JP2004012713A - Device and method for recognizing speech - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recognition device which can realize high recognition rate. <P>SOLUTION: This speech recognition device is provided with a storage means for storing a database correlating inquiry vocabulary candidates and response vocabulary candidates with probabilities with which both vocabulary candidates are connected, an identification processing means for performing identification processing of a speech word acquired and the vocabulary candidate by using the storage means, and a correcting means for correcting probability in the database stored in the storage means based on the result of the identification processing by the identification processing means. Consequently, high recognition rate can be realized by utilizing relation between the vocabulary candidate of the inquiry, the vocabulary candidate of the response and probability with which both vocabulary candidates are connected. <P>COPYRIGHT: (C)2004,JPO

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a voice recognition device and a voice recognition method.
[0002]
[Prior art]
2. Description of the Related Art A speech recognition device that acquires speech as data, processes the data, and recognizes the uttered content has been put to practical use. For example, a speech recognition engine is built in an application for inputting a sentence by utterance or an in-vehicle navigation system operated by utterance. In the current speech recognition, even if the acquired speech is analyzed as it is (analysis by acoustic model: phoneme analysis), the recognition rate remains at 60% to 80%. In addition to this, the recognition rate is increased to 85% to 95% by performing analysis [analysis using a language model] of grammar theory and the connection of words (phonemes) to complement them.
[0003]
As an analysis method using the acoustic model described above, a method using a probabilistic model such as a hidden Markov model (HMM: Hidden Markov Model) is famous. As an analysis method using the above-mentioned language model, there are methods such as bigram and trigram. These are methods for estimating one vocabulary in the same utterance by the same person, and predicting the entire vocabulary from the first two or three sounds (phonemes) in one vocabulary.
[0004]
[Problems to be solved by the invention]
Based on the situation of speech recognition, the inventor has conducted intensive research aiming at further improving the recognition rate, and has invented a new effective method as an analysis method using a language model. That is, an object of the present invention is to provide a speech recognition apparatus and method capable of realizing a high recognition rate.
[0005]
[Means for Solving the Problems]
The speech recognition apparatus according to claim 1, wherein the storage unit stores a database that associates a vocabulary candidate to be asked, a vocabulary candidate for a response, and a probability that both vocabulary candidates are associated with each other, and a word and a vocabulary of a speech acquired using the storage unit. It is characterized by comprising identification processing means for performing identification processing with a candidate, and correction means for correcting the probability in the database stored in the storage means based on the result of the identification processing by the identification means.
[0006]
According to a second aspect of the present invention, in the speech recognition apparatus according to the first aspect, the database includes a short-term conversation data set and a long-term conversation data set, and the correction unit outputs the result of the identification processing. On the basis of this, correction is performed on the short-term conversation data set and the long-term conversation data set, respectively.
[0007]
The speech recognition method according to claim 3 performs an identification process on the acquired speech word using a database in which a vocabulary candidate to be queried, a vocabulary candidate for a response, and a probability that both vocabulary candidates are associated are identified. It is characterized in that the probability in the database is corrected based on the processing result.
[0008]
According to a fourth aspect of the present invention, in the voice recognition method according to the third aspect, the database has a short-term conversation data set and a long-term conversation data set, and the short-term conversation data set is used based on a result of the identification processing. The present invention is characterized in that the conversation data set and the long-term conversation data set are respectively corrected.
[0009]
Note that the probability referred to here is not limited to the probability in the academic sense and the probability in a strict sense in terms of statistics (probability such as 1/10 or 10% when there is a possibility that it occurs once in 10 times). It also has the meaning of a broad sense (including the narrow sense described above) that relatively indicates the possibility of an event occurring.
[0010]
BEST MODE FOR CARRYING OUT THE INVENTION
As described above, the speech recognition uses both the analysis by the acoustic model and the analysis by the language model to improve the recognition rate. In the analysis using the conventional language model, the above-mentioned bigram, trigram, and the like are used. However, these methods do not reflect the interaction between people. In other words, the analysis based on the conventional language model is a micro-prediction that predicts the entire vocabulary (word) by halfway utterance. When trying to improve the recognition rate by focusing on the grammar of the language (for example, Japanese) and the rules of thumb of speech, predictions that focus on not only such microscopic predictions but also dialogues of responses to questions, The inventors have found that a concept that can be called macro prediction is effective for micro prediction. The present invention is based on such findings.
[0011]
The speech recognition device of the present invention has an input unit, an arithmetic processing unit (identification processing unit / correction unit), and a storage unit (storage unit). The arithmetic processing unit and the storage unit are configured as an electronic control unit (ECU) including a CPU, a ROM, a RAM, and the like. Note that the storage unit may be an exchangeable hard disk, an optical disk such as a CD-ROM, a DVD-ROM, or the like, or may be an externally attached one. The database and the like described later are stored in these storage units. A microphone as an input unit is connected to the ECU described above. The CPU in the ECU performs various calculations, and various data at the time of the calculations are held in a cache memory or a RAM in the CPU.
