JP4459729B2 - In-vehicle speech recognition system - Google Patents

Description

  The present invention relates to an in-vehicle voice recognition system for recognizing a user's voice in a vehicle.

  2. Description of the Related Art Conventionally, in a field where the noise environment changes variously, there is a speech recognition system that can always increase a good recognition rate even if the noise environment changes variously. Specifically, in this system, for example, different voice models corresponding to different noise environments assumed in the field to be used, and each voice model performs voice recognition processing independently, and symbols recognized in the voice recognition processing. In addition to the above, the numerical value indicating the certainty is also calculated, and the speech recognition unit that outputs the certainty is compared with the certainty of the recognized symbol, the largest one is extracted, and the recognition symbol corresponding to it is extracted from the system. A probability comparison unit that outputs a recognition result, and performs a speech recognition process using the selected speech model. In this way, multiple voice models are prepared corresponding to different noise environments, and the optimum voice model is selected according to the noise environment at the time of recognition and voice recognition is executed. Therefore, even if the noise environment changes variously, a good recognition rate can always be raised following the change. In addition, the recognition rate can be increased while greatly reducing the load of voice recognition processing, and real-time voice recognition is possible (see, for example, Patent Document 1).

On the other hand, a device for recognizing the voice of a user (speaker) similarly includes a dictionary corresponding to various mental states in the speech recognition dictionary, and estimates the mental state of the speaker during speech recognition. Some use a dictionary corresponding to the target state to improve the recognition rate. Specifically, this device includes a sensor device that detects the speaker state and the running state of the automobile, and a mental state estimation device that estimates the mental state of the speaker from the output thereof. A speech recognition dictionary group including dictionaries created based on pronunciation under various tensions is provided. At the time of speech input, the dictionary selection device selects an optimal dictionary from the speech recognition dictionary group according to the mental state estimated value such as a margin estimated value, and the speech recognition means uses this optimal dictionary. Perform voice recognition. Thereby, since speech recognition is performed using the optimal dictionary according to a speaker's mental state, the recognition rate can be improved (for example, refer patent document 2).
JP 2000-75889 A JP 2002-149191 A

  By the way, a high noise resistance is required for a voice recognition device of an in-vehicle device mounted on a vehicle. Therefore, conventionally, as a countermeasure, speech recognition is performed using an acoustic model including noise characteristics as in the system described in Patent Document 1, but the noise included in the prepared acoustic model is actually used. The input noise does not always match, and if the two do not match, the recognition performance may be degraded, and the recognition performance may be degraded even in environments with little noise or acoustically different vehicles. There was a problem.

  On the other hand, in the apparatus described in Patent Document 2, although a recognition dictionary to be used for speech recognition is determined based on output signals from various sensors, since only one dictionary is selected, selection of a recognition dictionary is appropriate. If not, there is a problem that the recognition rate may decrease.

  The present invention has been made in view of the above problems, and an object of the present invention is to provide an in-vehicle speech recognition system that performs speech recognition in response to noise that varies depending on the environment.

In order to solve the above-mentioned problem, an in-vehicle voice recognition system according to the invention of claim 1 includes a voice input unit (for example, a microphone 1 in an embodiment to be described later) for inputting voice uttered by a user, and vehicle identification. As an environmental condition , a vehicle sensor unit that detects at least one of a vehicle speed, an air conditioner air volume, a wiper speed, a window opening degree, and an audio volume (for example, a vehicle speed sensor 5, an air conditioner ECU 6, an audio ECU 7, etc. in an embodiment described later) And a speech recognition dictionary unit (for example, described later) corresponding to a plurality of acoustic models (for example, an acoustic model a231, an acoustic model b232,..., An acoustic model n233, etc. described later) stored in advance for each specific environmental condition. The speech recognition dictionary unit 24) and the speech signal input from the speech input unit are speech-recognized using the speech recognition dictionary unit. The recognition process section for obtaining the recognition score was scored as (e.g. acoustic pattern recognition unit 25 of the embodiment described later), from the output of the vehicle sensor unit, the real environment determination section determines the actual vehicle environmental conditions (e.g. and the process of step S3) to execute the environment evaluation unit 4 of the embodiment described later, the actual vehicle environment condition determined by the actual environment determination section, a plurality of acoustic previously stored for each of the specific vehicle environmental conditions A model reliability calculation unit that calculates the model reliability of the acoustic model that decreases as the difference from the environmental condition of the model increases (for example, the process in step S8 executed by the environment evaluation unit 4 in an embodiment described later); confidence the model reliability calculated in the calculation unit and the recognition score of the acoustic model calculates the actual environmental score by accumulating, overall basis of the actual environmental score Characterized in that it comprises a final result calculation unit that calculates a voice recognition result (e.g., the final result of the embodiment described later calculation unit 9).

