JP2003255972A - Speech recognizing device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve recognition precision even if a speech is broken during recognition of successive words. <P>SOLUTION: A matching means 5 performs a matching process by using a sound analysis result from a sound analyzing means 2, a vocabulary to be recognized from a recognized object vocabulary dictionary storage means 3, and a sound model from a sound model storage means 4, and computes the likelihood of each partial hypothesis showing each state of each vocabulary to be recognized and a next speech section wait decision means 6 sets a wait time up to a next speech section corresponding to the partial hypothesis with the maximum likelihood computed by the matching means 5, receives notice of speech section detection determination from a speech section detecting means 1, and instructs the matching means 5 to match the next speech section continuously when the next speech section is detected within the set wait time. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device which inputs a voice signal and outputs a recognition result.

[0002]

2. Description of the Related Art When performing voice recognition, a voice section of an input voice signal is detected, and the detected voice section is collated with a vocabulary to be recognized (hereinafter referred to as a recognition target vocabulary). As a general example, the one disclosed in Japanese Patent Laid-Open No. 59-21099797 is shown in FIG. 5, and FIG. 5 is a block diagram showing the configuration of such a voice recognition device.

In FIG. 5, 11 is a voice section detecting means for detecting a voice section of an input voice signal, and 12 is an acoustic analyzing means for performing an acoustic analysis on the voice signal of the voice section detected by the voice section detecting means 11. , 13 is a recognition target vocabulary dictionary storage unit that stores a recognition target vocabulary that is a recognition target and syntax information that defines a connection relationship between each recognition target vocabulary, and 14 is an acoustic model that is a minimum unit of recognition and waiting for each acoustic model. Acoustic model storage means for storing time information, 15 is an acoustic analysis result by the acoustic analysis means 12, recognition target vocabulary stored in the recognition target vocabulary dictionary storage means 13, and acoustics stored in the acoustic model storage means 14. It is a matching unit that performs matching using a model, calculates a likelihood, and outputs a recognition result.

Next, the operation will be described. The voice section detection means 11 detects the voice section of the input voice signal. Here, it is assumed that the voice section is detected by a predetermined threshold value of the power of the voice signal, for example. FIG. 6 is a diagram for explaining a start / end detection algorithm of a voice section. As shown in FIG. 6, the voice section detecting means 11 detects a section in which the power of the input voice signal is equal to or higher than a predetermined threshold as a voice section candidate, and the pause section between the voice section candidates has a predetermined threshold value, for example, If it is less than 350 msec, the two voice section candidates are detected as one voice section, and the voice signal of the detected voice section is output to the acoustic analysis unit 12.

The recognition target vocabulary stored in the recognition target vocabulary dictionary storage means 13 is, for example, “Tokyo”, “Kana garan”, “Kamakurashi”, “Kesenuma”, “Yukuhashi”, ... This is syntactic information that defines the connection relation of each recognition target vocabulary, such as the word "-" and the connection from "Kana garan" to "Kamakurashi" and from "Tokyo" to "Marunouchi". The recognition target vocabulary stored in the recognition target vocabulary dictionary storage unit 13 may be replaced for each recognition.

FIG. 7 shows an HMM (Hidden Marko).
v Model), in which the HMM in the case of being connected to “Kana ga ren” or “Kamakurashi”
Is shown. In the figure, each circle represents each state of the HMM and is stored in the acoustic model storage means 14, and an arrow represents each transition and is stored in the recognition target vocabulary dictionary storage means 13.
Further, / L1 / represents an acoustic model corresponding to a silent section at the beginning of a word (before utterance), / L2 / represents an acoustic model corresponding to a silent section at the end of a word (after utterance), and / L3 / represents between words (speaking). It represents an acoustic model corresponding to the silence interval of (middle).

The acoustic analysis means 12 is a voice section detection means 1
The voice signal in the voice section detected in 1 is cut out at a constant length (frame length) at regular intervals (frame periods),
The cut-out voice data (voice data for each frame) is analyzed, and the time-series data that is the acoustic analysis result is output to the matching unit 15.

The matching means 15 uses the acoustic analysis result by the acoustic analysis means 12, the recognition target vocabulary stored in the recognition target vocabulary dictionary storage means 13, and the acoustic model stored in the acoustic model storage means 14. Then, the likelihood in the final state of all the recognition target vocabularies is obtained, and the recognition target vocabulary having the maximum likelihood in the final state is output as the recognition result.

Here, the matching means 15 obtains the likelihood by performing the following calculation, for example. The acoustic model corresponding to the nth state of the recognition target vocabulary l stored in the recognition target vocabulary dictionary storage unit 13 is dic (l, n), and the recognition target vocabulary l is the nth at time (frame) t. In this case, the likelihood of one frame of the analysis result at this time is lklhd (l, t, n).

FIG. 8 is a diagram showing an example of a recognition path for the recognition target vocabulary 1, and shows an input time (frame) and a route in a state of a certain recognition target vocabulary 1. Although there are a plurality of possible routes, FIG. 8 shows one of them, and the likelihood when the state at the frame t is the n-th is lklhd (l, t, n) for the recognition target vocabulary l. I am trying.

