JP5037018B2 - Speech recognition apparatus and speech recognition method - Google Patents

Description

  The present invention relates to a voice recognition apparatus and a voice recognition method for recognizing a significant recognition target such as a word from a voice signal. The present invention also relates to a voice instruction device, an information device, and a vehicle system provided with such a voice recognition device.

  Some car navigation devices mounted on automobiles include a voice instruction device (voice input interface). The voice instruction device includes a voice recognition device that recognizes a driver's voice and a command generation unit that generates an instruction command corresponding to the recognized voice. In accordance with the instruction command generated by the command generation unit, the car navigation device executes destination setting, route search, and other operations.

Since the voice instruction device is an eye-free and hands-free interface, the driver can operate the car navigation device without hindering safe driving of the vehicle.
There is a demand to use a navigation device and other information devices not only in automobiles but also in motorcycles. Also in this case, as in the case of an automobile, an eye-free and hands-free interface is required, and the voice instruction device as described above is the most likely candidate.

A general word speech recognition apparatus includes a speech analysis unit that sequentially converts input speech to be recognized into acoustic feature parameters, and the obtained acoustic feature parameters for each word that has been created in advance. Compared with an acoustic library (recognition dictionary) that is a set of standard acoustic features, a matching unit that identifies a word represented by the input speech is included.
Generally, when creating an acoustic library, speech data of many speakers are recorded. By using the recorded data as learning data, an acoustic library having a style suitable for the recognition algorithm employed by the collation unit is created.
JP 2003-114696 A JP 2005-134436 A

However, it is impossible to collect learning data that covers all possible situations. In addition, depending on the feature parameters and the recognition (collation) algorithm employed, it is inevitable that a kind of “癖” occurs in the recognition process. For this reason, it is impossible to recognize all words with a uniform probability, and words that are easy to recognize and words that are difficult to recognize are inevitably generated.
Accordingly, an object of the present invention is to provide a speech recognition device and a speech recognition method that reduce non-uniformity in recognition probability between recognition objects (words, speech units, etc.) and realize more reliable speech recognition.

  Another object of the present invention is to provide a voice instruction device, an information device, and a vehicle system provided with such a voice recognition device.

In order to achieve the above object, the invention according to claim 1 recognizes an input speech signal in a forward direction according to an input time series, and the input speech signal corresponds to a recognition candidate (for each of a plurality of recognition candidates, for example). Forward direction recognition means for generating forward likelihood information representing likelihood, and an input speech signal is recognized in a reverse direction according to a reverse time series opposite to the input time series (for example, for each of a plurality of recognition candidates). The reverse speech recognition means for generating reverse likelihood information representing the likelihood that the input speech signal corresponds to the recognition candidate, and the outputs of the forward recognition means and the backward recognition means are integrated to correspond to the input speech signal. A speech recognition apparatus including an integrated determination unit that generates a recognition result. The integrated determination unit combines forward likelihood information and backward likelihood information generated by the forward direction recognition unit and the backward direction recognition unit, respectively, and combines the input speech signal representing the likelihood of being a recognition candidate. It includes likelihood information combining means for generating likelihood information and recognition result determination means for evaluating the combined likelihood information generated by the likelihood information combining means to obtain a recognition result. Also, the likelihood information coupling means, seen including weighting means for combining weighted respectively forward likelihood information and the backward likelihood information, the weighting coupling means, variations of the recognition probability between recognition candidate predetermined for each recognition candidate by learning as reduced, including the recognition candidate adaptive weighting means for combining by applying a weight that depends on the recognition candidates in the forward likelihood information and / or reverse likelihood information .
According to this configuration, not only the voice signal is recognized in the forward direction according to the input time series, but also the reverse direction according to the time series opposite to the input time series is recognized. Can be reduced. Thereby, more reliable voice recognition processing can be performed. In addition, according to the present invention, since the recognition result can be generated by comprehensively determining the recognition processing result in the forward direction and the backward direction, a recognition probability with little variation among recognition objects can be realized, and more reliable recognition processing can be performed. Become. Furthermore, since the recognition result is obtained by evaluating the combined likelihood information obtained by combining the forward likelihood information and the backward likelihood information, the recognition result reflecting the recognition processing result in the forward direction and the reverse direction is obtained. Thereby, variation in recognition probability for a plurality of recognition targets is reduced, and more reliable speech recognition processing can be performed. Since the combination of the recognition processing results in the forward direction and the backward direction is executed by appropriately weighting them, the recognition probability can be further improved.

  The forward direction and reverse direction recognition means each include, for example, a voice analysis means for generating a feature sequence of the input voice signal, a recognition dictionary (acoustic library: acoustic model, word dictionary, etc.), and a matching means. The recognition dictionary stores standard feature sequences created in advance for a plurality of recognition candidates. The matching means compares the feature sequence obtained by the speech analysis with the standard feature sequence stored in the recognition dictionary for each recognition candidate, and represents the likelihood that the input speech signal represents each recognition candidate. Generate likelihood information. More specifically, the matching unit obtains likelihood information according to a recognition algorithm such as a DP matching method (dynamic programming), a neural network method, a Bayes discriminant function method, a vector quantization method, or the like. Likelihood information is information representing the similarity between a feature sequence obtained by speech analysis and a standard feature sequence of recognition candidates stored in the recognition dictionary, and depends on the recognition algorithm executed by the matching means. Format information. Examples are likelihood and distance between feature sequence vectors.

