JP2006071792A - Speech recognition device for vehicle - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To precisely recognize an uttered speech without letting a user perform stepwise input operation. <P>SOLUTION: A dictionary retrieval part 12 performs retrieval from a word dictionary where corresponding words are previously set for a speech extracted by a speech recognition part 11 to select words as recognition candidates. A Mahalanobis distance arithmetic part 13 reads vehicle information set in each recognition candidate, reads current vehicle information, and calculates Mahalanobis distances of respective groups to respective recognition candidates of the vehicle information. An uttered phrase determination output part 14 determines the group whose Mahalanobis distance becomes minimum as the group that the vehicle information belongs to, determines it as a recognition result corresponding to the uttered phrase of a driver, and outputs it. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a vehicle voice recognition apparatus for accurately recognizing a voice uttered in a passenger compartment.

  2. Description of the Related Art In recent years, various systems have been developed in vehicles, in which troublesome switch input and the like are omitted, and the sound generated by the driver can be sensed to operate the corresponding on-vehicle device for the convenience of the driver.

For example, in Japanese Patent Laid-Open No. 2000-20090, in a speech recognition device for a vehicle that performs speech recognition by searching for and specifying a word spoken by a user from a word dictionary that stores a plurality of words in advance, the user Is initially estimated as at least one primary request, a current state and a future state of the user are estimated from the primary request, and another request is estimated from the estimated state. .
JP 2000-20090 A

  However, in the speech recognition apparatus disclosed in the above-mentioned Patent Document 1, in addition to the utterance for estimating the primary request, the user needs to input personal information, which takes time and is troublesome to set. is there. Also, even if the user's request is estimated and the search range of the word dictionary is narrowed or the order is changed, it is ultimately limited to the word rank set in the word dictionary, so accurate recognition is possible. There is a problem that there is a limit in obtaining the result. In other words, the word dictionary ranking is often influenced by the frequency of word usage up to the previous time, and even if the user is placed in a completely different situation this time, the previous situation is considered. May be misconfigured as a result of being set.

  The present invention has been made in view of the above circumstances, and it is an object of the present invention to provide a vehicle voice recognition device that can easily recognize a user's utterance phrase with high accuracy without causing the user to perform stepwise input operations. Objective.

  The present invention relates to a voice input means for inputting voice, a vehicle information input means for inputting own vehicle information, and a word dictionary for selecting a recognition candidate by searching a word dictionary in which the input voice is preset. Mahalanobis distance that reads the information group set in advance for each recognition candidate and calculates the Mahalanobis distance between the read information group and the current vehicle information when there are a plurality of recognition candidates and the search means It is characterized by comprising a calculation means and an utterance phrase determination means for deciding an utterance phrase from among each of the recognition candidates based on the Mahalanobis distance calculated for each of the recognition candidates.

  According to the vehicle speech recognition apparatus of the present invention, it is possible to easily recognize a user's utterance phrase with high accuracy without causing the user to perform stepwise input operations.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
1 to 6 show an embodiment of the present invention, FIG. 1 is a functional block diagram of a vehicle voice recognition device, FIG. 2 is a flowchart of a voice recognition program, and FIG. 3 is information set for each utterance phrase. FIG. 4 is an explanatory diagram of a table of values for calculating the Mahalanobis distance, FIG. 4 is an explanatory diagram of a table of values to calculate the Mahalanobis distance, information set with “thick” and “Atsugi” as examples of utterance phrases, and FIG. 4 is an explanatory diagram of the own vehicle position information in FIG. 4, and FIG. 6 is an explanatory diagram showing the distribution of the own vehicle driving information and the own vehicle position information when obtaining the Mahalanobis distance in FIG.

  In FIG. 1, reference numeral 1 denotes a vehicle voice recognition device. The voice recognition device 1 includes in-vehicle CAN communication (communication based on the Controller Area Network (ISO standard)) as own vehicle information input means built in the vehicle. Connected to the network 2. Various in-vehicle control units (for example, an engine control unit, a transmission control unit, a brake control unit, etc.) are connected to the in-vehicle CAN communication network 2, and an in-vehicle temperature sensor, a vehicle speed sensor, and a wiper switch provided in the vehicle. Sensors such as brake switches, signals from switches, and data calculated by each control unit can be shared.

