JPH04271397A - Voice recognizer - Google Patents

Abstract

PURPOSE:To recognize voice by the voice recognizer in consideration of the combination, of connections between successive obtained phonemes by learning and acquiring only necessary standard patterns corresponding to the combinations of connections between the successive phonemes while performing the recognition even if the standard patterns corresponding to all the combinations of connections between the successive phonemes by preparing the standard patterns corresponding to the phonemes of certain extent. CONSTITUTION:After a recognition result is determined at the time of the recognition of a voice, a phoneme pattern corresponding to an input voice pattern is taken out and environment information (combination information on successive phonemes, i.e., three combination phoneme labels in concrete) showing which position in a word the phoneme pattern is present with a phoneme label obtained from a word dictionary 16 is stored in a phoneme environment table 17; and the phoneme standard pattern is obtained and stored in a phoneme standard pattern memory 15.

Description

[Detailed description of the invention]

0001

[Field of Industrial Application] The present invention relates to a speech recognition device that recognizes speech using segments such as phonemes or syllables as recognition units, and in particular, speech recognition that efficiently selects standard patterns of recognition unit segments and performs identification processing such as pattern matching. This relates to a recognition device.

0002

2. Description of the Related Art Conventionally, speech recognition devices that have been put to practical use are word speech recognition devices that handle speech patterns in units of words and perform recognition on a per-word basis. In this method, patterns of word sounds to be recognized are stored in advance as standard patterns for each word, and recognition processing such as pattern matching is performed for each word.

[0003] Although such a word speech recognition device can obtain a recognition rate that is practical, the vocabulary it can recognize is limited to words that have been registered as standard patterns, so of course it cannot recognize words that have not been registered. . Therefore, in order to enable recognition of as many words as possible, it is necessary to register a large number of words. Therefore, considering that the user has to generate hundreds or thousands of voices during this registration process, such a voice recognition device cannot be said to be practical at all.

[0004] On the other hand, continuous speech recognition devices, which are currently being researched, select speech segments (about several dozen types) that are smaller than words, such as phonemes or syllables, as recognition units, and By recognizing speech segment by segment, it is possible to recognize speech of any vocabulary.

[0005] In such a continuous speech recognition device, it is possible to employ a level building method using automaton control. The level building method using automaton control has already been adopted in technology for continuous word recognition using words as recognition units; for example, in Japanese Patent Publication No. 1-59600.
I am familiar with the issue.

The configuration and operation of a speech recognition apparatus using a level building method using automaton control will be explained below with reference to FIGS. 2 to 5.

The speech recognition devices shown in FIGS. 2 to 5 have a standard pattern for each phoneme and recognize continuous speech. It can be seen as replacing phoneme recognition with a word dictionary in addition to its automaton memory.

First, in FIG. 2, the input voice input from the microphone 21 is extracted as a parameter by a voice analysis section 22 using an analysis method such as LPC cepstral analysis.
An input audio pattern consisting of a time series of parameter vectors is created and stored in the pattern buffer 23. The input speech pattern thus obtained is introduced into the phoneme recognition section 24, and recognition processing is performed using the phoneme standard pattern memory 25 and word dictionary 26.

FIG. 3 shows the internal configuration of the phoneme recognition section 24. In the phoneme recognition unit 24 in the figure, the word dictionary control unit 244
As shown in FIG. 5, a phoneme label string describing a word is read out from the word dictionary 26 having a tree-structured chain of phonemes.

That is, first, the phoneme label at the beginning of the phoneme label string is taken out, and the phoneme standard pattern to which the same phoneme label is attached is read out from the phoneme standard pattern memory 25 in FIG. 4 by the phoneme standard pattern control section 243. ,
DP matching is performed from the beginning of the input audio pattern in the pattern buffer 23 by the level building method. Following this, the phoneme standard pattern control unit 24 extracts phoneme labels in order according to the order of the phoneme label string, and selects the phoneme standard patterns to which the same phoneme labels are attached.
3, the phoneme standard pattern memory 25 is read out and level building is continued.

