable.
However, even when a speaker is specified and dedicated identification dictionary data is used, the recognition accuracy may not be sufficient for the driver to operate or set the vehicle using voice. . Therefore, the speech recognition apparatus 100 according to the present embodiment improves the recognition accuracy by considering the driver state, as will be described below.

B. Speech recognition processing:
FIG. 6 shows a flowchart of speech recognition processing performed by the speech recognition apparatus 100 of the present embodiment.
As shown in the drawing, in the voice recognition process, first, the driver state is detected (S101). FIG. 7 illustrates the driver state to be detected. For example, state A is a normal state, state B is tired, and state C is sleepy. State D is an impatient state, state E is an excited state, state F is a happy state, and state G is a sad state. These driver states are detected by analyzing the driver's face image taken by the in-vehicle camera 20. Of course, the driver's state may be detected by detecting the driver's biological information and driving behavior.

Subsequently, it is determined whether or not the driver state has changed since the previous detection (S102). For example, if the previously detected driver state is state A and the current driver state is also state B, it is determined that the driver state has changed (S102: yes). In this case, identification dictionary data corresponding to the current driver state is read (S103).
As described above, the identification dictionary data is data describing a distribution range of feature amounts for each type of sound (vowels and consonants to be identified). In a memory (not shown) of the speech recognition apparatus 100, identification dictionary data corresponding to each of the states A to G is stored.

In FIG. 8, the distribution range of the feature amount set in the identification dictionary data corresponding to the state A and the identification dictionary data corresponding to the state B are set for five types of vowels “A” to “O”. The distribution range of the feature amount is shown. Thus, the identification dictionary data changes depending on the driver state. This is due to the following reason.
First, when a person speaks, the type of sound is changed by changing how the mouth is opened, the shape of the mouth, and the strength to exhale. Therefore, the identification dictionary data used to specify the type of sound reflects differences in how to open the mouth, the shape in the mouth, the strength to exhale. Since there are individual differences in the size of the mouth, the shape in the mouth, the strength to exhale, etc., individual differences also appear in the identification dictionary data. For this reason, the recognition accuracy of voice recognition can be improved by using dedicated identification dictionary data depending on the speaker.

  However, how to open the mouth, the shape in the mouth, the strength to exhale, etc. vary greatly depending on the state of the speaker. For example, when you are tired and you are excited, you can immediately understand that the way you open your mouth and the strength of your breath differ greatly. Even in the shape of the mouth, when the user is excited, the back part of the palate tends to be pulled up, and when tired, it tends to hang down. Therefore, the identification dictionary data varies greatly depending on the state of the speaker.

Here, it can be intuitively understood that the recognition accuracy is improved by using different identification dictionary data for the case where the speaker is Mr. A and the case where Mr. B is used. However, even if you are tired or excited, Mr. A is Mr. A and Mr. B is Mr. B, so there is no need to use different identification dictionary data depending on whether you are tired or excited. , Can not be considered in common sense.
However, as described above, the reason why the recognition accuracy can be improved by using different identification dictionary data depending on whether the speaker is Mr. A or Mr. B. This is because the method of opening, the shape of the mouth, the strength to exhale, etc. are different. And from the point that the way of opening the mouth, the shape in the mouth, the strength to exhale, etc. are different, there is no significant difference between the speaker and the speaker.
In addition, since the speaker in the present embodiment is a driver, children and elderly people will not become speakers, and the range of speakers that are subject to speech recognition is narrower than in the case of general speech recognition. It is done. Furthermore, the driver who is driving can take various states both physically and psychologically. This means that there is a higher possibility of misrecognition due to a different driver state than a possibility of misrecognition due to a different speaker. In other words, when the speaker is a driver, the recognition accuracy of voice recognition is better when the identification dictionary data is used differently depending on the state of the speaker (driver state) than when the identification dictionary data is used differently by the speaker. It is thought that it is possible to improve.

In the voice recognition apparatus 100 of the present embodiment, identification dictionary data is stored for each driver state of the state A to the state G shown in FIG. Therefore, in the voice recognition process shown in FIG. 6, when it is determined that the driver state has changed (S102: yes), the identification dictionary data corresponding to the new driver state is read from the memory (S103). .
On the other hand, when the driver state has not changed (S102: no), the identification dictionary data that has already been read can be used continuously, and therefore the process of reading the identification dictionary data (S103) is omitted.

Subsequently, it is determined whether or not the driver speaks (S104). When the driver speaks, the microphone 10 detects the voice. Therefore, it can be immediately determined from the output of the microphone 10 whether the driver speaks.
As a result, if the driver does not speak (S104: no), the process returns to the top of the process, the driver state is detected again (S101), and then the series of subsequent processes (S102 to S104) described above are performed. Run.