  The recognition dictionary can be created by learning using previously recorded audio data as learning data. In this case, the recognition dictionary for the forward direction recognition processing means is created by learning using a feature sequence according to the input time series of the speech data, and the recognition dictionary for the backward direction recognition means is used when the speech data is input. It must be created by learning using a feature sequence according to a time sequence opposite to the sequence.

The weighting / combining means uses the weights depending on the recognition candidates, which are determined in advance for each recognition candidate so as to reduce the variation in the recognition probability among the recognition candidates, in the forward likelihood information and / or the reverse likelihood information. Recognition candidate adaptive weighting combining means for adding and combining. Therefore, since an appropriate weight can be assigned to each recognition candidate, variation in recognition probability among recognition candidates can be further reduced, and more reliable voice recognition can be performed. The weight for each recognition candidate is preferably determined in advance so that variation in recognition probability is reduced by learning, for example, and is registered in the recognition dictionary in association with the recognition candidate.
The invention according to claim 2 is the speech recognition apparatus according to claim 1, further comprising reverse reproduction means for reproducing the input voice signal according to a time series reverse to the input time series and inputting the input voice signal to the reverse direction recognition means. . With this configuration, the reverse direction recognizing unit can perform the input speech signal recognition processing according to the time series reverse to the input time series. More specifically, the backward direction recognizing means analyzes the voice signal reproduced in the reverse direction by the backward reproducing means and outputs the feature series, a recognition dictionary, and the feature series output by the voice analyzing means. Can be configured to include a matching unit that matches the contents of the recognition dictionary. The recognition dictionary in this case is preferably created by learning using a process in which learning data is reversely reproduced by reverse reproduction means and the output is analyzed by voice analysis means to obtain a feature series.

According to a third aspect of the present invention, the backward direction recognizing unit analyzes the input speech signal according to the input time series, generates a feature sequence according to the input time series, and a feature sequence generated by the speech analysis unit. The speech recognition apparatus according to claim 1, further comprising: a feature sequence inversion unit that converts the feature sequence according to a time sequence opposite to the input time sequence and inputs the converted feature sequence to the backward direction recognition unit. With this configuration, the reverse direction recognizing unit can perform the input speech signal recognition processing according to the time series reverse to the input time series. More specifically, the backward direction recognition means includes the voice analysis means, the feature series inversion means, a recognition dictionary, and a matching means for matching the feature series output from the feature series inversion means with the contents of the recognition dictionary. It can be set as the structure containing. The recognition dictionary in this case is preferably created by learning using a process in which the learning data is analyzed by the voice analysis means according to the input time series, and is inverted on the time axis by the feature series inversion means.

In the case of this configuration, the voice analysis process is a process common to both the forward direction and backward direction recognition processing means, and therefore, the voice analysis means may be shared by them .

The likelihood information combining means may be a linear combination (primary combination) of the forward likelihood information and the reverse likelihood information, or may be a non-linear combination (for example, a width equal to or larger than the square of the likelihood information). (Combining including a multiplicative term) .

According to a fourth aspect of the present invention, in the integrated determination unit, does the single processing result by at least one of the forward direction recognition unit and the backward direction recognition unit indicate the presence of a recognition candidate having a certain likelihood or more? When the one-way recognition determination unit and the one-way recognition determination unit determine whether the single processing result indicates the existence of a recognition candidate having a certain likelihood or more, the recognition candidate is determined. it is intended to include a means for outputting as a recognition result, a speech recognition apparatus according to any one of claims 1-3.

According to this configuration, since the recognition result may be output using the single processing result of the forward direction recognition unit or the backward direction recognition unit, the processing can be simplified and speeded up.
The invention according to claim 5 is characterized in that the one-way recognition determination means assigns a weight depending on a recognition candidate to the single processing result, and indicates that there is a recognition candidate having a likelihood that the single processing result is greater than or equal to a predetermined value. The speech recognition apparatus according to claim 4 , further comprising weighting determination means for determining whether or not a user is present. With this configuration, by assigning an appropriate weight to each recognition candidate, the probability of recognition processing based on the single processing result can be increased. The weight for each recognition candidate is preferably obtained in advance by learning and registered in advance in the recognition dictionary in association with the recognition candidate.

The invention according to claim 6 is the speech recognition device according to any one of claims 1 to 5 , speech signal input means for inputting a speech signal to the speech recognition device, and a recognition result by the speech recognition device. A voice instruction device including command data generating means for converting into instruction command data for input to a predetermined device. In this configuration, as a result of reducing the variation in recognition probability among recognition candidates, the probability that instruction command data corresponding to the user's intention is generated increases. Thereby, an excellent voice input interface can be provided.

A seventh aspect of the invention is an information device including the voice instruction device according to the sixth aspect and command processing means that operates in accordance with the instruction command data generated by the command data generation means. With this configuration, the information device can be comfortably operated by voice instructions.
The invention according to claim 8 is a vehicle system in which the command processing means of the information equipment according to claim 7 is mounted on a vehicle body. With this configuration, it is possible to provide a vehicle system that can comfortably operate the in-vehicle information device by voice instructions.

