JP3003355B2 - Standard pattern making device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a standard pattern creating apparatus for creating a standard pattern using learning data in speech recognition.

[0002]

2. Description of the Related Art At present, in the field of speech recognition, an unspecified speaker recognition system aiming to be able to recognize anyone's voice has been actively researched and developed. In these recognition systems, recognition methods such as a hidden Markov model and a neural network are widely used. The details of the hidden Markov model are described in detail, for example, in "Speech Recognition by Probabilistic Model", written by Seiichi Nakagawa, IPSJ (1988). Also, regarding speech recognition by a neural network, for example,
"Speech / Hearing and Neural Network Model" Shunichi Amari, eds. 199
In 0, he was familiar with Ohmsha (hereinafter referred to as Reference 2). In these methods, learning is performed using learning data composed of words, sentences, and the like uttered in advance by a large number of speakers, and high-performance recognition is realized by creating a more accurate standard pattern. .

At present, in a speech recognition system, the vocabulary to be recognized is limited depending on the use of the recognition device due to hardware limitations of the recognition device or to enhance recognition performance. The upper limit is about 1000 words. For example, in the use of ticket reservation, an artist name, a concert venue name, a vocabulary used for utterance of a reservation, and the like are target vocabularies, and the entire utterance composed of them is a recognition target. Here, as an example of the “task” in the claims, this limited recognition target will be described and described below. In this case, different tasks mean that the recognition target vocabulary and the recognition target formed therefrom are different.

When a standard pattern is created in word units or sentence units, when recognizing a different task, it does not exist in the original task but does not exist in the task. It is necessary to learn a new vocabulary to be recognized. However, learning a standard pattern in an unspecified speaker recognition system requires the invention of many speakers. Therefore, learning a new vocabulary every time a task is changed requires a great deal of work in constructing a database. It requires cost and time and is not realistic.

Accordingly, in a general speaker-independent system,
Sub-words (phonemes, syllables, etc.) smaller than words are used as recognition units, and models of these recognition units are learned during learning. New vocabulary models for different tasks can easily be created by concatenating these subword models.

In an unspecified speaker recognition apparatus, there are many examples in which a reference task is selected so that all recognition units appear as evenly as possible, and a model of the recognition unit is learned.

[0007]

In the recognition system using the above-described subword as a recognition unit, when a task different from the original task used for learning is to be recognized, compared with the case where the original task is recognized, It has recently been reported that cognitive performance is degraded.

In a new task, the actual utterance corresponding to a subword is deformed due to a difference in its surrounding environment, such as what the preceding and following subwords are, and becomes different from that at the time of learning. Is considered to be the cause.

As a target of the above-mentioned problem, it is conceivable that learning is performed again using the utterance of a new task. However, preparing a database in which many speakers utter the task every time the task changes,
Requires great effort.

According to the present invention, in an unspecified speaker recognition system, even when a task is changed, in addition to the utterance data learning data of the tasks of many speakers, the original task and the new task of a small number of speakers are provided. The object of the present invention is to create a high-accuracy standard pattern for a new task without using many speakers' new task utterances, and to improve the performance of the recognizer for the new task.

[0011]

According to a first aspect of the present invention, there is provided a standard pattern creating apparatus for learning a standard pattern using learning data uttered by a plurality of speakers in speech recognition. A standard pattern for a reference speaker is learned using reference pattern creation means for creating a pattern, and utterance data of a first task of a reference speaker who has uttered both the first task and a second task different from the first task. A first reference pattern creating unit, a second reference pattern creating unit that learns a standard pattern for the reference speaker using the utterance data of the second task of the reference speaker, and the reference from the first reference pattern. A task creating means for creating a mapping to the second pattern, and a task dependent pattern creating a task dependent pattern from the reference pattern using the mapping created by the mapping creating means. The turn creating means, in that they are composed of and wherein.

According to a second aspect of the present invention, there is provided the standard pattern creating apparatus according to the first aspect, wherein the reference pattern creating means, the reference first pattern creating means,
A task quasi-dependent pattern is created by mixing the reference second pattern creating means, the mapping creating means, the task dependent pattern creating means, and the reference pattern and the task dependent pattern at a certain ratio. And task quasi-dependent pattern creation means.

[0013]

The operation of the apparatus for producing a standard pattern according to the first invention will be described below. Here, a limited recognition target is given as an example of the “task”, and that case will be described below.