For example, the likelihood when the route as shown in FIG. 8 is taken is calculated by the following equation. Assuming that the length of the voice section detected by the voice section detecting means 11 is T frames, the cumulative likelihood Lklhd ′ (l, n) for a certain route up to the nth state of the recognition target vocabulary l is expressed.

[Equation 1] Here, k (t) indicates the order of the state assigned to the frame t.

Here, the cumulative likelihood Lklhd (l, n), which is the maximum likelihood in the route reaching the nth state of the recognition target vocabulary l for this input speech, is expressed by the following equation.

[Equation 2] Further, the likelihood LK (l) of the recognition target vocabulary l is N in the final state.
If (l), then LK (l) = Lklhd (l, N (l)) (3). Maximum likelihood LK (L) in the likelihood of each recognition target vocabulary
The recognition target vocabulary L is output as the recognition result.

[Equation 3]

As described above, in the conventional voice recognition device shown in FIG. 5, the voice section of the voice signal input by the voice section detection means 11 is detected, and the acoustic analysis means 12 is detected by the voice section detection means 11. The voice signal in the voice section is cut out for each frame, the cut-out voice data for each frame is analyzed, and the analysis result is output to the matching means 15. The matching means 15 stores in the recognition target vocabulary dictionary storage means 13 for each frame. The recognition target vocabulary stored and the acoustic model storage means 1
Matching is performed using the acoustic model stored in 4,
The likelihoods of all recognition target vocabularies in the final state are obtained, and the recognition target vocabulary that takes the maximum likelihood in the final state is output as the recognition result.

[0014]

Since the conventional speech recognition apparatus is constructed as described above, if the speech is interrupted by a pause when recognizing consecutive words, the speech following it cannot be recognized. There was a problem that the recognition accuracy deteriorates. In order to deal with this, if a method of waiting for the input of the next voice section for a certain period of time is used, the response time from the utterance to the output of the voice recognition result is delayed, and the noise in the pause section following the voice is reduced. There is a problem that the recognition accuracy is deteriorated by picking up and recognizing.

The present invention has been made in order to solve the above problems, and provides a voice recognition device capable of improving the recognition accuracy even when the voice is interrupted due to a pause or the like when recognizing continuous words. The purpose is to get.

[0016]

In the voice recognition device according to the present invention, the next voice section wait determination means sets and sets the waiting time until the next voice section corresponding to the partial hypothesis having the maximum likelihood. When the next voice section is detected within the waiting time, the collating means is instructed to continue the collation of the next voice section.

In the speech recognition apparatus according to the present invention, the next speech segment waiting determination means waits for the next speech segment in correspondence with the syntactic position of the partial hypothesis having the maximum likelihood or the position on the recognition target vocabulary. It sets the time.

In the voice recognition device according to the present invention, the next voice section wait determination means outputs the final result as a recognition result with the waiting time set to 0 when the partial hypothesis having the maximum likelihood is the final result. is there.

In the voice recognition apparatus according to the present invention, the next voice section wait determination means sets a first waiting time until the next voice section when the partial hypothesis having the maximum likelihood is an intermediate result. When the next voice section is detected within the set first waiting time, the collating means is instructed to continue the collation for the next voice section, and the next speech section is detected within the set first waiting time. If no voice section is detected, the intermediate result is output as the recognition result.

In the speech recognition apparatus according to the present invention, the next speech segment waiting determination means, when the partial hypothesis having the maximum likelihood is in the intermediate state, the second speech segment until the next speech segment shorter than the first waiting time. When the next voice section is detected within the set second wait time, the collating means is instructed to continue the collation of the next voice section, and the set second wait time is set. When the next voice section is not detected within the waiting time of, the intermediate state is output as the recognition result or no recognition result is output.

In the voice recognition device according to the present invention, the next voice section wait determination means sets the waiting time until the next voice section in correspondence with the last acoustic model of the partial hypothesis having the maximum likelihood. .

In the speech recognition apparatus according to the present invention, when the next speech section waiting determination means does not detect the next speech section within the set waiting time, the partial hypothesis with the maximum likelihood becomes the recognition target vocabulary. It judges whether or not it can be adopted as a recognition result based on the syntax information that defines the connection relation, and outputs the partial hypothesis with the maximum likelihood as the recognition result.

In the speech recognition apparatus according to the present invention, when the next speech section waiting determination means cannot adopt the partial hypothesis having the maximum likelihood as the recognition result, the last sound of the partial hypothesis with the next highest likelihood is detected. The waiting time until the next voice section is set according to the model.

In the voice recognition apparatus according to the present invention, the next voice section wait determination means inputs the likelihood of each partial hypothesis from the matching means and the syntactic position of each partial hypothesis or the position on the recognition target vocabulary. Then, the waiting time until the next speech section is set corresponding to the position on the syntactical hypothesis or the position on the recognition target vocabulary with the maximum likelihood, and the next speech section is detected within the set waiting time. In this case, the matching means is instructed to continue matching the next speech section, and if the next speech section is not detected within the set waiting time, the partial hypothesis with the maximum likelihood is recognized as the recognition result. Is output as.