  The vehicle-mounted speech recognition system having the above configuration stores in advance a speech input unit for inputting speech uttered by a user, a vehicle sensor unit for detecting a specific environmental condition of the vehicle, and each specific environmental condition. A speech recognition dictionary unit corresponding to a plurality of acoustic models, and when a speech signal is input from the speech input unit, the speech recognition execution unit recognizes the speech signal using the speech recognition dictionary unit. In addition to scoring the result of speech recognition, calculating the model reliability of the determination result of the real environment determination unit that determines the actual vehicle environmental condition from the output of the vehicle sensor unit and the environmental conditions of a plurality of acoustic models stored in advance Compare each other and calculate the model reliability of the acoustic model. Then, the final result calculation unit corrects the score of the voice recognition result calculated by the voice recognition execution unit based on the model reliability calculated by the model reliability calculation unit, and calculates a comprehensive voice recognition result. The recognition result can be calculated by comprehensively judging the voice recognition result by a plurality of acoustic models and the vehicle state.

According to the in-vehicle speech recognition system according to claim 1, the score of the speech recognition result calculated by the speech recognition execution unit is corrected by the model reliability of the acoustic model determined from the actual vehicle environmental conditions, By calculating a simple speech recognition result, the recognition result can be calculated by comprehensively judging the speech recognition result by a plurality of acoustic models and the vehicle state.
Therefore, by accurately responding to noise that varies depending on the environment, it is possible to perform speech recognition that is less susceptible to environmental changes.

  Embodiments of the present invention will be described below with reference to the drawings.

(Device configuration)
FIG. 1 is a block diagram showing the overall configuration of an in-vehicle speech recognition system according to an embodiment of the present invention.
In FIG. 1, the in-vehicle voice recognition system of this embodiment includes a microphone 1 for inputting a user's voice, and the user's voice input from the microphone 1 is input to the voice recognition unit 2 and prepared. Speech recognition is performed sequentially or in parallel using some (or all) of the generated acoustic models. The speech recognition unit 2 is a processing unit that performs speech recognition to convert the input speech into “recognition result text” and output the “recognition result text” recognized by the speech recognition unit 2. , Voice recognition result table a31, voice recognition result table b32..., Voice recognition result table n33, etc.

  The speech recognition unit 2 will be described in detail. The speech recognition unit 2 samples and quantizes analog speech input from the microphone 1 and converts it into a digital speech signal. A frequency analysis unit 22 that performs spectrum analysis on the digital audio signal output from the / D conversion unit 21 and captures frequency characteristics of the audio, and a plurality of acoustic models such as an acoustic model a231, an acoustic model b232, and an acoustic model n233. While comparing the contents of the acoustic model storage unit 23 with the model and the “acoustic model” provided in the acoustic model storage unit 23 with the contents of the speech recognition dictionary unit 24, the digital speech signal output by the frequency analysis unit 22 is recognized. And an acoustic pattern recognition unit 25.

  The plurality of acoustic models provided in the acoustic model storage unit 23 is, for example, “runs at a speed of 60 [km / h] by the vehicle A and is operated by an air conditioner) in order to cope with noise that changes in the environment. The acoustic models suitable for specific environments such as “when the air volume is 30 [%]”, and the conditions for creating each acoustic model are referred to as “model environment values” in this embodiment. The speech recognition dictionary unit 24 stores speech recognition dictionaries corresponding to a plurality of acoustic models a231, acoustic models b232,. ing. Furthermore, the acoustic pattern recognition unit 25 recognizes the “recognition score” obtained when the recognition dictionary for each acoustic model used for speech recognition and the corresponding recognition dictionary for each “model environment value” are recognized and executed in an actual environment. Are output to the speech recognition result storage unit 3 as one or more recognition results.

  On the other hand, the environment evaluation unit 4 outputs output signals from a vehicle speed sensor 5 that detects the traveling speed of the vehicle, an air conditioner ECU 6 that controls an air conditioner mounted on the vehicle, an audio ECU 7 that controls an acoustic device mounted on the vehicle, and the like. And the sensor values of the vehicle speed sensor 5, air conditioner ECU 6, audio ECU 7, etc. and the state of each control ECU are temporarily stored in the temporary storage unit 8 and the acoustic model provided in the acoustic model storage unit 23. a processing unit that calculates “model reliability” of each of a231, acoustic model b232..., acoustic model n233.