The utterance of the speaker is converted into a time series of feature vectors for each unit called a frame having a certain time length through processes such as AD conversion and voice analysis. The feature vector is obtained by extracting a very small amount of the voice spectrum at that time, and usually has 10 to 100 dimensions. The time series of the feature vectors is referred to as utterance data here.

As a recognition method, a case where a continuous probability density distribution is used as an output probability distribution in a hidden Markov model will be described as an example. A Gaussian mixture distribution is used as the continuous probability density distribution. In this case, the parameters to be learned include the average vector of the mixed continuous distribution, the variance, the branch probability, and the transition probability. Here, the case of learning the average vector of each distribution will be described as an example. That is, the standard pattern in the present invention is a set of the average vector of each distribution of each state in the hidden Markov model corresponding to each recognition unit, covering all recognition units, all states, and all distributions. That is, {μ (i, j, k) | j = 1, ..., I, j = 1, ..., J (i), k = 1, ..., K (i, j)}. (1) where μ is an average vector, I, J (i), K (i,
j) is the total number of recognition units, the number of states in recognition unit i, and the number of distributions in state j of recognition unit i, respectively.

First, a standard pattern for a standard task A is created. In the case of the Hidden Markov Model, this standard pattern is usually created by preparing utterance data of task A of many speakers and learning the parameters of each recognition unit using the utterance data. The learning method is described in detail in Reference 1. The pattern created here is called a reference pattern. Speakers used to create reference patterns include:
The reference speaker described below may or may not be included. The above corresponds to the reference pattern creating means in the first invention.

Next, utterance data of a speaker who has uttered both the task A and the task B different from the task A are prepared. This speaker is called a reference speaker. The number of reference speakers may be one or more. Here, a plurality of reference speakers will be described as an example.

A standard pattern is created by using the utterances of task A of these reference speakers. The creation method may be a normal learning method or a method of learning only the average vector among the parameters. Also, using the reference model, the utterance is segmented for each recognition unit, and after identifying the utterance data corresponding to each distribution of each state of the hidden Markov model, by averaging the feature vectors of the utterance data, A method of creating an average vector for each distribution is also possible. The standard pattern thus formed is referred to as a reference first pattern. Only one reference first pattern may be created for all reference speakers, or one reference first pattern may be created for each reference speaker.
Further, the reference speakers may be divided into several groups, and a first reference pattern may be created for each group. In this case, a set of reference speakers for each first reference pattern is called a reference speaker group. For example, when the first reference pattern is created for each of the reference speakers, the number of reference speaker groups is equal to the number of reference speakers, and one reference first pattern is included in the entire reference speakers.
When making a pattern, there is only one reference speaker group. The above corresponds to the reference first pattern creating means in the first invention.

Next, in the same manner as the above-described first reference pattern creating means, a standard pattern is created by using the utterance of the task B of the reference speaker. The creation method is the same as the method used for the first reference pattern creation means. The reference second pattern is
It is created so as to correspond one-to-one with the reference first pattern created earlier. That is, in the corresponding reference first pattern and reference second pattern, the reference speaker group used for the creation is the same. The above corresponds to the reference second pattern creating means in the first invention.

Next, a mapping from the first reference pattern to the second reference pattern is created. Now, the recognition unit, the state of the model of the recognition unit, the distribution in the state, and the like are collectively referred to as a mapping unit. A mapping is created for each mapping unit.

The method of defining the mapping unit is as described above.
In addition to the three types, for example, a certain recognition unit and another recognition unit may be the same mapping unit, or a certain recognition unit may be a mapping unit as a whole and another recognition unit may indicate a state therein. Or a certain state of a certain recognition unit and a certain state of another recognition unit can be the same mapping unit. It is also possible to use all the standard patterns as one mapping unit. That is, the method of determining the mapping unit is arbitrary. For example, as a mapping unit,
When each distribution in the state of the recognition unit is taken, the average vector of the corresponding distribution of the corresponding state in the model of the same recognition unit may be made to correspond one-to-one.

The mapping unit may be created separately for each reference speaker group, or may be common to all reference speaker groups. However, if a different mapping unit is defined for each reference speaker group, a mapping common to all reference speaker groups cannot be created.

The mapping may be linear, or a non-linear mapping may be created using a neural network or the like.

The above corresponds to the mapping creating means in the first invention.