In the speech recognition apparatus according to the present invention, the next speech section wait determination means uses the likelihood of each of the above partial hypotheses from the matching means and the waiting time information of each acoustic model stored in the acoustic model storage means. , The syntactic information that defines the connection relation of each recognition target vocabulary stored in the recognition target vocabulary dictionary storage unit, and the waiting time of the last acoustic model of the partial hypothesis that is the maximum likelihood If the waiting time to the section is set and the next voice section is detected within the set waiting time, the collating means is instructed to continue the collation of the next voice section, and the waiting time is less than the set waiting time. If the next speech segment is not detected in, it is judged whether the partial hypothesis with the maximum likelihood can be adopted as the recognition result by the syntax information, and the partial hypothesis with the maximum likelihood is output as the recognition result. .

[0026]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the present invention will be described below. Embodiment 1. 1 is a block diagram showing a configuration of a voice recognition device according to a first embodiment of the present invention. In FIG. 1, reference numeral 1 indicates a voice section of an input voice signal and outputs a voice signal of the detected voice section. A voice section detection means for outputting a voice section confirmation notification indicating that the voice section has been confirmed; 2 an acoustic analysis means for performing acoustic analysis on the voice signal of the voice section detected by the voice section detection means 3;
Is a recognition target vocabulary dictionary storage unit that stores a recognition target vocabulary to be recognized and syntax information that defines a connection relationship between each recognition target vocabulary, and 4 is an acoustic model as a minimum unit of recognition and waiting time information of each acoustic model. Is an acoustic model storage means for storing

Further, in FIG. 1, 5 is an acoustic analysis means 2.
Acoustic analysis result, the recognition target vocabulary stored in the recognition target vocabulary dictionary storage unit 3, and the acoustic model storage unit 4
Each of the recognition target vocabularies stored in the recognition target vocabulary dictionary storage means 3 is compared with each other by using the acoustic model stored in the recognition target vocabulary to calculate the likelihood in each partial hypothesis indicating each state of each recognition target vocabulary. It is a collation unit that obtains the syntactic position of each partial hypothesis or the position on the recognition target vocabulary from the syntactic information that defines the connection relation of.

Further, in FIG. 1, reference numeral 6 is a portion for inputting the likelihood of each partial hypothesis from the collating means 5 and the syntactic position of each partial hypothesis or the position on the recognition target vocabulary to obtain the maximum likelihood. The waiting time until the next speech section is set in correspondence with the position on the syntactical hypothesis or the position on the recognition target vocabulary, and upon receipt of the speech section detection confirmation notification from the speech section detection means 1, less than the set waiting time. If the next voice segment is detected in
The collating means 5 is instructed to continue the collation of the next speech section, and if the next speech section is not detected within the set waiting time, the partial hypothesis having the maximum likelihood is output as the recognition result. It is a voice section waiting determination means.

As described above, the acoustic analysis unit 2, the recognition target vocabulary dictionary storage unit 3, and the acoustic model storage unit 4 are the conventional acoustic analysis unit 12, recognition target vocabulary dictionary storage unit 13, and acoustic model storage unit shown in FIG. It is equivalent to 14.

Next, the operation will be described. As shown in FIG. 6, the voice section detecting means 1 detects a section in which the power of the input voice signal is equal to or more than a predetermined threshold as a voice section candidate, and the pause section between the voice section candidates has a predetermined threshold value, for example, If it is less than 350 msec, the two voice section candidates are detected as one voice section, and the voice signal of the detected voice section is output to the acoustic analysis unit 2. Further, the voice section detection means 1 detects the start of the voice section, and the power of the voice signal having a predetermined threshold value or more is kept for a predetermined time, for example, 50 mse.
When c continues, the voice section is decided to be a voice section, and a voice section confirmation notice is output to the next voice section wait determination means 6.

The acoustic analysis means 2 performs acoustic analysis on the voice signal of the voice section detected by the voice section detection means 1 as in the conventional case. That is, the acoustic analysis unit 2 cuts out the voice signal of the voice section detected by the voice section detection unit 1 at a frame length for each frame period, analyzes the cut-out voice data of each frame, and outputs the result of the acoustic analysis. The time series data is output to the matching means 5.

The conventional collating means 15 outputs the recognition target vocabulary having the maximum likelihood as the final state as the recognition result, whereas the collating means 5 outputs all recognitions stored in the recognition target vocabulary dictionary storage means 3. Using the acoustic model stored in the acoustic model storage unit 4 corresponding to each state of the HMM of the target vocabulary and the acoustic analysis result by the acoustic analysis unit 2, the likelihood in each partial hypothesis indicating each state of each recognition target vocabulary. The degree is calculated, and the syntactical position of each partial hypothesis or the position on the recognition target vocabulary is obtained from the syntax information defining the connection relation of each recognition target vocabulary stored in the recognition target vocabulary dictionary storage unit 3, The likelihood in each partial hypothesis and the syntactical position of each partial hypothesis or the position on the recognition target vocabulary are output to the next speech section waiting determination means 6.