[0012]
Further, the voice recognition device of the present embodiment is incorporated in a car navigation system, and attempts to recognize an operator's response to a question (question) issued by the car navigation system. For this reason, it is not necessary to perform a recognition process on the question sentence (word) issued by the car navigation system, and the apparatus itself grasps the question sentence (word) from the beginning.
[0013]
FIG. 1 schematically shows the structure of a database used for speech recognition according to the present embodiment. FIG. 2 is a flowchart illustrating a speech recognition process in the speech recognition apparatus and method according to the present embodiment.
[0014]
First, the structure of a database used in the speech recognition of the present embodiment will be described with reference to FIG. As shown in FIG. 1, there are two sets of matrix-like data sets in which vocabulary candidates to be queried are arranged in the row direction, and vocabulary candidates for responses are arranged in the column direction. Then, the probability that the vocabulary candidate related to each response appears for the vocabulary candidate for each question is stored at the intersection.
[0015]
As described above, there are two sets of data sets. A short-term conversation data set and a long-term conversation data set. The short-term conversation data set is one in which the above-described probabilities are stored as probabilities that appear in a series of relatively short conversations. On the other hand, a long-term conversation data set is defined as a probability that the above-mentioned probability appears during a conversation when the conversation is performed (here, this is defined as a long-term conversation corresponding to a short-term conversation). Is what is being done. The two data sets differ only in the probability of being stored at the intersection (it may be temporarily the same), and the arrangement of the matrix-like vocabulary candidates is the same.
[0016]
For example, in a series of dialogues conducted in a short period of time, it is rare (the probability is low) that exactly the same questions (and responses to them) are made. On the other hand, in the long run, the same question is more likely to be asked again (the probability is high) in consideration of the human nature of the speaker of the conversation, the object of interest (hobbies and interests), and the like. That is, the recognition rate can be improved by taking into account the difference between the probabilities of a short-term view and a long-term view. Therefore, here, two data sets are prepared, and the recognition rate is improved by effectively using these data sets. The specific usage will be described later.
[0017]
Next, a description will be given based on the flowchart shown in FIG. First, a voice is acquired with the above-described microphone or the like (step 200). The acquired voice is first analyzed by an acoustic model (step 205). Examples of the analysis by the acoustic model include a hidden Markov model and an analysis by an island (island) analysis proposed by the present inventors. Next, it is determined whether or not the recognition rate in the analysis using the acoustic model is sufficiently high (step 210). Here, it is assumed that if the recognition rate is larger than α, it is sufficiently high.
[0018]
If the recognition rate is sufficiently high, that is, if step 210 is affirmative, it is determined that recognition was possible only by analysis using the acoustic model, and the recognition is determined (step 215), and the process exits the flowchart in FIG. On the other hand, when step 210 is denied, it is subsequently determined whether or not the recognition rate in the analysis by the acoustic model is too low (step 220). Here, it is assumed that the case where the recognition rate is smaller than β is too low. In this case, it is determined that recognition of the acquired voice is difficult (even if analysis is performed using a language model in addition to analysis using an acoustic model), and after issuing a re-speak instruction (step 225), the process exits the flowchart in FIG. .
[0019]
The re-speaking instruction may be an error sound (beep sound) or an error display indicating that the recognition has failed, or may output "Please re-enter" as a synthesized voice or recorded voice from the speaker of the device. May be. The reuttered voice is acquired, and the voice recognition processing is performed again from step 200. On the other hand, if step 220 is denied, it can be said that the recognition rate is higher than β described above but does not reach α. In this case, analysis is performed using a language model that takes into account the dialogue (question and response) described above (step 230).
[0020]
As a result, it is determined whether or not the recognition rate is sufficiently high (step 235). If the recognition rate is sufficiently high, that is, if step 235 is affirmed, the recognition is determined (step 240). On the other hand, if step 235 is denied, it is determined that recognition of the acquired voice is difficult (analysis using the language model is difficult), and after issuing a re-speak instruction (step 225), the process exits the flowchart in FIG. .
[0021]
Here, the description of each step partially overlaps with that described above, but the analysis using the language model after step 230 described above will be described in detail below.
[0022]
When the device is used for the first time, each probability in the two data sets shown in FIG. 1 has been reset to an initial value of 100. The two data settings are corrected each time speech recognition is performed from this initial state, the corrected data set is stored (only the data set for long-term conversation is described later), and the subsequent voice recognition is continued. used. In the case of the present embodiment, first, a sentence (or word, the same applies hereinafter) to be queried is emitted from the navigation system by a synthesized voice or a recorded voice [a in FIG. 1]. This sentence is stored as a candidate vocabulary in the vocabulary candidate in question in FIG.
[0023]
On the other hand, a response sentence is uttered by the speaker. The device acquires the uttered vocabulary using a microphone or the like (step 200), and first performs analysis using an acoustic model (step 205). As a result, vocabulary candidates for the response are detected. The vocabulary candidate at this time may not completely match the vocabulary candidate of the response in FIG. 1, but in this case, the most similar one is selected [b in FIG. 1]. In this way, (a, b) in the matrix is determined (step 230). If the probability of (a, b) is within the upper 10% of the probabilities in all elements in row a of the short-term conversation table (step 235), it is recognized that the response to question a is b. (Step 240).