  In addition, a navigation device 3 is also connected to the voice recognition device 1 as own vehicle information input means. This navigation device 3 is composed of a GPS receiver, four wheel speed sensors, and other necessary sensors and switches, and vehicle travel information obtained from these and map information recorded on a CD-ROM or the like. Are calculated while performing map matching, etc., and the current position of the vehicle and its surrounding map, the optimum route to the destination, distance, direction, etc. are displayed on the liquid crystal display, or route guidance by voice from the speaker Go and guide.

  Furthermore, the speech recognition apparatus 1 is connected with a microphone 4 as a speech input means for capturing speech from a driver (user).

  Then, as will be described later, the voice recognition device 1 recognizes an utterance phrase by performing a recognition process on the voice from the driver input from the microphone 4 based on information from the in-vehicle CAN communication network 2 and the navigation device 3, The recognition result is output to, for example, the input system of the navigation device 3 or the input system of the vehicle air conditioner, and the settings of these necessary devices are varied.

  That is, the speech recognition apparatus 1 is mainly composed of a speech extraction unit 11, a dictionary search unit 12, a Mahalanobis distance calculation unit 13, and an utterance phrase determination output unit 14.

  The voice extraction unit 11 removes noise and the like from the voice input from the microphone 4, extracts only the voice, and outputs it to the dictionary search unit 12.

  The dictionary search unit 12 is provided as a word dictionary search unit, searches a word dictionary in which a corresponding word is set in advance for the voice input from the voice extraction unit 11, and selects it as a recognition candidate. And output to the Mahalanobis distance calculation unit 13. Here, the dictionary search unit 12 is not limited to homonyms such as “thick” and “hot”, but even if the ending part such as “hot” and “Atsugi” is unclear, All of these words are searched as recognition candidates, and are ranked according to the frequency of use so far and output.

  The dictionary search unit 12 receives a recognition result corresponding to the utterance phrase V of the driver from the utterance phrase determination output unit 14 and learns and updates the order of recognition candidates in a preset word dictionary.

  The Mahalanobis distance calculation unit 13 receives a recognition candidate from the dictionary search unit 12, and inputs own vehicle driving information such as in-vehicle temperature and vehicle speed from the in-vehicle CAN communication network 2, and the direction of the recognition candidate and the own vehicle from the navigation device 3. The vehicle position information such as the angle with the direction of travel and the destination information such as the distance of the vehicle to the recognition candidate are input.

Processing performed by the Mahalanobis distance calculation unit 13 will be described based on the table of FIG.
That is, let V be the actual utterance phrase of the driver, and A, B, and C be the recognition candidates searched by the dictionary search unit 12 for this utterance phrase V. A plurality of vehicle information (Ai, Bi, Ci: i = 1, 2,...) Is set in advance for these recognition candidates A, B, and C, and the respective vehicle information (Ai, Bi, Ci: i = i = 1, 2,..., Own vehicle driving information Pi (i = 1, 2,...) From the in-vehicle CAN communication network 2, own vehicle position information Qi (i = 1, 2,...) From the navigation device 3, It consists of destination information Ri (i = 1, 2,...) (Ai = (P1, P2,..., Q1, Q2,..., R1, R2,...), Bi = (P1, P2,. , Q2, ..., R1, R2, ...), Ci = (P1, P2, ..., Q1, Q2, ..., R1, R2, ...)).

  A group is formed for each vehicle information (Ai, Bi, Ci: i = 1, 2,...). That is, the group GA is composed of A1, A2,..., The group GB is composed of B1, B2,..., And the group GC is composed of C1, C2,.