Here, in the calculation of the likelihood in each partial hypothesis, for example, in FIG. 7, the likelihood in the partial hypotheses up to / ka /, the likelihood in the partial hypotheses up to / ka /, / na /, / ka / , / Na /, / ga / such as the likelihood in the partial hypothesis are sequentially calculated, and / ka /, / n
a /, / ga /, / wa /, / ke /, / N /, / L3
/, / Ka /, / ma /, / ku /, / ra /, / si
Likelihoods in the partial hypotheses up to /, / L2 / are calculated.

The next voice section wait determination means 6 is the collation means 5
Of the partial hypotheses received from, the waiting time until the next speech section is set corresponding to the position on the syntactical position or the recognition target vocabulary of the partial hypothesis having the maximum likelihood, and the speech section detection means 1 When the next voice section is detected within the set waiting time after receiving the voice section detection confirmation notification of
When the collating means 5 is instructed to continue the collation of the next speech section and the next speech section is not detected within the set waiting time, the partial hypothesis having the maximum likelihood is output as the recognition result.

FIG. 2 is a flow chart showing the judgment processing of the next voice section waiting judgment means. Here, in the case where the partial hypothesis having the maximum likelihood is in a position where there is no succeeding vocabulary in the syntax or the recognition target vocabulary. Is set as the final result, the partial hypotheses at positions previously designated as poses in the recognition target vocabulary dictionary storage unit 3 are set as intermediate results, and the partial hypotheses at other positions are set as intermediate states. That is, in the example of FIG. 7, when the partial hypothesis having the maximum likelihood is located at the position of / L2 /, it is the final result, and when it is at the position of / L3 /, it is the intermediate result, and / L1 /, / L2 If the position is other than /, / L3 /, the intermediate state is set.

The next voice section wait decision means 6 corresponds to the position on the syntactic or recognition target vocabulary where the partial hypothesis having the maximum likelihood corresponds to from the end of the verified voice section to the start of the next voice section. The waiting time is set. For example, the next voice section wait determination unit 6 sets the waiting time to 0 because it is not necessary to continuously recognize the subsequent voice section when the partial hypothesis having the maximum likelihood is the final result. If it is an intermediate result, the waiting time ThTime1 is set to, for example, 3 seconds, and if it is an intermediate result, the waiting time ThTime2 is set to, for example, 1 second.

In step ST11 of FIG. 2, the next speech section waiting determination means 6 receives from the matching means 5 the likelihood in the partial hypothesis, the syntactic position of the partial hypothesis, or the position in the recognition target vocabulary. In step ST12, the partial hypothesis having the maximum likelihood among the received partial hypotheses is set as the determination partial hypothesis. In step ST13, it is determined whether the partial hypothesis for judgment is the final result from the received syntactical position of each partial hypothesis or the position on the recognition target vocabulary. Since the time is 0, the judgment partial hypothesis is immediately output as the recognition result in step ST14.

If the judgment partial hypothesis is not the final result in step ST13, it is judged in step ST15 whether the judgment partial hypothesis is an intermediate result. If it is an intermediate result, the process proceeds to step ST16 and the intermediate result is judged. If not, that is, if it is in the intermediate state, the process proceeds to step ST19. When the judgment partial hypothesis is an intermediate result, in step ST16, upon receipt of the voice section confirmation notification from the voice section detecting means 1, the waiting time from the end of the verified voice section to the start of the next voice section. Is less than ThTime1, eg T
Check if hTime1 = less than 3 seconds.

In step ST16, when the waiting time ThTime1 has elapsed without detecting the next voice section,
In step ST17, the intermediate result is output as the recognition result as the timeout process. On the other hand, step ST
If the next voice section is detected at less than ThTime1 at 16, the collating means 5 is instructed to continue the recognition from the previous collation state at step ST18.

Here, in the case of the intermediate result, the waiting time (ThTime1) of the next voice section is set to be a long time, for example, 3 seconds, when the partial hypothesis is the intermediate result, the pause (silent section) is originally set. ) Is expected to be inserted, so there is a high possibility that the silent section is long.

If the judgment partial hypothesis is in the intermediate state in step ST15, in step ST19,
Upon receiving the voice section confirmation notification from the voice section detecting means 1,
The waiting time from the end of the verified voice section to the start of the next voice section is less than ThTime2, for example ThTime.
e2 = 1 Check if it is less than 1 second, ThTime
If less than 1 and the next voice section is detected, step ST
At 18, the collating means 5 is instructed to continue the recognition from the previous collation state.

If it is determined in step ST19 that the waiting time has exceeded the predetermined time ThTime2 without detecting the next voice section, a timeout process is performed in step ST20. In this time-out process, the maximum likelihood that can be output as the recognition result among the partial hypotheses in the previous speech section is output as the recognition result. Further, in this timeout processing, it may be rejected because there is no recognition result.