[0024]
After (a, b), the pair of data sets is corrected. First, regarding the data set for short-term conversation, the data stored in (a, b) is set to -10, and all data other than (a, b) in row a is set to +3. As described above, during the short-term conversation, it is considered that there is a low possibility that the exchange once spoken is performed again. Therefore, the probability of the data (a, b) is corrected to be low. In addition, the other data in the row a appear in the conversation because the vocabulary not yet used in the short-term conversation is more likely to appear in the conversation than the vocabulary already used. Modify to increase the probability of possibility. The short-term conversation may include setting a destination between the navigation system and the operator, searching for a destination, obtaining information around the destination, and the like.
[0025]
On the other hand, regarding the long-term conversation data set, +5 is set for data stored in (a, b), and -2 is set for all data except (a, b) in row a. As described above, similar topics are often spoken when considered over a long span, and it is highly likely that the exchange once recognized will be performed again. Modify the probability to be higher. As for the other data in the row a, the vocabulary of a conversation that is rarely performed in some assumed conversations is unlikely to appear in future conversations. Modify to lower the probability.
[0026]
Thus, the two data sets are modified (step 245). As described above, in step 235, when determining whether or not recognition was possible from the probability of (a, b), the determination based on the short-term conversation data set instead of the long-term conversation data is performed in two steps. This is because each data set has the above-described properties. Then, it is determined whether or not the topic has changed (step 250). If the topic has not changed, the flowchart of FIG. 2 ends. If the topic has changed, the short-term conversation that has been used up to that point is used. The long-term data set is discarded once (step 255), and the long-term data set at that point is copied as it is (step 260).
[0027]
At that moment, the two data sets are the same, but by performing the recognition process again, different weights are given to the respective data sets, thereby contributing to an improvement in the recognition rate. There are various ways to determine whether the topic has changed. For example, it may be determined that the topic has changed when the interval between the utterances is long, or it may be determined that the topic has changed when a word that frequently appears during the conversation has changed. Alternatively, it may be determined that the topic has been changed when a word that frequently appears at the time of changing the topic (for example, “Okay,” etc.) is detected. In the case of the present embodiment, it is premised on the exchange of the question of the navigation system and the response of the operator, so that it is easy for the navigation system to grasp the change of the topic.
[0028]
Note that the present invention is not limited to the embodiment described above. For example, the device of the above-described embodiment is integrated into a navigation system, and the vocabulary to be asked is known from the beginning by the device. However, the present invention can also be applied to a case where both the inquiring side and the responding side are to be grasped by voice recognition. However, as described above, the form of recognizing the response to the question issued by the device can be said to be a form to which the present invention can be more effectively applied since the vocabulary of the questioner can be specified.
[0029]
【The invention's effect】
According to the speech recognition apparatus and method according to the first and third aspects, the speech recognition is performed by using the relationship between the vocabulary to be queried, the vocabulary candidate for the response, and the probability that both vocabulary candidates are connected, thereby achieving high recognition. Rate can be obtained. In particular, if the target is specialized in an interactive format, it is possible to realize a recognition rate higher than that of the related art.
[0030]
According to the speech recognition apparatus and method of the present invention, a data set for long-term conversation and a data set for short-term conversation are prepared in a database, and speech recognition is performed using the two data sets. Therefore, it is possible to realize a higher recognition rate by utilizing the difference between the utterance rate of a specific vocabulary in a short span and the utterance rate in a long span.
[Brief description of the drawings]
FIG. 1 is an explanatory diagram schematically showing a database structure in an embodiment of a speech recognition apparatus and method of the present invention.
FIG. 2 is an explanatory diagram schematically showing a database structure in one embodiment of the speech recognition device of the present invention.

Claims (4)

Storage means for storing a database in which a vocabulary candidate to be asked, a vocabulary candidate for a response, and a probability of association between the vocabulary candidates are stored;
Identification processing means for performing identification processing between the words of the voice acquired using the storage means and the vocabulary candidates, and
A speech recognition apparatus, comprising: a correction unit that corrects the probability in the database stored in the storage unit based on a result of the identification processing by the identification unit. The database has a short-term conversation data set and a long-term conversation data set,
2. The speech recognition apparatus according to claim 1, wherein the correction unit corrects each of the short-term conversation data set and the long-term conversation data set based on a result of the identification processing. 3. The acquired speech word is identified using a database that associates the vocabulary candidate to be asked, the vocabulary candidate for the response, and the probability that the two vocabulary candidates are associated with each other, and based on the result of the identification process, A speech recognition method characterized by correcting a probability. The database has a short-term conversation data set and a long-term conversation data set,
The speech recognition method according to claim 3, wherein the short-term conversation data set and the long-term conversation data set are respectively corrected based on the result of the identification processing.

Images