[0011] If the end of the phoneme label string is reached and the level building reaches the end of the input speech pattern in the pattern buffer 23, the matching distance value and the phoneme label string representing the word are used as output candidates for word determination. The information is sent to section 242.

[0012] The word determining unit 242, which has received the output candidates in this manner, selects the minimum matching distance value from all matching distance values, and outputs the phoneme label string of the corresponding word as a recognition result.

In this example, the phoneme standard pattern memory 25
is the phoneme label /a/, /k/, ~, /N as shown in Figure 4.
The word dictionary 26 is a memory in a table format in which / and corresponding phoneme standard patterns are associated with each other, and the word dictionary 26 is a memory in which phonemes are stored in a tree structure as shown in FIG.

[0014] Such a continuous speech recognition device has the following drawbacks, which are current problems to be solved.

That is, in the conventional device described above, phoneme standard patterns are retrieved from the phoneme standard pattern memory in accordance with the order of the phoneme label string in the word dictionary, and matching is performed. Even labels are affected by the phonemes before and after the utterance in a word, and the phoneme pattern is deformed, so even if the phonemes have the same phoneme label, it is difficult to match them using the same standard phoneme pattern. is inappropriate.

However, if we try to take into account the environment of connected phonemes before and after a phoneme, there are thousands of combinations of phoneme labels including the phonemes before and after the phoneme, and it is difficult to create standard phoneme patterns at once from audio data that satisfies all of these combinations. The disadvantage was that a large amount of audio data had to be processed first.

[0017]

[Problems to be Solved by the Invention] The present invention has been made in view of the above-mentioned disadvantages, and it is intended that if a standard pattern corresponding to a certain number of segments is prepared as an initial state, all of the connections before and after the segments can be completed. Even if there is a lack of standard patterns that correspond to the combination, while performing recognition processing,
The purpose of this invention is to realize a speech recognition device that can learn and acquire as many standard patterns as necessary corresponding to combinations of front and rear connections of segments.

[0018]

[Means for Solving the Problems] The speech recognition device of the present invention uses segments such as phonemes or syllables constituting speech as recognition units, and uses standard pattern data that numerically expresses standard patterns of recognition unit segments. A front and rear environment information table is stored in which the front and rear environment information regarding the recognition unit segments before and after the recognition unit segment of the standard pattern are associated with each other, and a large number of words are stored, each consisting of a segment chain with the recognition unit segment as a heading. At the time of recognition, a standard pattern of the recognition unit is selected and recognized based on the surrounding environment information of the recognition unit in the above-mentioned surrounding environment information table. By clustering the input speech, a standard pattern and a front/back environment information table are created from the input speech, and during additional learning,
By performing voice clustering using a voice pattern obtained by weighting the standard pattern with a weighting coefficient based on the front and rear environment information stored in the front and rear environment information table and frequency information based on the number of learning samples, a new This is to create a standard pattern and a front/back environment information table.

[0019]

According to the speech recognition apparatus of the present invention, when a user specifies a learning function during speech recognition, additional learning of phoneme standard patterns and tables can be performed using speech input during the learning period. This makes it possible to learn unlearned phoneme standard patterns and gradually improve recognition performance. In addition, such additional learning
It is also possible to perform this in parallel with normal recognition operations, and additional learning can be performed while performing recognition to improve recognition performance.

[0020]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows the configuration of an embodiment of a speech recognition apparatus according to the present invention.

In the figure, 11 to 16 correspond to 21 to 2 in FIG.
Similarly to 6, "microphone" to "word dictionary" are shown, and the device of this embodiment differs from the device of FIG. 1 in that a phoneme environment table 17 is added.

The configuration and operation of the embodiment of the present invention shown in the figure will be explained below. First, the voice input from the microphone 11 is extracted as parameters by an analysis method such as LPC cepstrum analysis in the voice analysis unit 12, and an input voice pattern consisting of a time series of parameter vectors is created and stored in the pattern buffer 13. .