On the other hand, when the driver speaks (S104: yes), the driver's voice is acquired as it is (S105).
Then, the speech is recognized using the read identification dictionary data (S106). That is, the acquired speech signal is divided into a plurality of phonemes, and a plurality of feature amounts are extracted from each phoneme (see FIG. 3). Then, the type of sound represented by the phoneme (which vowel or consonant) is specified by comparing the plurality of feature values extracted for each phoneme with the read identification dictionary data (see FIG. 4). ). Thereafter, the speech is recognized by combining a plurality of phonemes (see FIG. 5).

When the voice is recognized in this manner, the result is output to an external device (for example, a control device (not shown) of the vehicle 1) (S107), and then it is determined whether or not the voice recognition is to be ended (S108).
As a result, when it is determined that the speech recognition is not finished (S108: no), the process returns to the top of the process, and after detecting the driver state again (S101), the above-described series of processes (S102 to S108) are performed. Run. If it is determined that the voice recognition is to be ended while repeating such a process (S108: yes), the voice recognition process of FIG. 6 is ended.

  As described in detail above, in the voice recognition processing of the present embodiment, appropriate identification dictionary data is properly used according to the driver's state, so that the recognition accuracy of voice recognition can be greatly improved.

C. Modified example:
In the above-described embodiment, the identification dictionary data corresponding to the driver state is described as being stored in the memory in advance. However, identification dictionary data corresponding to the driver state may be generated. In this way, the identification dictionary data corresponding to the driver state can be generated for each driver, so that the recognition accuracy of voice recognition can be further improved.

  FIG. 9 shows a rough internal structure of a modified speech recognition apparatus 200 that generates identification dictionary data corresponding to the driver state for each driver. The speech recognition apparatus 200 of the illustrated modification is different from the speech recognition apparatus 100 of the present embodiment described above with reference to FIG. 2 with a speech learning unit 201, an identification dictionary data generation unit 202, and a standard identification dictionary data storage unit. The difference is that 203 is added.

Of these, the voice learning unit 201 learns the driver's voice based on the driver's voice signal acquired by the voice acquisition unit 104.
The identification dictionary data generation unit 202 generates identification dictionary data corresponding to the driver state based on the learned driver's voice. When generating identification dictionary data, first, identification dictionary data (standard identification dictionary data) set for each driver state assuming a standard driver is acquired from the standard identification dictionary data storage unit 203. . Subsequently, the acquired standard identification dictionary data is corrected according to the deviation between the standard driver's voice and the learned driver's voice to identify the driver according to the driver's condition for the learned driver. Generate dictionary data. Then, the identification dictionary data generation unit 202 stores the generated identification dictionary data in the identification dictionary data storage unit 103. The process of generating the learned driver identification dictionary data will be described in detail later.

  Other points are the same as those of the speech recognition apparatus 100 of the present embodiment described above with reference to FIG. Hereinafter, in brief, the driver state detection unit 101 detects the driver state and outputs the result to the identification dictionary data selection unit 102. Then, the identification dictionary data selection unit 102 selects identification dictionary data corresponding to the driver state from the identification dictionary data stored in the identification dictionary data storage unit 103. The voice recognition unit 105 performs voice recognition using the voice signal acquired by the voice acquisition unit 104 while referring to the identification dictionary data selected by the identification dictionary data selection unit 102.

FIG. 10 shows a flowchart of identification dictionary data generation processing in which the voice recognition device 200 according to the modification generates identification dictionary data corresponding to the driver state for each driver.
In the identification dictionary data generation process, first, the driver state is detected (S201). Here, the driver state is detected by analyzing the driver's face image captured by the in-vehicle camera 20, but the driver state may be detected by other methods.

Then, it is determined whether or not the driver state is a normal state (S202). As a result, when the driver state is not the normal state (S202: no), it is considered that it is not suitable for generating the identification dictionary data corresponding to the driver state. Therefore, after returning to the top of the process and detecting the driver state (S201), it is determined again whether or not the driver state is the normal state (S202).
If it is determined that the driver's state is the normal state while repeating such determination (S202: yes), driving is performed by a method such as reading out a preset word or sentence to the driver. The user's learning voice is acquired (S203).

  Based on the driver's voice learned in this way, identification dictionary data for the normal state is generated (S204). That is, a distribution range in the feature amount space is generated for each of a plurality of vowels and consonants. Various methods are known for generating identification dictionary data based on learned speech, and any method may be used. Since the identification dictionary data generated in this way is generated based on the voice when the driver is in the normal state, it becomes the identification dictionary data for the normal state.