Here, the waiting time ThTime2 of the next voice section in the intermediate state is set to be shorter than the waiting time ThTime1 in the intermediate result, because a pause such as a sentence or word break is assumed in advance. This is because it is assumed that a pose that is not that long compared to the place where it is done is not included.

In the first embodiment, for convenience of explanation, the syntactical position or the position on the recognition target vocabulary has three types of final result, intermediate result, and intermediate state. For example, FIG. "Kanagawa" of the recognition target vocabulary "Kanagawa"
You may make more detailed settings, such as setting another waiting time between "ken" and "ken".

As described above, according to the first embodiment, the waiting time until the next speech section is associated with the syntactic position of the partial hypothesis having the maximum likelihood or the position on the recognition target vocabulary. By setting, it is possible to improve the recognition accuracy even when the voice is interrupted due to a pause or the like when recognizing continuous words.

Further, according to the first embodiment, when the utterance is completed up to the end of the syntax, the response time from the utterance to the output of the voice recognition result can be shortened. To be

Embodiment 2. 3 is a block diagram showing the configuration of a voice recognition device according to a second embodiment of the present invention. In the figure, reference numeral 5 a is used to perform matching using the acoustic analysis result by the acoustic analysis unit 2, the recognition target vocabulary stored in the recognition target vocabulary dictionary storage unit 3, and the acoustic model stored in the acoustic model storage unit 4. It is a matching means for performing the likelihood calculation in each partial hypothesis indicating each state of each recognition target vocabulary.

Further, in FIG. 3, 6a is a collating means 5a.
The likelihood of the partial hypothesis from, the waiting time information of each acoustic model from the acoustic model storage unit 4, and the syntax information defining the connection relation of each recognition target vocabulary from the recognition target vocabulary dictionary storage unit 3 are input. , The waiting time until the next voice section is set based on the waiting time information of the last acoustic model of the partial hypothesis having the maximum likelihood, the voice section detection confirmation notification from the voice section detecting means 1 is received, and the set waiting time is set. When the next voice section is detected within the time, the collating means 5a is instructed to continue the collation of the next voice section, and when the next voice section is not detected within the set waiting time, It is a next voice section waiting determination unit that determines whether the partial hypothesis having the maximum likelihood can be adopted as a recognition result based on the syntax information from the recognition target vocabulary dictionary storage unit 3 and outputs the recognition result.

Further, in FIG. 3, the voice section detection means 1, the acoustic analysis means 2, the recognition target vocabulary dictionary storage means 3 and the acoustic model storage means 4 are equivalent to the configuration shown in FIG. 1 of the first embodiment.

Next, the operation will be described. In the first embodiment, the waiting time until the next speech section is set in correspondence with the syntactic position of the partial hypothesis having the maximum likelihood or the position on the recognition target vocabulary. In 2, corresponding to the last acoustic model of the partial hypothesis with maximum likelihood,
The waiting time until the next voice section is set.

For example, in FIG. 7, the waiting time of the voice section is set to 1 second for the acoustic model / L1 / corresponding to the silent section at the beginning of the word (before utterance), and it corresponds to the silent section at the end of the word (after utterance). The waiting time of the voice section is 0 seconds for the acoustic model / L2 /, and the waiting time of the voice section is 3 seconds for the acoustic model / L3 / corresponding to the silent section between words (during utterance). For other acoustic models such as / ka /, / na /, etc., 1 second or the like is set. Further, for example, 2 seconds is set for an acoustic model corresponding to a whisper, and further, an acoustic model that is apt to be mistakenly determined to be a silent section by voice section detection in a noisy environment, for example, “shi”, “hi”, which is easily devoiced,
For “fu”, “chi”, etc., it is set to 1.5 seconds, for example.
These waiting times are stored in the acoustic model storage means 4 as waiting time information of each acoustic model.

FIG. 4 is a flow chart showing the judgment processing of the next voice section waiting judgment means. In step ST21, the next voice section wait determination means 6a receives from the matching means 5a,
Receive the likelihood for each partial hypothesis. Step ST22
In, the partial hypothesis that takes the maximum likelihood among the received partial hypotheses is the determination partial hypothesis.

In step ST23, the waiting time information of the final acoustic model p of the judgment partial hypothesis is extracted from the acoustic model storage means 4, and the waiting time ThTime (p) of the next speech section is calculated from the extracted waiting time information. Set.
For example, in FIG. 7, when the last acoustic model is / L2 /, ThTime (p) = 0 seconds is set, and / L3 /
In this case, ThTime (p) = 3 seconds is set.

In step ST24, upon receipt of the voice section confirmation notification from the voice section detecting means 1, the waiting time from the end of the verified voice section to the start of the next voice section exceeds ThTime (p). Check if not. At step ST24, the waiting time ThTime
If the next voice section is detected under (p), the collating means 5a is instructed to continue the recognition from the previous collating state in step ST25.

On the other hand, in step ST24, when the waiting time ThTime (p) has elapsed without the detection of the next voice section, the next step ST26 to step ST29.
Timeout process up to. As this timeout processing, for example, the following processing is performed.