The invention according to claim 9 recognizes an input speech signal in a forward direction according to an input time series, and generates a forward likelihood information indicating the likelihood that the input speech signal corresponds to a recognition candidate; A reverse direction recognition step for recognizing a speech signal in a reverse direction according to a reverse time series opposite to an input time series, and generating reverse likelihood information indicating likelihood that the input speech signal corresponds to a recognition candidate; and the forward direction An integrated determination step that integrates the results of the recognition step and the backward recognition step to generate a recognition result corresponding to the input voice signal, and the integrated determination step includes the forward direction recognition step and the backward direction recognition step, respectively. Likelihood information combination that combines the generated forward likelihood information and reverse likelihood information to generate combined likelihood information representing the likelihood that the input speech signal corresponds to the recognition candidate. And a recognition result determination step for obtaining a recognition result by evaluating the combined likelihood information generated in the likelihood information combining step, wherein the likelihood information combining step includes the forward likelihood information and the reverse likelihood information. look including a weighted combination step of combining by weighting each time information, said weighted combination step, variations in the recognition probability between recognition candidate is predetermined by learning for each recognition candidate as reduced, it depends on the recognition candidate weighting the forward likelihood information and / or recognition candidate adaptive weighting combiner steps including binding to impart the opposite direction likelihood information for a speech recognition method. By this method, the effect described in relation to the invention of claim 1 can be achieved .
In the invention according to claim 10 , in the integration determination step, does the single processing result by at least one of the forward direction recognition step and the backward direction recognition step indicate the presence of a recognition candidate having a certain likelihood or more? When the one-way recognition determination step and the one-way recognition determination step determine whether the single processing result indicates the presence of a recognition candidate having a certain likelihood or more, the recognition candidate is determined. The speech recognition method according to claim 9 , further comprising a step of outputting as a recognition result.
The invention according to claim 11 is characterized in that the one-way recognition determination step assigns a weight depending on a recognition candidate to the single processing result, and indicates that the single processing result has a likelihood of being more than a predetermined likelihood. The speech recognition method according to claim 10 , further comprising a weighting determination step for determining whether or not the image is present.

  Of course, the invention similar to the invention of the speech recognition apparatus can be applied to the invention of the speech recognition method.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.
FIG. 1 shows an overall configuration of a vehicle system according to an embodiment of the present invention. This vehicle system includes a vehicle body 1 of a two-wheeled vehicle as an example of a vehicle, a voice instruction device body 10A attached to the vehicle body 1, and a command (command) attached to the vehicle body 1 from the voice instruction device body 10A. An information processing unit 51 as command processing means that receives and operates, a microphone 5 provided in a helmet 3 worn by an occupant 2 (usually a driver, but may be a passenger), and a helmet And a speaker 6 attached to 3. The microphone 5 functions as a voice signal input unit that inputs a voice signal to the voice instruction apparatus body 10A. The microphone 5 and the voice instruction apparatus main body 10A constitute a voice instruction apparatus 10. The voice instruction apparatus 10 and the information processing unit 51 constitute an in-vehicle information device 50. The voice instruction device main body 10A and the information processing unit 51 may be integrated, or may be separate devices.

Thus, the in-vehicle information device 50 is an information device capable of performing an instruction operation by voice. Examples of such information equipment include a navigation device (preferably capable of voice guidance), a mobile phone, and a sound reproduction device (for example, an MD player, a CD player, or other audio device).
The microphone 5 is provided at the mouth portion of the helmet 3, and the speaker 6 is provided at the ear portion of the helmet 3. The microphone 5 detects a voice emitted by the occupant and inputs a voice signal, which is an electrical signal corresponding to the voice, to the voice instruction apparatus main body 10A. The speaker 6 receives the audio signal from the information processing unit 51 and converts the audio signal into sound. Thus, the vehicle occupant can give a voice instruction to the in-vehicle information device 50 by voice and can listen to sound information generated by the in-vehicle information device 50. In this way, a hands-free and eyes-free interface is constructed.

The voice instruction device main body 10A may be mounted on the helmet 3. However, in order to reduce the number of accessories of the helmet 3 as much as possible, the voice instruction device main body 10A may be configured as a portable device held on the occupant's clothes or the like, It is preferable to configure as an in-vehicle device attached to the vehicle.
The connection between the microphone 5 and the voice instruction device main body 10A may be made by a wired connection using a cable included in the harness 7, or when the voice instruction device main body 10A is configured as an in-vehicle device, the connection is wireless. They can also be connected by communication. Similarly, the connection between the speaker 6 and the information processing unit 51 may be performed by wired connection using a cable included in the harness 7 or may be performed by wireless communication connection. As a wireless communication method, for example, Bluetooth, infrared communication, or other short-range wireless communication methods can be employed.

  FIG. 2 is a block diagram showing an electrical configuration of the in-vehicle information device 50. The voice instruction device main body 10A includes a voice recognition device 11 that recognizes a voice signal input from the microphone 5, a command data generation unit 12 (command data generation unit) that converts a recognition result of the voice recognition device 11 into command data. It has. The command data generation unit 12 generates command data in a format that can be input to the in-vehicle information device 50, and provides the command data to the information processing unit 51 of the in-vehicle information device 50.