In the phoneme recognition unit 14, as shown in FIG.
Based on the level information of 1, the phoneme label at the corresponding position and the triplet phoneme label of the preceding phoneme label and the following phoneme label are read out.

As the word dictionary 16, for example, a memory in which a tree-structured phoneme label string as shown in FIG. Phoneme label /k/(*
2), and the triple phoneme label of the preceding phoneme label /a/ (*1) and the following phoneme label /i/ (*3) are obtained simultaneously.

Next, this triplet phoneme label /a/, /k
/, /i/ (phoneme environment), phoneme environment table 17
A phoneme standard pattern corresponding to that phoneme environment is read out from the phoneme environment.

Such a phoneme environment table 17, for example, as shown in FIG.
A memory is used in which a phoneme standard pattern label] and its [phoneme environment] are stored in association with each other. For example, /a
For /, /k/, /i/ (denoted as /aki/ in the figure), search the triplet phoneme column and find the cluster number. 1 /k/ is found as the corresponding phoneme standard pattern label. The total number of such triplet phonemes is on the order of several thousand, whereas the total number of phoneme standard pattern labels is about 200 at most.

This is because matching processing time increases as the number of phoneme standard patterns increases, so the number of phoneme standard patterns is limited and a plurality of phoneme environments (triad phonemes) are assigned to one phoneme standard pattern. A table that describes this assignment is called a phoneme environment table. This assignment (correspondence) is statistically determined based on a large number of learning phoneme patterns.

The phoneme standard pattern control unit 143 controls the phoneme standard pattern label /k/ indicated by the phoneme environment table 17.
(Cluster No. 4) makes phoneme standard pattern memory 1
Cluster No.5 in 5. The phoneme standard pattern for the four-phoneme label /k/ is read out and sent to the level building section 141. The level building section 141 is
DP matching is performed by the level building method from the beginning of the input speech pattern in the pattern buffer 13, and as shown in FIG. 10, the phoneme sent from the phoneme standard pattern control unit 143 is Matching with standard pattern.

Further, if the matching reaches the end of the phoneme label string in the word dictionary and the level building reaches the end of the input speech pattern in the pattern buffer 13, the level building unit 141 outputs the matching distance value as an output candidate. The phoneme label string representing the word is processed by the word determination unit 1.
Send to 42.

[0030] The word determination unit 142 selects the minimum matching distance value from all matching distance values, and outputs the phoneme label string (recognition result) of the corresponding word.

[0031] In this manner, when selecting a standard pattern for a certain phoneme, a phoneme environment table is used to effectively select a pattern with a matching phoneme environment.

The initial learning of the phoneme standard pattern and phoneme environment table will now be explained.

First, data for each phoneme section (referred to as phoneme segment data) is collected from the initial learning speech data in which phoneme labels are attached for each phoneme section. Clustering is performed based on these phoneme segment data. Clustering is performed as follows.

(1) As an initial cluster, phoneme segments having the same phoneme label are set as one cluster, and the centroid of each cluster is determined. If the phoneme label is as shown in FIG. 8, the initial number of clusters N is 40.

(2) Find the average distortion for each cluster, and set the largest distortion as DNmax. If the number of clusters N is a predetermined number, the process ends.

(3) Divide the cluster with the maximum distortion into two by the phoneme segment that is the longest distance within the cluster, and find the centroid of each. Set N=N+1 and repeat from (2). Note that the average strain Dn of cluster n is

[0037]

[Math 1]

[0038] Here, Mn represents the number of elements (phoneme segments) included in cluster n, Cn represents the centroid of cluster n, and Xni represents the i-th element (phoneme segment) of cluster n. Further, the final number N of clusters is approximately 128 to 256.