In step ST26, it is determined whether the determination partial hypothesis can be adopted as a recognition result.
Whether or not the judgment partial hypothesis can be adopted as a recognition result is stored in the recognition target vocabulary dictionary storage means 3 as syntax information that defines the connection relation of each recognition target vocabulary. As the syntactic information, for example, in FIG. 7, only the partial hypotheses reaching the acoustic model / L2 / corresponding to the silent section of the ending (after utterance) is adopted as the recognition result, or the ending (after utterance) silence is used. Acoustic model / L2 / corresponding to a section, or acoustic model / L3 corresponding to a silent section between words (during vocalization)
The partial hypothesis reaching / is adopted as the recognition result.

If the result of determination in step ST26 can be adopted as the recognition result, in step ST27
The judgment partial hypothesis is output as the recognition result. Step S
If the determination result of T26 cannot be adopted as the recognition result, it is checked in step ST28 whether or not there is a partial hypothesis with the next highest likelihood of the determination partial hypothesis.

If there is a partial hypothesis with the next highest likelihood in step ST28, the partial hypothesis with the next highest likelihood is set as a new partial hypothesis for determination in step ST29, and the process returns to step ST23 to The process of is repeated.
On the other hand, in step ST28, if there is no partial hypothesis with the next highest likelihood, the recognition result is rejected and the process ends.

Here, there is a case where the partial hypothesis with the next highest likelihood does not exist, because the calculation amount of all the partial hypotheses is enormous, and the maximum likelihood is calculated for each frame by a method called beam search. The calculation amount is reduced by not calculating the partial hypotheses that have a certain likelihood difference from the degree, or by only calculating the partial n hypotheses from the maximum likelihood to the top n. is there.

As described above, according to the second embodiment, by setting the waiting time until the next speech section in correspondence with the last acoustic model of the partial hypothesis having the maximum likelihood, continuous processing is performed. When recognizing a word, even if the voice is interrupted due to a pause, etc., the recognition accuracy can be improved, and it is possible to cope with different pauses and different silence intervals due to jumbles, etc. It is possible to deal with a case where a silent section is erroneously detected, and the recognition accuracy can be improved.

Further, according to the second embodiment, by setting the waiting time until the next speech section in correspondence with the last acoustic model of the partial hypothesis having the maximum likelihood, the portion having the maximum likelihood is obtained. It is necessary to set the waiting time in detail when changing the syntax or the recognition target vocabulary, rather than setting the time for the next voice section wait in accordance with the position of the hypothesis in the syntax or the position of the recognition target vocabulary. The effect of disappearing is obtained.

It goes without saying that each means of the speech recognition apparatus of the above-mentioned first and second embodiments can be constituted by either hardware or software. Further, when the software is used, a medium in which the software is recorded is required.

[0063]

As described above, according to the present invention, the next voice section wait determination means sets the waiting time until the next voice section in accordance with the partial hypothesis having the maximum likelihood, and the set wait time is set. If the next voice segment is detected within the time, the collating means is instructed to continue the collation of the next voice segment so that the voice may be interrupted by a pause when recognizing consecutive words. Also, there is an effect that the recognition accuracy can be improved.

According to the present invention, the next speech segment waiting determination means sets the waiting time until the next speech segment in correspondence with the syntactic position of the partial hypothesis having the maximum likelihood or the position on the recognition target vocabulary. By doing so, when recognizing continuous words, even if the voice is interrupted due to a pause or the like, there is an effect that the recognition accuracy can be improved.

According to the present invention, when the next voice section wait determination means determines that the partial hypothesis having the maximum likelihood is the final result,
By outputting the final result as the recognition result with the waiting time set to 0, the response time from the utterance to the output of the speech recognition result can be shortened when the utterance is completed up to the end of the syntax. There is.

According to the present invention, when the next voice section wait determination means determines that the partial hypothesis having the maximum likelihood is an intermediate result,
The first waiting time until the next voice section is set, and when the next voice section is detected within the set first waiting time, the collating means continues to collate the next voice section. If the next voice section is not detected within the set first waiting time, the intermediate result is output as the recognition result, and the voice is interrupted by a pause when recognizing continuous words. However, there is an effect that the recognition accuracy can be improved.

According to the present invention, the next voice section wait determination means, when the partial hypothesis having the maximum likelihood is in the intermediate state, waits for the second voice section until the next voice section shorter than the first wait time. When the time is set and the next voice section is detected within the set second waiting time, the collating means is instructed to continue the collation of the next voice section, and the set second waiting time is set. If the next voice segment is not detected within the time, the intermediate state is output as the recognition result, or the absence of the recognition result is output to interrupt the voice in the intermediate state when recognizing consecutive words. Also, there is an effect that the recognition accuracy can be improved.

According to the present invention, the next speech segment waiting determination means sets the waiting time until the next speech segment corresponding to the last acoustic model of the partial hypothesis having the maximum likelihood, so that consecutive words When recognizing a voice, even if the voice is interrupted due to a pause, etc., the recognition accuracy can be improved, and it is possible to deal with different silence intervals that are assumed due to pauses, jumbles, etc. Therefore, it is possible to deal with the case where the silent section is mistakenly set, and the recognition accuracy can be improved.