  Specifically, the sensor values stored in the temporary storage unit 8 and the states of the control ECUs are referred to as “actual environment values” in the present embodiment, and the environment evaluation unit 4 includes the “actual environment values” and “ The model reliability is calculated by the following equation (1) using the “model environment value” and the “vehicle coefficient” which is a value unique to the vehicle. The “vehicle coefficient” is a coefficient value for correcting individual differences between the vehicle when each acoustic model is set and the vehicle actually used. In the following formula (1), “ABS ()” means to take the absolute value of the value in parentheses.

  The meanings of symbols in the above formula (1) are as shown in Table 1 below.

  In addition, the environment evaluation unit 4 calculates “actual environment scores” for all “recognition result texts” corresponding to each acoustic model, using “model reliability” and “recognition score”.

  Further, the final result calculation unit 9 compares all “recognition result texts” stored in each speech recognition result table of the speech recognition result storage unit 3 and outputs a result that seems to be optimal as a final recognition result. Specifically, when the same “recognition result text” exists in each speech recognition result table of the speech recognition result storage unit 3, the “real environment value” stored in the temporary storage unit 8 is a processing unit. , The “real environment score” of the “recognition result text” obtained corresponding to each acoustic model is added, and the final result storage unit 10 is used as the “final evaluation value” of each “recognition result text”. Is stored in the final result table 101. Then, “recognition result text” is output in descending order of “final evaluation value”. Details of the processing of the final result calculation unit 9 will be described later.

(Voice recognition processing)
Next, the operation of the in-vehicle speech recognition system of this embodiment will be described in detail with reference to the drawings. FIG. 2 is a flowchart showing the speech recognition processing operation of the in-vehicle speech recognition system of the present embodiment.
In FIG. 2, when voice is first input from the microphone 1 (step S1), the voice recognition unit 2 notifies the environment evaluation unit 4 of the start / end of voice input (step S2).
On the other hand, the environment evaluation unit 4 monitors output signals from the vehicle speed sensor 5, the air conditioner ECU 6, the audio ECU 7, and the like, and the “actual environment value” such as the state of the control ECU and the sensor value during voice input is stored in the temporary storage unit 8. Save (step S3).

On the other hand, the voice input from the microphone 1 is converted into a digital voice signal by the A / D converter 21 and the frequency analysis for analyzing the spectrum is executed by the frequency analyzer 22 (step S4).
Further, the frequency analysis unit 22 extracts the frequency characteristics of the digital audio signal from the frequency analysis result (step S5).
When the frequency characteristics of the digital audio signal are extracted, the acoustic pattern recognition unit 25 then selects the acoustic model a231, acoustic model b232,..., Acoustic model n233 provided in the acoustic model storage unit 23, respectively. Using the speech recognition dictionary unit 24, speech recognition is executed for each acoustic model (step S6).

In addition, the acoustic pattern recognition unit 25 uses the “recognition result text” and the “recognition score” to generate a speech recognition result table a31, a speech recognition result table b32,. In n33 and the like, each acoustic model is stored (step S7).
On the other hand, the environment evaluation unit 4 calculates “model reliability” for each acoustic model from the “real environment value”, “vehicle coefficient”, and “model environment value” using the above-described equation (1) (step S8). .
Furthermore, the environment evaluation unit 4 calculates the “real environment score” from the “model reliability” and the “recognition score” for all the recognition results, and stores them in each speech recognition result table (step S9).

  Here, “model environment value” and “model reliability” for each acoustic model stored in the speech recognition result table a31, the speech recognition result table b32..., The speech recognition result table n33, etc. of the speech recognition result storage unit 3. "Recognition result text", "Recognition score", "Real environment score", and "Candidate ranking" are shown in Tables 2 to 6 below. For example, the example shown in Table 2 is the content of the speech recognition result table a31. The example shown in Table 3 is the contents of the speech recognition result table b32. Similarly, the example shown in Table 4 is the contents of a speech recognition result table c (not shown). The example shown in Table 5 is the contents of the speech recognition result table d (not shown). The example shown in Table 6 is the contents of the speech recognition result table e (not shown).

  In the examples shown in Tables 2 to 6, the actual environment is actually running at 55 [km / h] per hour, the air volume of the air conditioner is 20 [%], the wiper is unused, and the vehicle window Is a recognition result of the voice “Ozawa to Utsunomiya” uttered with the audio volume set to 20 [dB] in the fully closed state. In this case, the variables in the actual environment are V = 55, T = 20, W = 0, O = 0, and A = 20, respectively. In the examples shown in Tables 2 to 6, the vehicle coefficient C is 1.2. The vehicle coefficient is a value for correcting a difference in acoustic characteristics between the vehicle when the acoustic model environment is set and the vehicle in the actual environment.