The phoneme segment data that is the centroid of each cluster obtained in this manner is used as a phoneme standard pattern, and the phoneme environment and number of phoneme segment data included in each cluster is used as a phoneme environment table. An example of such a table is shown in FIG. 11, where cluster no. , phoneme labels, and phoneme environment frequency information. The number following each phoneme environment represents frequency information of the phoneme environment, and represents the number of elements (phoneme segments) included in each cluster during learning.

However, as the frequency information, for example, a value obtained by normalizing the cumulative number of appearances of the phoneme environment to a maximum value of 100 may be used.

At the time of recognition, this number of appearances is used to select a phoneme standard pattern. For example, the triple phoneme label /_ki/ appears in /k/ in clusters 5 and 6, respectively, and the number of occurrences is 27 and 33. If this is normalized to a maximum value of 100, it becomes 81,100. Here, if the selection threshold is 85, /_k
The phoneme standard pattern label having the phoneme environment of i/ becomes cluster 6, and the corresponding phoneme standard pattern is read from the phoneme standard pattern memory. Depending on the threshold value, there may be cases where matching is performed with multiple phoneme standard patterns in the same phoneme environment, but in that case, the one with the smallest distance from the matching result with the input speech pattern is ultimately selected.

[0042] In this way, the standard pattern of a certain phoneme is 1
There are multiple phoneme standard patterns for one phoneme, and each phoneme standard pattern has multiple phoneme environments. However, unlike conventional multi-template phoneme standard patterns, this method has the advantage that which phoneme standard pattern is selected is determined by the phoneme environment, and frequency information is also included in the criteria for this determination. Therefore, unlike the conventional multi-template method, standard patterns (templates) can be effectively selected.

Next, additional learning of phoneme standard patterns and phoneme environment tables will be explained. The phoneme environment table and phoneme environment table obtained through initial learning reflect the phoneme environment included in the initial learning data, but as shown earlier, they do not include all the phoneme environments of speech. do not have. Additional learning is to learn phoneme environments that are not included in this initial learning data, and is performed as follows.

(1) The phoneme segment data in the input speech automatically extracted from the matching result during recognition is used as additional learning data 1 together with its phoneme environment information.

(2) The phoneme standard pattern weighted by the number of cluster elements in the phoneme environment table is used as additional learning data 2. (This can be considered as approximation data of the initial training data with average distortion for each cluster.) (3
) From the additional learning data 1 and 2, the phoneme standard pattern and phoneme environment table are re-created using the same method as the initial learning described above.

[0046] The phoneme standard pattern created in this way,
and the phoneme environment table includes the phoneme environment included in both the initial learning data and the additional learning data,
The phoneme standard pattern is obtained as a new centroid,
The phoneme environment table has phoneme environments included in both and their frequency information.

Further, additional learning can be performed any number of times to acquire new phoneme environments. At that time, the amount of data is N
Since it is represented by (=128 to 256) phoneme standard patterns, no matter how many times additional learning is performed, it can be processed using N+the number of additional learning data. Moreover, the number of data to be stored is only N phoneme standard patterns, and the amount of processing for additional learning from the second time onwards does not increase.

Furthermore, in order to acquire data for additional learning and perform additional learning, an additional learning period must be specified at the time of recognition.

In the speech recognition device of the present invention, as a method of specifying an additional learning period, when a learning function is specified in the recognition state, a predetermined period thereafter (until the end instruction is given by the user, or The recognition result of the voice input during a period of time or a certain number of inputs, etc.
Alternatively, based on the correction result, phoneme segment data is acquired from the input speech as additional learning data and additional learning is performed. Additional learning can be performed in parallel with normal recognition operations, in which case the speech recognition device will gradually adapt to the user's own voice without the user being aware of the learning. .

[0050] In the example, an example was shown in which the level building method was used with phoneme segments as recognition units, but in this part, segmentation for each phoneme is first performed and matching with the phoneme standard pattern is performed for each phoneme interval. Similar effects can be obtained depending on the method used. At this time, the configuration of the phoneme recognition section is as shown in FIG. 12, and the level building section 141 is replaced with a phoneme segmentation section 145 and a phoneme matching section 146.