According to the present invention, when the next voice section wait determination means does not detect the next voice section within the set waiting time, the partial hypothesis having the maximum likelihood indicates the connection relation of each recognition target vocabulary. It is possible to improve the recognition accuracy by determining whether the recognition result can be adopted as the recognition result based on the defined syntactic information and outputting the partial hypothesis having the maximum likelihood as the recognition result.

According to the present invention, when the next speech segment wait judgment means cannot adopt the partial hypothesis having the maximum likelihood as the recognition result, it corresponds to the last acoustic model of the partial hypothesis with the next highest likelihood. By setting the waiting time until the next voice section, even if the voice is interrupted by a pause when recognizing consecutive words, it is possible to improve the recognition accuracy, and also assume a pause or a whisper. It is possible to deal with different silent sections that are generated, and to deal with a case where a voice that is easily devoiced is erroneously made into a silent section by voice section detection, and it is possible to improve recognition accuracy.

According to the present invention, the next voice section waiting judgment means inputs the likelihood of each partial hypothesis from the matching means and the syntactic position of each partial hypothesis or the position on the recognition target vocabulary,
When the waiting time to the next speech section is set corresponding to the position on the syntactical position or the recognition target vocabulary of the maximum likelihood and the next speech section is detected within the set waiting time. , The collating means is instructed to continue to collate the next speech segment, and if the next speech segment is not detected within the set waiting time, the partial hypothesis with the maximum likelihood is output as the recognition result. By doing so, when recognizing continuous words, even if the voice is interrupted due to a pause or the like, there is an effect that the recognition accuracy can be improved.

According to the present invention, the next voice section wait determination means is the likelihood of each partial hypothesis from the matching means, the waiting time information of each acoustic model stored in the acoustic model storage means, and the recognition target vocabulary. Input the syntactic information that defines the connection relation of each recognition target vocabulary stored in the dictionary storage means, and wait until the next speech section by the waiting time information of the last acoustic model of the partial hypothesis with the maximum likelihood. When the time is set and the next voice section is detected within the set waiting time, the collating means is instructed to continue the collation of the next voice section, and the next voice section is set within the set waiting time. If no interval is detected, it is determined whether the partial hypothesis with the maximum likelihood can be adopted as the recognition result based on the syntactic information, and the partial hypothesis with the maximum likelihood is output as the recognition result to recognize consecutive words. Let In this case, the recognition accuracy can be improved even if the voice is interrupted due to a pause, etc. There is an effect that it is possible to cope with the case where there is no sound section and the recognition accuracy can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a voice recognition device according to a first embodiment of the present invention.

FIG. 2 is a flowchart showing a determination process of a next voice section wait determination means of the voice recognition device according to the first embodiment of the present invention.

FIG. 3 is a block diagram showing a configuration of a voice recognition device according to a second embodiment of the present invention.

FIG. 4 is a flowchart showing a determination process of a next voice section wait determination means of the voice recognition device according to the second embodiment of the present invention.

FIG. 5 is a block diagram showing a configuration of a conventional voice recognition device.

FIG. 6 is a diagram illustrating a start / end detection algorithm of a voice section.

FIG. 7 is a diagram showing an example of an HMM.

FIG. 8 is a diagram showing an example of a recognition path for a recognition target vocabulary.

[Explanation of symbols]

1 voice section detection means, 2 acoustic analysis means, 3 recognition target vocabulary dictionary storage means, 4 acoustic model storage means, 5, 5
a collating means, 6 and 6a next voice section waiting determination means.

Claims (10)