  The speech recognition device 11 is a pair of recognition processing units that perform input speech recognition processing according to, for example, an HMM (Hidden Markov Model) word recognition algorithm, that is, a forward direction recognition processing unit 20 (forward direction recognition means) and a backward direction recognition process. The unit 30 (reverse direction recognition means) is provided. Further, the speech recognition apparatus 11 integrates the outputs of the reverse playback unit 15 (reverse playback means) that reverses and plays back the audio signal on the time axis, and the forward and reverse direction recognition processing units 20 and 30 to obtain a recognition result. And an integrated determination unit 40 (integration determination means). The speech signal to be recognized is input to the forward direction recognition processing unit 20 as a forward direction waveform along the time axis. On the other hand, the audio signal having the reverse waveform inverted on the time axis by the reverse reproduction unit 15 is input to the reverse direction recognition processing unit 30.

The forward direction recognition processing unit 20 includes a speech analysis unit 21 (speech analysis unit), an acoustic library 22 (recognition dictionary), and a collation unit 23 (collation unit). Similarly, the reverse direction recognition processing unit 30 includes a speech analysis unit 31 (speech analysis unit, acoustic library 32 (recognition dictionary), and collation unit 33 (collation unit).
The voice analysis unit 21 analyzes a voice signal (forward waveform) input from the microphone 5 and extracts a feature parameter series (for example, MFCC (Mel Frequency Cepstrum Coefficient)) representing the acoustic feature. Similarly, the voice analysis unit 31 analyzes the voice signal (reverse waveform) input from the reverse reproduction unit 15 and extracts a feature parameter series (for example, MFCC) representing the acoustic feature.

  The acoustic library 22 includes an acoustic model 24 and a word dictionary (language model) 25. Similarly, the acoustic library 32 includes an acoustic model 34 and a word dictionary (language model) 35. The acoustic models 24 and 34 are obtained by modeling the acoustic features of the standard speech pattern for each predetermined unit (for example, word) of speech, and reference information for evaluating the acoustic similarity with the input speech pattern. It is. Further, the word dictionaries 25 and 35 are information for giving restrictions on connection of the acoustic model. A typical example of such information is a probability that another word (phoneme) appears after a certain word (phoneme).

  The acoustic models 24 and 34 and the word dictionaries 25 and 35 are created by learning using learning data recorded in advance. The learning data is, for example, voice data that records utterances of various speakers in various situations. However, the acoustic model 24 and the word dictionary 25 of the forward direction recognition processing unit 20 are created by learning using forward waveform speech data (for example, “yamaha”, “hoNda”, “suzuki”, etc.) as learning data. On the other hand, the acoustic model 34 and the word dictionary 35 of the backward direction recognition processing unit 30 have the backward waveform voice data (that is, the voice data reproduced by being reversed on the time axis. For example, the above example is reversed. "Ahamay", "adNoh", "ikuzus" etc.) as learning data.

The collating units 23 and 33 collate the feature parameter series extracted by the speech analyzing units 21 and 31 with the recognition candidates of the acoustic libraries 22 and 32, and generate a likelihood representing the likelihood of the recognition candidate by the HMM algorithm. To do.
The integrated determination unit 40 comprehensively determines the outputs of the forward direction recognition processing unit 20 and the backward direction recognition processing unit 30, outputs a recognition result, and delivers it to the command data generation unit 12.

FIG. 3 is a flowchart for explaining a first processing example in the integrated determination unit 40.
The forward direction recognition processing unit 20 generates a forward likelihood LFi for the i-th word (recognition candidate) registered in the word dictionary 25. Similarly, the reverse direction recognition processing unit 30 generates a reverse likelihood LRi for the i-th word (recognition candidate) registered in the word dictionary 35. However, it is assumed that a plurality of common words are registered in the word dictionaries 25 and 35 in the same order.

In this case, the integrated determination unit 40 calculates a combined likelihood Li obtained by linearly combining the forward and backward likelihoods LFi, LRi according to the following equation (1) (step S1: likelihood information combining unit, weighted combining unit). ).
Li = αLFi + βLRi (1)
However, the weighting coefficients α and β are real numbers (constants), and α ≠ 0 and β ≠ 0. By appropriately determining these weighting coefficients α and β, the forward direction recognition result and the backward direction recognition result can be appropriately combined, and an appropriate recognition result can be obtained. The weighting coefficients α and β may be determined in advance by learning.

Further, the weighting coefficients α and β may be values αi and βi determined in advance for each word. The weighting coefficients αi and βi in this case may be stored in the word dictionaries 25 and 35 together with the recognition candidate words. In this case, in step S1, the joint likelihood Li is obtained according to the calculation by the following equation (1a) (recognition candidate adaptive weighting combining means).
Li = αi · LFi + βi · LRi (1a)
Next, the integrated determination unit 40 finds k that satisfies the following expression (2) (step S2).

k = arg _i max (Li) …… (2)
That is, the kth word is the word that has the maximum joint likelihood Li.
Further, the integration determining unit 40 determines whether or not the maximum joint likelihood Lk satisfies the condition of the following equation (3) (step S3).
Lk> Θ (3)
The threshold Θ is a real number, may be a constant, or may be a different value (Θk) depending on the word.