  Next, the final result calculation unit 9 compares all “recognition result texts” stored in the respective speech recognition result tables of the speech recognition result storage unit 3 and if the same “recognition result text” exists. , While referring to the “real environment values” stored in the temporary storage unit 8, all the “real environment scores” of the “recognition result texts” obtained corresponding to the respective acoustic models are added, and each “recognition result” The “final evaluation value” of “text” is stored in the final result table 101 of the final result storage unit 10 (step S10).

  Table 7 below is a table showing an example of each “recognition result text”, its “final evaluation value”, and “candidate ranking” stored in the final result table 101 of the final result storage unit 10. The “recognition result text” and the “real environment score” respectively obtained corresponding to the above correspond to the examples shown in Tables 2 to 6 above.

  As shown in Table 7 above, the in-vehicle speech recognition system according to the present embodiment ranks and outputs “recognition result text” in descending order of “final evaluation value” (step S11).

  In this embodiment, when the recognition process is performed on a plurality of acoustic models, the calculation load on the speech recognition unit 2 is high (heavy), so the process is distributed to an external server, another on-vehicle ECU, or the like. The distributed processing may be performed to reduce the load on the speech recognition unit 2 and accelerate the speech recognition processing.

  In the present embodiment, the environment evaluation unit 4 includes an actual environment determination unit and a model reliability calculation unit. Specifically, the process of step S3 executed by the environment evaluation unit 4 corresponds to a real environment determination unit, and the process of step S8 executed by the environment evaluation unit 4 corresponds to a model reliability calculation unit.

  As described above, according to the in-vehicle voice recognition system of the present embodiment, the microphone 1 for inputting the voice spoken by the user, the vehicle speed sensor 5 for detecting a specific environmental condition of the vehicle, the air conditioner ECU 6, the audio A vehicle sensor unit such as an ECU 7 and a speech recognition dictionary unit 24 corresponding to a plurality of acoustic models stored in advance for each specific environmental condition. When a speech signal is input from the microphone 1, the acoustic pattern recognition unit 25 However, the speech recognition dictionary unit 24 is used to recognize the speech signal, score the result of speech recognition, and the environment evaluation unit 4 determines the actual vehicle environment condition from the output of the vehicle sensor unit, The model reliability of the acoustic model is calculated by comparing the determination result with environmental conditions of a plurality of acoustic models stored in advance. Then, the final result calculation unit 9 corrects the score of the speech recognition result calculated by the acoustic pattern recognition unit 25 based on the calculated model reliability, and calculates a comprehensive speech recognition result.

Thereby, the score of the speech recognition result calculated by the acoustic pattern recognition unit 25 is corrected by the model reliability of the acoustic model determined from the actual vehicle environmental conditions, and a comprehensive speech recognition result can be calculated. .
Therefore, by accurately responding to noise that varies depending on the environment, it is possible to perform speech recognition that is less susceptible to environmental changes.

1 is a block diagram illustrating an overall configuration of an in-vehicle speech recognition system according to an embodiment of the present invention. It is a flowchart which shows the speech recognition process operation | movement of the vehicle-mounted speech recognition system of the Example.

Explanation of symbols

1 Microphone (voice input unit)
4 Environmental evaluation department (real environment judgment department, model reliability calculation department)
5 Vehicle speed sensor (vehicle sensor)
6 Air conditioner ECU (vehicle sensor unit)
7 Audio ECU (vehicle sensor unit)
9 Final result calculation unit (final result calculation unit)
24 Voice recognition dictionary part (voice recognition dictionary part)
25 Acoustic pattern recognition unit (voice recognition execution unit)
S3 Real environment determination unit S8 Model reliability calculation unit 231 Acoustic model a (acoustic model)
232 Acoustic model b (acoustic model)
233 Acoustic model n (acoustic model)

Claims (1)

A voice input unit for inputting voice spoken by the user;
As a specific environmental condition of the vehicle, a vehicle sensor unit for detecting at least one of a vehicle speed, an air conditioner air volume, a wiper speed, a window opening, and an audio volume ;
A speech recognition dictionary corresponding to a plurality of acoustic models stored in advance for each of the specific environmental conditions;
A voice recognition execution unit that recognizes a voice signal from the voice input unit using the voice recognition dictionary unit, scores a result of the voice recognition, and obtains a recognition score ;
An actual environment determination unit that determines an actual vehicle environmental condition from the output of the vehicle sensor unit;
The real and actual vehicle environment condition determined by the environment determination section, the model reliability of the acoustic model which the difference becomes more larger lower the environmental conditions of the plurality of acoustic models stored in advance for each of the specific vehicle environmental conditions A model reliability calculation unit for calculating
The real environment score is calculated by accumulating the model reliability calculated in the model reliability calculation unit and the recognition score of the acoustic model, and a final speech recognition result is calculated based on the real environment score A vehicle-mounted speech recognition system comprising a result calculation unit.

Images