  The average MA is set for each piece of information in the group GA, and MA = (Pav1, Pav2,..., Qav1, Qav2,..., Rav1, Rav2,...). Here, Pav1 is an average of own vehicle driving information P1 in vehicle information Ai, Pav2 is an average of own vehicle driving information P2 in vehicle information Ai, Qav1 is an average of own vehicle position information Q1 in vehicle information Ai, and Qav2 is vehicle information Ai. The average of the vehicle position information Q2 in the vehicle information, Rav1 is the average of the destination information R1 in the vehicle information Ai, and Rav2 is the average of the destination information R2 in the vehicle information Ai.

  Similarly, the average MB is set for each piece of information in the group GB, and MB = (Pav1, Pav2,..., Qav1, Qav2,..., Rav1, Rav2,...). Here, Pav1 is the average of the own vehicle driving information P1 in the vehicle information Bi, Pav2 is the average of the own vehicle driving information P2 in the vehicle information Bi, Qav1 is the average of the own vehicle position information Q1 in the vehicle information Bi, and Qav2 is the vehicle information Bi. The average of the vehicle position information Q2 at Rv1, Rav1 is the average of the destination information R1 in the vehicle information Bi, and Rav2 is the average of the destination information R2 in the vehicle information Bi.

  Further, the average MC is set for each piece of information in the group GC, and MC = (Pav1, Pav2,..., Qav1, Qav2,..., Rav1, Rav2,...). Here, Pav1 is an average of own vehicle driving information P1 in vehicle information Ci, Pav2 is an average of own vehicle driving information P2 in vehicle information Ci, Qav1 is an average of own vehicle position information Q1 in vehicle information Ci, and Qav2 is vehicle information Ci. The vehicle position information Q2 in the vehicle information Ci, Rav1 is the average of the destination information R1 in the vehicle information Ci, Rav2 is the average of the destination information R2 in the vehicle information Ci.

  The components of the variance-covariance matrix are the variance and covariance calculated from each component of the vehicle information in each group. σA is dispersion for each Pi, dispersion for each Qi, dispersion for each Ri, covariance for each combination of Pi, covariance for each combination of Qi, and each combination of Ri in the group GA Is a matrix with the covariance of. σB is the variance for each Pi, variance for each Qi, variance for each Ri, covariance for each combination of Pi, covariance for each combination of Qi, and each combination of Ri in the group GB Is a matrix with the covariance of. σC is the variance for each Pi, variance for each Qi, variance for each Ri, covariance for each combination of Pi, covariance for each combination of Qi, and each combination of Ri in the group GC Is a matrix with the covariance of. Specifically, σA will be described. ΣA is P1 dispersion, P2 dispersion,..., Q1 dispersion, Q2 dispersion,... R1 dispersion, R2 dispersion,..., P1 and P2 covariance , P2 and P3 covariance, ..., Q1 and Q2 covariance, Q2 and Q3 covariance, ..., R1 and R2 covariance, and R2 and R3 covariance, ... It has. Since σB and σC are the same as σA, specific descriptions thereof are omitted.

Then, in order to determine which group GA, GB, GC the vehicle information In inputted from the in-vehicle CAN communication network 2 when the driver utters the utterance phrase V belongs to, the vehicle information In and each group's Mahalanobis Calculate the distance. The vehicle information In and the Mahalanobis distance DA of the group GA are calculated by the following equation (1).
DA · DA = XA · (σA) ⁻¹ · YA (1)
Here, (σA) ⁻¹ is an inverse matrix of the variance-covariance matrix σA, XA is a row vector display of the difference between the vehicle information In and the average MA, and YA is a vertical vector display of the difference between the vehicle information In and the average MA.

Similarly, the vehicle information In and the Mahalanobis distance DB of the group GB are calculated by the following equation (2).
DB · DB = XB · (σB) ⁻¹ · YB (2)
Here, (σB) ⁻¹ is an inverse matrix of the variance-covariance matrix σB, XB is a row vector display of the difference between the vehicle information In and the average MB, and YB is a vertical vector display of the difference between the vehicle information In and the average MB.