If the condition of the expression (3) is satisfied, that is, if the joint likelihood Lk exceeds a predetermined threshold value Θ, the integrated determination unit 40 uses the kth word as a recognition result, and data representing this word Is output (step S4). Otherwise, the recognition result is rejected and no data representing any word is output) Step S5).
In this way, in this example, the joint likelihood Li is obtained by weighting the forward and backward likelihoods LFi, LRi obtained by the forward direction recognition process and the backward direction recognition process with the weighting coefficients α, β and linearly combining them. Is required. Then, a recognition result is obtained based on the joint likelihood Li. In this way, by combining and using the recognition results in the forward direction and the reverse direction, it is possible to suppress variations related to ease of recognition.

  FIG. 4 is a flowchart for explaining a second processing example in the integrated determination unit 40. In FIG. 4, steps in which processing similar to the steps shown in FIG. 3 is performed are denoted by the same reference numerals as in FIG. 3. In this example, the integrated determination unit 40 performs determination using the output of the forward direction recognition processing unit 20 that is one of the pair of recognition processing units 20 and 30 with priority. That is, first, the integrated determination unit 40 uses the forward likelihood LFi generated by the forward direction recognition processing unit 20 to find j that satisfies the following expression (4) (step S10).

j = arg _i max (LFi) (4)
That is, the j-th word is the word that has the maximum forward likelihood LFi.
Further, the integration determination unit 40 determines whether or not the maximum forward likelihood LFj satisfies the condition of the following equation (5) (step S11, one-way recognition determination unit).
LFj> θ (5)
The threshold value θ is a real number and may be a constant or a different value (θj) depending on the word.

If the condition of the expression (5) is satisfied, that is, if the forward likelihood LFj exceeds the predetermined threshold θ, the integrated determination unit 40 regards the jth word as a recognition result (step S12). . Otherwise, the process (steps S1 to S5) described with reference to FIG. 3 is executed, and the recognition process based on the joint likelihood Li is executed.
In this way, when the determination can be made based only on the single processing result of the forward direction recognition processing unit 20, the processing using the joint likelihood Li can be omitted, so that the recognition processing can be speeded up and the voice It is possible to reduce the calculation load of the processing device constituting the recognition device 11.

Of course, it is also possible to preferentially use the single processing result of the backward direction recognition processing unit 30 and perform the same processing.
FIG. 5 is a flowchart for explaining a third processing example in the integrated determination unit 40. In FIG. 5, steps in which processing similar to the steps shown in FIG. 4 is performed are denoted by the same reference numerals as in FIG. 4.

In this example, when it is determined in step S11 that the maximum forward likelihood LFj exceeds the threshold θ (or θi), a determination according to the following equation (6) is further performed (step S15, one-way recognition). Determination means).
LFi ≦ θ (or θi) for ∀i ≠ j (6)
That is, for any i other than j, the j-th word is set as a recognition result on the condition that the forward likelihood LFi is equal to or less than the threshold θ (or θi) (step S12). That is, when only one word having a forward likelihood LFi exceeding the threshold θ (or θi) is determined, the j-th word is output as a recognized word without performing the determination based on the joint likelihood Li.

When the condition of the expression (6) is not satisfied (step S15: NO), that is, when there are two or more words having a forward likelihood exceeding the threshold θ (or θi), the above-described FIG. The processing described with reference (steps S1 to S5) is executed, and recognition processing based on the joint likelihood Li is executed.
However, in this case, it is preferable that the recognition process (steps S1 to S5) based on the joint likelihood Li is performed in a range of i that satisfies LFi> θ (or θi). Thereby, since the determination process based on the joint likelihood Li can be simplified, the processing speed can be increased, and the calculation load of the processing apparatus constituting the speech recognition apparatus 11 can be further reduced.

  Thus, even in this processing example, when the determination can be made based only on the single processing result of the forward direction recognition processing unit 20, the processing using the joint likelihood Li can be omitted. In addition, only when a single word can be identified based on the forward likelihood LFi, word recognition based on only the single processing result of the forward recognition processing unit 20 is allowed, so the certainty of the recognition result Can be increased.

Of course, as in the case of the processing example of FIG. 4, the same processing can be performed using the single processing result of the backward direction recognition processing unit 30 preferentially.
As indicated by a two-dot chain line in FIG. 5, in step S11, when the forward maximum likelihood LFj does not exceed the threshold θ (or θj), the determination process based on the combined likelihood Li is not performed. The recognition result may be rejected (step S5).

FIG. 6 is a flowchart for explaining a fourth processing example in the integrated determination unit 40. In FIG. 6, steps in which processing similar to the steps shown in FIG. 5 is performed are denoted by the same reference numerals as in FIG. 5.
In this example, instead of the determination in step S15 in FIG. 5, determination by the following equation (7) is performed (step S16, one-way recognition determination means).