Subsequently, identification dictionary data for the normal state is acquired from the identification dictionary data (standard identification dictionary data) set for each driver state assuming a standard driver (S205). Standard identification dictionary data is stored in advance in the memory of the speech recognition apparatus 200 according to the modification.
Then, the normal state identification dictionary data generated in S204 is compared with the normal state identification dictionary data read in S205 (S206). By doing so, the correction amount of the identification dictionary data for generating the learned driver identification dictionary data can be obtained from the standard driver identification dictionary data.
Thereafter, the learned identification dictionary data for the driver is generated by correcting the standard identification dictionary data with the correction amount of the identification dictionary data thus obtained (S207).

  FIG. 11 illustrates a state where the learned identification dictionary data for the driver is generated by correcting the standard identification dictionary data. As shown in FIG. 7, here, there are seven driver states of state A to state G. Therefore, the identification dictionary data of these seven driver states is set in the standard identification dictionary data. Has been. However, in order to avoid the illustration from becoming complicated, in FIG. 11, identification dictionary data for three driver states of state A to state C is displayed. In addition, in each identification dictionary data, a distribution range in the feature amount space is described for each of a plurality of vowels and consonants. In order to avoid complicated illustration of these, in FIG. The distribution range in the feature amount space is displayed for the vowel of “a”.

For example, the driver identification dictionary data learned in the state A (normal state) is a distribution indicated by a broken line in FIG. 11A, and the standard identification dictionary data is a distribution indicated by a solid line in the figure. Suppose. This difference in distribution is due to individual differences among drivers. Therefore, it is considered that the difference due to the individual difference that occurs in the state A (normal state) also occurs in other driver states.
The speech recognition apparatus 200 according to the modification generates identification dictionary data corresponding to each driver state based on such a concept. For example, for the state B, if the standard identification dictionary data has a distribution indicated by a solid line in FIG. 11B, the learned identification dictionary data for the driver is considered to have a distribution indicated by a broken line in the figure. . Similarly, for the state C, if the standard identification dictionary data has a distribution indicated by a solid line in FIG. 11C, the learned identification dictionary data for the driver is considered to have a distribution indicated by a broken line in the figure. .
In S207 of FIG. 10, identification dictionary data corresponding to the learned driver state for the driver is generated from the standard identification dictionary data in this way. Then, the generated identification dictionary data for the driver is stored in a memory (not shown) (S208), and the identification dictionary data generation process of FIG. 11 is terminated.

  If the identification dictionary data corresponding to the driver state is generated for each driver as described above, it is possible to recognize the voice in consideration of not only the driver state but also the difference due to the individual difference of the driver. As a result, the recognition accuracy can be further improved.

  As mentioned above, although the present Example and the modification were demonstrated, this invention is not restricted to said Example and a modification, It can implement in a various aspect in the range which does not deviate from the summary.

1 ... Vehicle, 10 ... Microphone, 20 ... In-vehicle camera,
DESCRIPTION OF SYMBOLS 100 ... Voice recognition apparatus, 101 ... Driver state detection part,
102: Identification dictionary data selection unit 103: Identification dictionary data storage unit,
104 ... voice acquisition unit, 105 ... voice recognition unit, 200 ... voice recognition device,
201 ... voice learning unit, 202 ... identification dictionary data generation unit,
203: Standard identification dictionary data storage unit.

Claims (4)

A voice recognition device that obtains a driver's voice, divides the voice into phonemes of a predetermined time width, and recognizes the voice by comparing the phonemes with identification dictionary data,
A driver state detector for detecting the driver state of the driver;
An identification dictionary data storage unit storing the identification dictionary data corresponding to the driver state;
An identification dictionary data selection unit for selecting the identification dictionary data corresponding to the driver state from the stored identification dictionary data;
A voice recognition device comprising: a voice recognition unit that recognizes the voice using the selected identification dictionary data. The speech recognition device according to claim 1,
The driver recognition unit is a detection unit that detects the driver state by analyzing a face image of the driver. The speech recognition device according to claim 1 or 2,
A standard identification dictionary data storage unit for storing standard identification dictionary data set assuming a standard driver;
An identification dictionary data generating unit that generates the identification dictionary data of the driver from the standard identification dictionary data by learning the driver's voice and detecting a correction amount with respect to the standard driver's voice; A speech recognition apparatus comprising: A voice recognition method for acquiring a driver's voice, dividing the voice into phonemes of a predetermined time width, and recognizing the voice by comparing the phonemes with identification dictionary data,
Detecting the driver state of the driver;
Selecting the identification dictionary data according to the driver state from the identification dictionary data stored according to the driver state;
Recognizing the sound using the selected identification dictionary data.

Images