Further, the vehicle information In and the Mahalanobis distance DC of the group GC are calculated by the following equation (3).
DC · DC = XC · (σC) ⁻¹ · YC (3)
Here, (σC) ⁻¹ is an inverse matrix of the variance-covariance matrix σC, XC is a row vector display of the difference between the vehicle information In and the average MC, and YC is a vertical vector display of the difference between the vehicle information In and the average MC.

  In this way, the Mahalanobis distances DA, DB, and DC calculated for the groups GA, GB, and GC are output to the utterance phrase determination output unit 14. That is, the Mahalanobis distance calculation unit 13 is provided as Mahalanobis distance calculation means.

  The utterance phrase determination output unit 14 receives the Mahalanobis distances DA, DB, and DC calculated for each group GA, GB, and GC from the Mahalanobis distance calculation unit 13 and the group having the smallest Mahalanobis distances DA, DB, and DC. The vehicle information In is determined as a group to which the vehicle information In belongs, is determined as a recognition result corresponding to the driver's utterance phrase V, and is output from the dictionary search unit 12 and the voice recognition device 1 to the outside. That is, the utterance phrase determination output unit 14 is provided as an utterance phrase determination means.

  Note that if the result of word dictionary search in the dictionary search unit 12 is that there is only one recognition candidate, the Mahalanobis distance calculation unit 13 does not calculate the Mahalanobis distance for the recognition candidate, and the utterance phrase is used as it is. The utterance phrase determination output unit 14 determines and outputs the recognition candidate as a recognition result corresponding to the utterance phrase V of the driver.

  In addition, when a group of Mahalanobis distances having exactly the same value is input from the Mahalanobis distance calculation unit 13, the utterance phrase determination output unit 14 selects the highest recognition candidate in the recognition candidate order searched by the dictionary search unit 12. It is determined and output as a recognition result corresponding to the utterance phrase V of the driver.

Next, the speech recognition program executed by the speech recognition apparatus 1 will be described with reference to the flowchart of FIG.
First, in step (hereinafter abbreviated as “S”) 101, the voice extraction unit 11 removes noise and the like from the voice input from the microphone 4 and extracts only the voice.

  Next, in S102, the dictionary search unit 12 searches a word dictionary in which a corresponding word is set in advance for the voice extracted in S101, and selects it as a recognition candidate.

  And it progresses to S103, it is determined whether the recognition candidate of S102 is one, and when it is one, it progresses to S104, the utterance phrase determination output part 14 determines an utterance phrase as the recognition candidate, and progresses to S105. , Output the recognition result and exit the program.

  On the other hand, if it is determined in S103 that there is not one recognition candidate, that is, a plurality of recognition candidates, the process proceeds to S106, and the Mahalanobis distance calculation unit 13 sets the vehicle information (Ai, Bi, Ci: i = 1, 2,.

  Next, the process proceeds to S107, the Mahalanobis distance calculation unit 13 reads the current vehicle information In, and the process proceeds to S108, where the Mahalanobis distance calculation unit 13 corresponds to each recognition candidate A, B, C of the vehicle information In as described above. The Mahalanobis distance DA, DB, DC of each group GA, GB, GC is calculated.

  Next, proceeding to S108, the utterance phrase determination output unit 14 determines the group having the smallest Mahalanobis distance DA, DB, DC as the group to which the vehicle information In belongs, and corresponds to the utterance phrase V of the driver. Determine and output the recognition result and exit the program.

  The utterance phrase determination output unit 14 recognizes the recognition candidate corresponding to the utterance phrase V of the driver as the highest recognition candidate in the recognition candidate order searched by the dictionary search unit 12 when there is the same Mahalanobis distance. Determine and output as a result and exit the program.

  Next, speech recognition executed by the speech recognition apparatus 1 will be described with a more specific example. Here, a case where the driver speaks “hot”, that is, a case where the driver's utterance phrase V is “hot” will be described.