[0051] Furthermore, in the above explanation, a case has been described in which phoneme segments are used as recognition units, but the present invention is not limited to this. For example, similar effects can be obtained even if syllable segments are used as recognition units. It will be done.

As described above, according to the present invention, when selecting a standard pattern for a certain phoneme, a phoneme environment table is used to effectively select a pattern with a matching phoneme environment. Unlike the method in which one phoneme standard pattern has one phoneme standard pattern, or the simple multi-template method in which one phoneme label has multiple phoneme standard patterns, but each is treated the same, each phoneme standard pattern A speech recognition device that can effectively select a phoneme standard pattern (template) according to the phoneme environment in the dictionary can be realized.

[0053]

Effects of the Invention The speech recognition device of the present invention extracts the corresponding phoneme pattern of the input speech pattern after the recognition result is determined during speech recognition, and identifies the phoneme pattern according to the phoneme label obtained from the word dictionary. It is possible to acquire environmental information such as the position in a word and the standard phoneme pattern. Therefore, if standard patterns corresponding to a certain number of phonemes are prepared in the initial state of the device,
Even if there are not enough standard patterns that correspond to all combinations of phoneme connections, it is possible to learn and acquire as many standard patterns that correspond to combinations of phoneme connections as necessary while performing recognition processing. It is possible to realize speech recognition with high recognition accuracy that takes into account the combination of front and rear connections of phonemes.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram showing an embodiment of a speech recognition device according to the present invention.

FIG. 2 is a configuration diagram showing an example of a conventional speech recognition device.

FIG. 3 is a block diagram showing a phoneme recognition unit of the conventional speech recognition device shown in FIG. 2;

FIG. 4 is a configuration diagram showing a phoneme standard pattern memory of the conventional speech recognition device of FIG. 2;

FIG. 5 is a configuration diagram showing a word dictionary of the conventional speech recognition device shown in FIG. 2;

FIG. 6 is a configuration diagram showing a phoneme recognition unit of the speech recognition device according to the present invention.

FIG. 7 is a configuration diagram showing a word dictionary of the speech recognition device according to the present invention.

FIG. 8 is a schematic diagram showing an example of phoneme labels of the speech recognition device according to the present invention.

FIG. 9 is a configuration diagram showing a phoneme standard pattern memory of the speech recognition device according to the present invention.

FIG. 10 is a diagram showing matching between an input pattern of the speech recognition device according to the present invention and a phoneme standard pattern.

FIG. 11 is a configuration diagram showing a phoneme environment table of the speech recognition device according to the present invention.

FIG. 12 is a configuration diagram showing another embodiment of the phoneme recognition unit of the speech recognition device according to the present invention.

[Explanation of symbols]

11. Microphone 12 Speech analysis section 13 Pattern buffer 14 Speech recognition section 15 Phoneme standard pattern memory 16 Word dictionary 17 Phoneme environment table

Claims (1)

[Claims] Claim 1: In a speech recognition device that recognizes speech using segments such as phonemes or syllables constituting speech as recognition units, standard pattern data numerically expressing a standard pattern of recognition unit segments and recognition unit segments of the standard pattern. The recognition system is equipped with a front and rear environment information table that stores the front and rear environment information related to the recognition unit segments before and after the recognition unit segment connected to the target in association with each other, and a word dictionary that stores a large number of words consisting of segment chains with the recognition unit segment as the header. Sometimes, a recognition process is performed in which a standard pattern of the recognition unit is selected and recognized based on the surrounding environment information of the recognition unit in the above-mentioned surrounding environment information table, and during initial learning, the speech input in advance is clustered. , create a standard pattern and a front/back environment information table from the input audio, and during additional learning, the standard pattern is calculated using a weighting coefficient based on the front/back environment information stored in the front/back environment information table and frequency information based on the number of learning samples. Using weighted voice patterns,
A speech recognition device characterized by creating a new standard pattern and a front/back environment information table by performing clustering of speech.