LFj−LFi> Δ for ∀i ≠ j (7)
However, the threshold Δ is a real number (here, Δ> 0) and is a predetermined constant. A threshold value Δi for each word i may be used as the threshold value Δ.
The equation (7) represents a condition that the difference between the maximum forward likelihood LFj and any other forward likelihood LFi exceeds a threshold value Δ (or Δi). In other words, the difference between the maximum forward likelihood LFj and the next largest forward likelihood exceeds the threshold Δ (or threshold Δi), that is, the maximum forward likelihood LFj is compared with others. It is a condition that it protrudes.

If this condition is satisfied (YES in step S16), the j-th word is output as a recognized word without performing determination based on the joint likelihood Li (step S12).
When the condition of the expression (7) is not satisfied (step S15: NO), that is, when there are two or more words having the forward likelihood LFi exceeding the threshold θ (or θi), the above-described FIG. The processing described with reference to (Steps S1 to S5) is executed, and the recognition processing based on the joint likelihood Li is executed.

However, in this case, it is preferable that the recognition process (steps S1 to S5) based on the joint likelihood Li is performed in a range of i that satisfies LFi> θ (or θi). Thereby, since the determination process based on the joint likelihood Li can be simplified, the processing speed can be increased, and the calculation load of the processing apparatus constituting the speech recognition apparatus 11 can be further reduced.
Thus, even in this processing example, when the determination can be made based only on the single processing result of the forward direction recognition processing unit 20, the processing using the joint likelihood Li can be omitted. In addition, the word recognition based on only the single processing result of the forward recognition processing unit 20 is allowed only when it is recognized that the maximum forward likelihood LFj is prominent as compared with others. The certainty can be increased.

Of course, as in the case of the processing example of FIG. 4, the same processing can be performed using the single processing result of the backward direction recognition processing unit 30 preferentially.
As in the case of the processing example of FIG. 5, as shown by a two-dot chain line in FIG. 6, when the forward maximum likelihood LFj does not exceed the threshold θ (or θj) in step S11, the combined likelihood The recognition result may be rejected (step S5) without performing the determination process based on the degree Li.

FIG. 7 is a flowchart for explaining a fifth processing example in the integrated determination unit 40. In FIG. 7, steps in which the same processes as those in FIG. 4 are performed are denoted by the same reference numerals as those in FIG. 4.
In this example, when the determination based on the single processing result by the forward direction recognition processing unit 20 is preferentially performed, the weighted forward likelihood γi · LFi to which the weight γi is assigned for each word is used. That is, a word (j-th word) having the maximum weighted forward likelihood γi · LFi is obtained (step S20). That is,
j = arg _i max (γj · LFi) (8)
It is.

Then, it is determined whether or not the maximum weighted forward likelihood γj · LFj satisfies the following expression (9) (step S21: one-way recognition determination means; weight determination means).
γj ・ LFj> θ ...... (9)
The threshold value θ is a real number and is a constant. The threshold value θ may be a different value θj depending on the word, but since the likelihood LFi is weighted, there is no practical benefit of weighting the threshold value θ. For example, the weighting coefficient γi may be obtained in advance by learning and stored in the word dictionary 25 for each individual word.

If the condition of Equation (9) is satisfied, that is, if the weighted forward likelihood γj · LFj exceeds a predetermined threshold θ, the integrated determination unit 40 uses the jth word as a recognition result ( Step S12). Otherwise, the process (steps S1 to S5) described with reference to FIG. 3 is executed, and the recognition process based on the joint likelihood Li is executed.
In this way, by replacing the processes in steps S10 and S11 in FIG. 4 with the processes in steps S20 and S21 described above, the recognition by the weighted forward likelihood γi · LFi with the weight γi for each word can be performed. it can.

Similar modifications can be applied to the processes shown in FIGS.
FIG. 8 is a block diagram for explaining the configuration of a voice instruction apparatus according to the second embodiment of the present invention. 8, parts corresponding to the respective parts shown in FIG. 2 are given the same reference numerals as those in FIG.
In the first embodiment described above, the reverse reproduction unit 15 that inverts the input audio data on the time axis and reproduces and outputs the reverse audio data that is reversely reproduced by the reverse reproduction unit 15 is provided. Voice analysis processing is performed.

In contrast, in this embodiment, the voice analysis process is performed only on the forward direction voice data, and the voice analysis result is reversed on the time axis and delivered to the collation unit 33 of the backward direction recognition processing unit 30. I am doing so.
More specifically, the audio signal from the microphone 5 is given to the audio analysis unit 21 of the forward direction recognition processing unit 20 and the audio analysis unit 31 of the reverse direction recognition processing unit 30. The configuration of the forward direction recognition processing unit 20 is the same as that in the first embodiment described above.

The reverse direction recognition processing unit 30 includes a time axis inversion unit 37 that inverts the feature parameter series generated by the speech analysis unit 31 on the time axis and converts the feature parameter series into a reverse direction feature parameter series. The reverse direction feature parameter series generated by the time axis inversion unit 37 is given to the matching unit 33. Other configurations of the reverse direction recognition processing unit 30 are the same as those in the first embodiment.
Even with such a configuration, as in the case of the first embodiment, the recognition probability between words (recognition candidates) registered in the word dictionaries 25 and 35 using the forward direction recognition process and the backward direction recognition process in combination. The variation of can be suppressed.