[Claims] 1. A voice section detecting unit for detecting a voice section of an input voice signal and outputting a voice section confirmation notification indicating that the voice section has been fixed, and a voice section detected by the voice section detecting unit. Acoustic analysis means for performing acoustic analysis on a voice signal; recognition target vocabulary dictionary storage means for storing recognition target vocabulary to be recognized; acoustic model storage means for storing an acoustic model as a minimum unit of recognition; Matching is performed using the acoustic analysis result by the acoustic analysis unit, the recognition target vocabulary stored in the recognition target vocabulary dictionary storage unit, and the acoustic model stored in the acoustic model storage unit, and each recognition target vocabulary When the waiting time until the next speech section is associated with the matching means for calculating the likelihood in each partial hypothesis indicating each state of, and the partial hypothesis with the maximum likelihood calculated by the matching means When the next voice section is detected within the set waiting time after receiving the voice section detection confirmation notification from the voice section detection means, the next voice section is continued to the matching means. A voice recognition apparatus comprising: a next voice section wait determination means for instructing to perform matching. 2. The next speech segment waiting determination means sets the waiting time until the next speech segment in correspondence with the position on the syntactical position of the partial hypothesis or the recognition target vocabulary that has the maximum likelihood. The voice recognition device according to claim 1. 3. The next voice section wait determination means sets the waiting time to 0 when the partial hypothesis having the maximum likelihood is the final result in a position where no vocabulary following it is present in the syntactic or vocabulary to be recognized. The voice recognition device according to claim 2, wherein the final result is output as a recognition result. 4. The next voice section wait determination means determines the first waiting time until the next voice section in the case of an intermediate result in which the partial hypothesis having the maximum likelihood is in a position designated in advance as a pause. When the next voice section is detected within the set and the set first waiting time, the collating means is instructed to continue the collation of the next voice section, and the set first waiting time is set. The speech recognition apparatus according to claim 2, wherein the intermediate result is output as a recognition result when the next speech section is not detected within a time period. 5. The next speech section wait determining means is an intermediate except for an intermediate result other than the final result in which the partial hypothesis having the maximum likelihood is in a position where no following vocabulary exists in the syntax or the recognition target vocabulary. In the case of the state, the second waiting time until the next voice section shorter than the first waiting time is set, and when the next voice section is detected within the set second waiting time, If the collating means is instructed to continue the collation of the next voice section and the next voice section is not detected within the set second waiting time, the intermediate state is output as a recognition result, or Alternatively, the voice recognition device according to claim 4, which outputs no recognition result. 6. The next speech segment waiting determination means sets a waiting time until the next speech segment corresponding to the last acoustic model of the partial hypothesis having the maximum likelihood.
The voice recognition device described. 7. The next voice section wait determination means, when the next voice section is not detected within the set waiting time, the partial hypothesis with the maximum likelihood defines the connection information of each recognition target vocabulary. 7. The voice recognition apparatus according to claim 6, wherein the partial hypothesis having the maximum likelihood is output as a recognition result by determining whether or not it can be adopted as a recognition result. 8. The next speech section waiting determination means, when a partial hypothesis having the maximum likelihood cannot be adopted as a recognition result, corresponds to the last acoustic model of the partial hypothesis with the next highest likelihood, The voice recognition device according to claim 7, wherein a waiting time until a voice section is set. 9. A voice section detecting unit that detects a voice section of an input voice signal and outputs a voice section confirmation notification indicating that the voice section has been fixed, and a voice section detected by the voice section detecting unit. An acoustic analysis means for performing acoustic analysis on a speech signal, a recognition target vocabulary dictionary storage means for storing recognition target vocabulary to be recognized and syntax information defining connection relationships between each recognition target vocabulary, and a minimum unit of recognition The acoustic model storage means for storing an acoustic model, the acoustic analysis result by the acoustic analysis means, the recognition target vocabulary stored in the recognition target vocabulary dictionary storage means, and the acoustic model storage means. Matching is performed using the acoustic model, the likelihood in each partial hypothesis indicating each state of each recognition target vocabulary is calculated, and stored in the recognition target vocabulary dictionary storage means. Collation means for obtaining the syntactic position of the partial hypothesis or the position on the recognition target vocabulary from the syntactic information defining the connection relation of the recognition target vocabulary, the likelihood of each partial hypothesis from the collation means, and the above By inputting the syntactic position of the partial hypothesis or the position on the recognition target vocabulary, the waiting time until the next speech section corresponding to the syntactic position of the partial hypothesis or the position on the recognition target vocabulary When the next voice section is detected within the set waiting time after receiving the voice section detection confirmation notification from the voice section detection means, the next voice section is continued to the matching means. A voice recognition apparatus including a next voice section wait determination unit that outputs a partial hypothesis having the maximum likelihood as a recognition result when a next voice section is not detected within a set waiting time after instructing to perform matching. . 10. A voice section detecting unit that detects a voice section of an input voice signal and outputs a voice section confirmation notification indicating that the voice section has been fixed, and a voice section detected by the voice section detecting unit. An acoustic analysis means for performing acoustic analysis on a speech signal, a recognition target vocabulary dictionary storage means for storing recognition target vocabulary to be recognized and syntax information defining connection relationships between each recognition target vocabulary, and a minimum unit of recognition Acoustic models and acoustic model storage means for storing waiting time information of each acoustic model, acoustic analysis results by the acoustic analysis means, recognition target vocabulary stored in the recognition target vocabulary dictionary storage means, and the acoustic A matching unit that performs matching using the acoustic model stored in the model storage unit and calculates a likelihood in each partial hypothesis indicating each state of each recognition target vocabulary; Likelihood of each partial hypothesis from the matching means, waiting time information of each acoustic model stored in the acoustic model storage means, and connection of each recognition target vocabulary stored in the recognition target vocabulary dictionary storage means By inputting the syntactic information defining the relationship, the waiting time until the next speech section is set by the waiting time information of the last acoustic model of the partial hypothesis having the maximum likelihood, and the speech section from the speech section detecting means is set. When the next voice section is detected within the set waiting time after receiving the notification of detection confirmation, the collating means is instructed to continue the collation of the next voice section and the waiting time is set within the set waiting time. When the next speech segment is not detected, it is determined whether the partial hypothesis with the maximum likelihood can be adopted as the recognition result by the syntax information, and the partial hypothesis with the maximum likelihood is output as the recognition result. Speech recognition apparatus and a voice interval waiting determining means.