  In contrast, it is assumed that the dictionary search unit 12 searches a preset word dictionary and selects “hot” and “Atsugi” as recognition candidates. FIG. 4 shows a table created for this “hot” “Atsugi” corresponding to FIG. That is, the vehicle information preset for the recognition candidate “hot” is A1 to A4, and the vehicle information preset for the recognition candidate “Atsugi” is B1 to B4. And as own vehicle information, only P1 is set and this is in-vehicle temperature. Further, only Q1 is set as the vehicle position information, and this is an angle with the current vehicle traveling direction with the position of the place name of the recognition candidate as the destination (see FIG. 5). In addition, destination information is not set.

  FIG. 6 shows the in-vehicle temperature in the horizontal axis and the angle in the vertical axis. The distribution of the vehicle information A1 to A4 of the recognition candidate “hot” is plotted with white circles, and the distribution of the vehicle information B1 to B4 of the recognition candidate “Atsugi” is plotted with x points. Further, the In point (in-vehicle temperature 25 ° C., angle 60 °) in FIG. In such a distribution state, whether the current vehicle information In belongs to the distribution group GA of the vehicle information A1 to A4 of the recognition candidate “hot”, or the distribution group GB of the vehicle information B1 to B4 of the recognition candidate “Atsugi” Whether it belongs or not is determined by calculating the Mahalanobis distance.

  Thus, the Mahalanobis distance DA to the group GA of the distribution of the current vehicle information In and the recognition candidate “hot” vehicle information A1 to A4 is the distribution of the current vehicle information In and the vehicle information B1 to B4 of the recognition candidate “Atsugi”. If it is smaller than the Mahalanobis distance DB to the group GB, the current vehicle information In is judged to belong to the group GA of the distribution of the vehicle information A1 to A4 of the recognition candidate “hot”, and “hot” is recognized. The result is output.

  As described above, according to the embodiment of the present invention, it is determined by the statistical analysis of Mahalanobis distance to which of the recognition candidates the driver's utterance phrase V belongs, so that the accurate recognition result reflecting the current situation is obtained. Can be obtained.

  Further, since the determination is made by reflecting the current situation at the time of recognition, the driver does not need to perform an input operation other than the utterance and is easy to use.

  In the embodiment of the present invention, examples of “hot” and “Atsugi” are described as recognition candidates, but the present invention is not limited to this.

Functional block diagram of a vehicle voice recognition device Voice recognition program flowchart Explanatory drawing of the table of each value that calculates information set for each utterance phrase and Mahalanobis distance An explanatory diagram of the table of each value that calculates information and Mahalanobis distance set with "thick" and "Atsugi" as examples of utterance phrases Explanatory drawing of own vehicle position information in FIG. Explanatory drawing which shows distribution of the own vehicle driving information and the own vehicle position information when obtaining the Mahalanobis distance in FIG.

Explanation of symbols

1 Voice recognition device 2 In-vehicle CAN communication network (own vehicle information input means)
3. Navigation device (vehicle information input means)
4 Microphone (voice input means)
11 voice extraction unit 12 dictionary search unit (word dictionary search means)
13 Mahalanobis distance calculator (Mahalanobis distance calculator)
14 Utterance phrase determination output section (utterance phrase determination means)
Agent Patent Attorney Susumu Ito

Claims (2)

Voice input means for inputting voice;
Own vehicle information input means for inputting own vehicle information;
A word dictionary search means for selecting a recognition candidate by searching a word dictionary in which the input voice is set in advance;
When there are a plurality of recognition candidates, a Mahalanobis distance calculation means for reading a preset information group for each recognition candidate and calculating a Mahalanobis distance between the read information group and current vehicle information;
An utterance phrase determining means for determining an utterance phrase from the recognition candidates based on the Mahalanobis distance calculated for each of the recognition candidates;
A voice recognition device for a vehicle, comprising:   The vehicle speech recognition apparatus according to claim 1, wherein the information group set in advance for each recognition candidate includes any one of vehicle driving information, vehicle position information, and destination position information. .

Images