Next, the reference pattern is converted by using the mapping created by the above-described mapping creating means, and a task-dependent pattern is created. Task dependent pattern is 1 for reference pattern
Corresponds to one. For example, if the mapping unit is distribution,
Apply a mapping to each distribution to determine a task-dependent average vector. If a common mapping is defined for all distributions, an average vector suitable for the task is obtained using a common mapping for the average vector of all distributions of the reference pattern. The set of average vectors adapted to this task is in this case called a task-dependent pattern.

When a mapping is created for each reference speaker group, a plurality of patterns after the mapping can be formed. In such a case, the plurality of patterns are weighted by some method so that task-dependent patterns are obtained. Create a pattern. For example, when the mapping unit is each distribution, the average vector of the distribution of the reference pattern and the reference first
Using the distance between the pattern and the mean vector of the distribution,
A weighting method is conceivable.

[0027]

(Equation 1)

Here, p is a subscript representing a reference speaker group, μ ₁ (p) is an average vector of the distribution in the reference first pattern of the reference speaker group p, and μ ₂ (p) is a reference speaker group p , The average vector of the distribution in the reference second pattern of, μ is the average vector of the distribution in the reference pattern,

[0029]

(Equation 2)

D (p) is the distance between μ and μ (p),
m is an appropriate real number. As the distance, for example, the Euclidean distance between the average vectors is used, but any distance may be used as long as it represents the degree of dissimilarity between the distributions.

As a method for creating a task-dependent pattern, for example,

[0032]

(Equation 3)

The following is also conceivable. The definition of the symbol is the same as in the expression (2).

In the above description, the mapping unit is each distribution,
This is for the case where a mapping is created for each reference speaker group.However, when the mapping unit is other than each distribution, or when a mapping common to all reference speaker groups is created, , It is possible to create task dependent patterns. The above corresponds to the task dependent pattern creating means in the first invention.

The above is the description of the standard pattern forming apparatus of the first invention. With this device, when there is a task change, if there is a reference pattern that is a standard pattern for the original task and utterance data of the original task and a new task of a small number of reference speakers, they are It is possible to create a highly accurate standard pattern for a new task of the speaker.

In the standard pattern forming apparatus of the second invention,
The reference pattern and the task dependent pattern created by the standard pattern creating apparatus of the first invention are mixed to create a task semi-dependent pattern. The method of mixing is arbitrary, but here, as an example, a method of taking a weighted average of the average vector for each distribution will be described. That is,

[0037]

(Equation 4)

Where 0 <k <1 and

[0039]

(Equation 5)

The value of k may be determined for each distribution, for each state, or for each recognition unit. Alternatively, a common value may be used in all recognition units. In other words, the unit of audio for which one value of k is valid can be arbitrarily set, similarly to the mapping unit. The above corresponds to the task semi-dependent data creating means.

When the number of reference speakers is small, or when the amount of utterance data of task A or task B of the reference speaker is small, the first reference pattern or the second reference pattern is used.
The pattern is not created with high accuracy, and the accuracy of the mapping deteriorates, so that the accuracy of the task-dependent pattern may decrease. In such a case, in the standard pattern creating apparatus of the second invention, since the information of the reference pattern is also used at a certain ratio, the accuracy of the created task semi-dependent pattern may increase. However, if the amount of the utterance data of the reference speaker task A or task B is sufficiently large,
It is presumed that the standard pattern creation device of the first invention has higher accuracy.

[0042]

Embodiments of the present invention will be described below with reference to the drawings. The embodiment corresponds to the example described in the section of the operation, and the notation such as variables is the same as that given there. FIG. 1 is a block diagram showing one embodiment of the standard pattern learning device of the first invention.

The reference pattern creating means 101 receives the utterances of task A of many speakers, learns standard patterns, and outputs the learned standard patterns as reference patterns.

The first reference pattern creating means 102 receives the utterance data of task A of the reference speaker who uttered both task A and task B. Using them, standard patterns for task A for each reference speaker are learned and they are output as reference first patterns.

The reference second pattern creating means 103 receives the utterance data of the task B of the reference speaker. Using them, the standard patterns for task B for each standard speaker are learned and they are output as reference second patterns.

In the mapping creation means 104, the first reference
The pattern and the reference second pattern are input, and a mapping for each distribution between them is created and output.

The task-dependent pattern creation means 105 receives the reference pattern and the mapping created by the mapping creation means 104, creates a task-dependent pattern by using the equation (2) of the action section, and outputs the task-dependent pattern. You.