However, learning when creating the acoustic model 34 and the word dictionary 35 constituting the acoustic library 32 of the backward direction recognition processing unit 30 is performed in the same processing procedure as the recognition processing (procedure for time axis inversion after speech analysis). There is a need.
In this embodiment, the speech recognition units 21 and 31 are provided in the forward direction recognition processing unit 20 and the backward direction recognition processing unit 30, respectively. However, one speech analysis unit is used as the forward direction and backward direction recognition processing unit 20, 30 may be shared.
[Experimental example]
FIG. 9 is a diagram showing recognition rates when voice data is reproduced both in the forward direction and in the reverse direction, and a word recognition experiment is performed for each. In the recognition for the forward reproduction voice data, the recognition rate is low for “word 6” and “word 9”, whereas in the recognition for the backward reproduction voice data, the recognition of both of those words exceeds 80%. The rate was obtained. On the other hand, in “word 11”, the recognition rate for the forward reproduction sound data is slightly better than the recognition for the reverse reproduction sound data.

From this experimental result, the word recognition rate is improved by the above-described embodiment in which the word recognition processing is performed by combining the forward direction recognition and the backward direction recognition. It is understood that it can be done.
[Modification]
In the above-described embodiment, HMM word recognition is taken as an example, but any method such as a DP matching method, a neural network method, a Bayes discriminant function method, or a vector quantization method can be applied as a recognition algorithm. In this case, as the likelihood information, not “likelihood” but “distance” between feature vectors may be used, and in the conditional expressions (3), (5), (6), (7), (9), etc. It may be necessary to reverse the direction of the inequality sign as appropriate.

Further, the present invention can be similarly applied to continuous speech recognition even if not word recognition. In this case, the recognition target is an utterance unit such as “character” or “syllable” instead of “word”.
Furthermore, in the above-described embodiment, the joint likelihood Li is a linear combination of the forward likelihood LFi and the reverse likelihood LRi. However, the joint likelihood Li may be represented by a non-linear combination of these, and in general, the joint likelihood Li The degree Li can be expressed as the following equation (9) by the function f of the forward likelihood LFi and the backward likelihood LRi.

Li = f (LFi, LLi) ...... (9)
An example in which the joint likelihood Li is determined by the non-linear combination of the forward and backward likelihoods LFi and LRi is as shown in the following equations (10) and (11).
Li = a · LFi ² + b · LFi + c · LRi ² + d · LRi (10)
Li = m · LFi ³ + n · LFi ² + p · LFi + q · LRi ³ + r · LRi ² + s · LRi …… (11)
However, a, b, c, d, m, n, p, q, r, and s are coefficients (real numbers) and may be constants, or different values are set in advance for each recognition candidate (word). You may keep it.

  In the above-described embodiment, both the forward direction recognition process and the backward direction recognition process are performed in any case, but the forward direction recognition process result is prioritized as in the processes illustrated in FIGS. In this case, a negative determination is made in steps S11, S15, S16, and S21 after the processing (speech analysis and collation processing in the case of the configuration shown in FIG. May be performed for the first time). Of course, the same applies to the case where priority is given to the result of the reverse direction recognition process, and the process related to the forward direction recognition (speech analysis and collation process) may be postponed.

  Further, in the processing examples of FIGS. 4 to 7 described above, only the determination regarding one of the forward likelihood LFi and the reverse likelihood LRi is performed before the determination using the combined likelihood Li. The determination regarding both of the above may be performed. Then, a word having a sufficiently large value of any one of the likelihoods LFi, LRi is found, a word having a likelihood LFi, LRi equal to or greater than a predetermined threshold is determined, and compared to other words It is also possible to determine the recognition result on the condition that the likelihoods LFi and LRi are sufficiently large, and omit the subsequent determination process based on the combined likelihood Li.

Furthermore, in the above-described embodiment, the voice instruction device 10 for the in-vehicle information device 50 is taken as an example. However, the present invention is not limited to the in-vehicle information device and is applied for voice instruction to other information devices. You can also
In addition to the various modifications described in the section “Means for Solving the Problems”, various design changes can be made within the scope of the matters described in the claims.

1 is a diagram showing an overall configuration of a vehicle system according to an embodiment of the present invention. It is a block diagram which shows the electric constitution of the said vehicle system. It is a flowchart for demonstrating the 1st process example in an integrated determination part. It is a flowchart for demonstrating the 2nd process example in an integrated determination part. It is a flowchart for demonstrating the 3rd process example in an integrated determination part. It is a flowchart for demonstrating the 4th process example in an integrated determination part. It is a flowchart for demonstrating the 5th process example in an integrated determination part. It is a block diagram for demonstrating the structure of the voice instruction apparatus which concerns on 2nd Embodiment of this invention. It is a figure which shows the recognition rate at the time of reproducing | regenerating audio | voice data in both the forward direction and a reverse direction, and performing word recognition experiment about each.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Car body 2 Crew 3 Helmet 5 Microphone 6 Speaker 7 Harness 10 Voice indication device 10
DESCRIPTION OF SYMBOLS 10A Voice instruction | indication apparatus main body 11 Voice recognition apparatus 12 Command data generation part 15 Reverse reproduction | regeneration part 20 Forward direction recognition process part 21 Voice analysis part 22 Acoustic library 23 Collation part 24 Acoustic model 25 Word dictionary 30 Reverse direction recognition process part 31 Voice analysis part 32 acoustic library 33 collation unit 34 acoustic model 35 word dictionary 37 time axis inversion unit 40 integration determination unit 50 in-vehicle information device 51 information processing unit

Claims (11)

Forward direction recognition means for recognizing an input speech signal in a forward direction according to an input time series, and generating forward likelihood information representing the likelihood that the input speech signal corresponds to a recognition candidate;
A reverse direction recognizing means for recognizing an input voice signal in a reverse direction according to a reverse time series opposite to the input time series, and generating reverse likelihood information representing the likelihood that the input voice signal corresponds to a recognition candidate;
An integrated determination unit that integrates outputs of the forward direction recognition unit and the backward direction recognition unit to generate a recognition result corresponding to an input voice signal;
The integrated determination unit combines forward likelihood information and backward likelihood information generated by the forward direction recognition unit and the backward direction recognition unit, respectively, and combines the input speech signal representing the likelihood of being a recognition candidate. A likelihood information combining unit that generates likelihood information, and a recognition result determining unit that evaluates the combined likelihood information generated by the likelihood information combining unit and obtains a recognition result,
The likelihood information coupling means, seen including weighting means for combining weighted respectively forward likelihood information and the backward likelihood information,
The weighting / combining means sets the weights depending on the recognition candidates, which are predetermined by learning for each recognition candidate so as to reduce the variation of the recognition probability between the recognition candidates, in the forward likelihood information and / or the reverse likelihood information. bonded by applying the recognition candidate adaptive weighting combiner means including speech recognition device.   The speech recognition apparatus according to claim 1, further comprising reverse reproduction means for reproducing an input voice signal according to a time series reverse to an input time series and inputting the input voice signal to the reverse direction recognition means. The reverse direction recognition means includes
Voice analysis means for analyzing an input voice signal according to an input time series and generating a feature series according to the input time series;
2. The voice according to claim 1, further comprising: a feature series inversion means for converting the feature series generated by the voice analysis means into a feature series according to a time series opposite to the input time series and inputting the feature series to the backward direction recognition means. Recognition device. The integrated determination means includes
Unidirectional recognition determination means for determining whether a single processing result by at least one of the forward direction recognition means and the backward direction recognition means indicates the presence of a recognition candidate having a certain likelihood or more.
And a means for outputting the recognition candidate as a recognition result when the one-way recognition determination means determines that the single processing result indicates the presence of a recognition candidate having a certain likelihood or more. speech recognition apparatus according to any one of claims 1-3. The one-way recognition determination means assigns a weight depending on a recognition candidate to the single processing result, and determines whether the single processing result indicates the presence of a recognition candidate having a certain likelihood or more. The speech recognition apparatus according to claim 4 , comprising means. The speech recognition device according to any one of claims 1 to 5 ,
Voice signal input means for inputting a voice signal to the voice recognition device;
A voice instruction device comprising: command data generating means for converting a recognition result by the voice recognition device into instruction command data for inputting to a predetermined device. A voice instruction device according to claim 6 ;
Information processing equipment including command processing means that operates in accordance with instruction command data generated by the command data generating means. A vehicle system in which the command processing means of the information device according to claim 7 is mounted on a vehicle body. A forward recognition step for recognizing an input speech signal in a forward direction according to an input time series, and generating forward likelihood information representing the likelihood that the input speech signal corresponds to a recognition candidate;
A reverse direction recognition step for recognizing an input voice signal in a reverse direction according to a reverse time series opposite to the input time series, and generating reverse likelihood information representing the likelihood that the input voice signal corresponds to a recognition candidate;
Integrating the results of the forward direction recognition step and the backward direction recognition step to generate a recognition result corresponding to the input speech signal; and
In the integration determination step, the forward likelihood information and the backward likelihood information respectively generated in the forward direction recognition step and the backward direction recognition step are combined to represent the likelihood that the input speech signal corresponds to the recognition candidate. A likelihood information combining step for generating combined likelihood information, and a recognition result determining step for evaluating the combined likelihood information generated in the likelihood information combining step to obtain a recognition result,
The likelihood information binding step, seen including a weighted combination step of combining by weighting each of the forward likelihood information and the backward likelihood information,
In the weighting and combining step, the forward likelihood information and / or the backward likelihood information are weights that depend on the recognition candidate and are determined in advance for each recognition candidate by learning so as to reduce variation in recognition probability between recognition candidates. recognition candidate adaptive weighted combination step of including, speech recognition method of bonding imparted to. The integration determination step includes
A one-way recognition determination step for determining whether a single processing result by at least one of the forward direction recognition step and the backward direction recognition step indicates the presence of a recognition candidate having a certain likelihood or more;
A step of outputting the recognition candidate as a recognition result when it is determined by the one-way recognition determination step that the single processing result indicates the presence of a recognition candidate having a certain likelihood or more. Item 10. The speech recognition method according to Item 9 . The one-way recognition determination step assigns a weight depending on a recognition candidate to the single processing result, and determines whether the single processing result indicates the presence of a recognition candidate having a certain likelihood or more. The speech recognition method according to claim 10 , comprising steps.