FIG. 2 is a block diagram showing an embodiment of the standard pattern learning apparatus according to the second invention. The reference pattern creating means 101 receives the utterances of task A of many speakers, learns a standard pattern, and outputs the learned standard pattern as a reference pattern.

The first reference pattern creating means 102 receives the utterance data of task A of the reference speaker who has uttered both task A and task B. Using them, standard patterns for task A for each reference speaker are learned and they are output as reference first patterns.

The second reference pattern creating means 103 receives the utterance data of the task B of the reference speaker. Using them, the standard patterns for task B for each standard speaker are learned and they are output as reference second patterns.

In the mapping creation means 104, the first reference
The pattern and the reference second pattern are input, and a mapping for each distribution between them is created and output.

The reference pattern and the mapping created by the mapping creation means 104 are input to the task-dependent pattern creation means 105, and a task-dependent pattern is created and output using the equation (2) of the action section. You.

In the task semi-dependent pattern creating means 106, a reference pattern and a task dependent pattern are input, and a task semi-dependent pattern is created and output according to the equation (6) of the operation section.

[0054]

As described above, according to the present invention, when a task is changed, a reference pattern which is a standard pattern for the original task and utterances of the original task and the new task by a small number of speakers are provided. If you have data, use them
It is possible to create a standard pattern with higher recognition performance for a new task of many speakers.

In the operation section and the embodiment section, the case where the average vector of the Hidden Markov Model is used as the standard pattern has been described.
The present invention can be applied to a case where another feature amount such as a transition probability or a combination of a plurality of types of feature amounts is used. In addition, when using not only the hidden Markov model but also other methods such as a neural network other than the hidden Markov model as a speech recognition method, if a set of feature amounts is extracted as a standard pattern, the present invention can be applied. Is easy. Therefore, these applications are also within the scope of the present invention.

In the section of the operation and the section of the embodiment, the task has been described by taking only the “limited recognition target” as an example, but in addition, the “microphone used for voice input” or the “ It is easy to consider the magnitude of the ambient noise and the like as tasks, and to apply the present invention to the case where they are different as in the case of the difference of “limited recognition target”. Therefore, these applications are also within the scope of the present invention. Further, when the task is other than “limited recognition target”, it is possible to take a unit other than a subword, for example, a sentence or a word, as a recognition unit, and such a case is also within the scope of the present invention. .

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of a standard pattern creation device of the first invention.

FIG. 2 is a block diagram showing one embodiment of a standard pattern creation device of the second invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 101 Reference pattern creation means 102 Reference first pattern creation means 103 Reference second pattern creation means 104 Mapping creation means 105 Task dependent pattern creation means 106 Task semi-dependent pattern creation means

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-5-246991 (JP, A) JP-A-5-40496 (JP, A) JP-A-3-186899 (JP, A) Acoustical Society of Japan Heisei 2 Proceedings of the Fall Meeting of the Japanese Society of Science and Technology ▲ I ▼ 1-8-12 “Speaker Adaptation for Speech Recognition by Hemisyllable HMM” p. 23-24 (published on September 19, 1990) Proceedings of the Acoustical Society of Japan, Spring Meeting 2004, I-P1-P-15 “Task adaptation for speech recognition” p. 133-134 (Issued March 17, 1992) (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 3/00 521 JICST file (JOIS)

Claims (2)

(57) [Claims] In the speech recognition, when learning a standard pattern using learning data uttered by a plurality of speakers, a reference pattern creating means for creating a reference pattern of a first task; A reference first pattern creating means for learning a standard pattern using utterance data of a first task of a reference speaker who has uttered a second task different from the second task, and utterance of a second task of the reference speaker A reference second pattern creating means for learning a standard pattern using data; a mapping creating means for creating a mapping from the reference first pattern to the reference second pattern; and a mapping creating means from the reference pattern. And a task dependent pattern creating means for creating a task dependent pattern using the created mapping. 2. The reference pattern creating unit, the reference first pattern creating unit, the reference second pattern creating unit, the mapping creating unit, the data dependent pattern creating unit, the reference pattern and the task. 2. The standard pattern creating apparatus according to claim 1, further comprising: task quasi-dependent pattern creating means for creating a task quasi-dependent pattern by mixing the dependency pattern with a certain fixed ratio.