JP2022010410A - Speech recognition device, speech recognition learning device, speech recognition method, speech recognition learning method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a speech recognition device capable of achieving end-to-end speech recognition in consideration of context.

SOLUTION: A speech recognition device includes: a model parameter learning unit which uses a word sequence of interest as an observation value, uses a word sequence older than the word sequence of interest, an acoustic feature sequence corresponding to the word sequence of interest, and a model parameter θ as parameters, and which learns the model parameter θ by performing maximum likelihood estimation for a likelihood function of probability that an observation value occurs under the parameters; and a spoken voice recognition unit which uses a word sequence to be recognized as an observation value, uses an already recognized word sequence older than the word sequence to be recognized, an acoustic feature corresponding to the word sequence to be recognized, and the learned model parameter θ as parameters, and which repeats a process of recognizing the word sequence to be recognized in a chronological order by maximum likelihood reference for the likelihood function of the probability that the observation value occurs under the parameters.

SELECTED DRAWING: Figure 1

COPYRIGHT: (C)2022,JPO&INPIT

Description

The present invention relates to a voice recognition device, a voice recognition learning device, a voice recognition method, a voice recognition learning method, and a program.

With the progress of deep learning technology, a modeling method of speech recognition called end-to-end speech recognition, in which the input is speech and the output is text, has appeared, and technological progress is progressing. There are three types of speech recognition that have been widely used so far: an acoustic model that models the relationship between speech and phoneme sequences, a pronunciation model that models the relationship between phoneme sequences and words, and a language model that models the relationships between words. It was composed of a combination of models, and the speech recognition algorithm (device) was constructed by learning each model independently using different data. On the other hand, in end-to-end speech recognition, a speech recognition algorithm (device) can be configured with only one model that models the relationship between speech and text, and the data used for learning is only paired data of speech and text. Is.

The configuration of the prior art will be described. end-to-end The acoustic feature sequence that can be automatically extracted from the input speech of speech recognition is X = (x ₁ ,…, x _T ), and the output word sequence is W = (w ₁ ,…, w _N ). And model P (W | X, θ). Here, θ represents a model parameter. The modeling of P (W | X, θ) is expressed by the following equation.

In the speech recognition algorithm (device) by this modeling, the word sequence W ^ of the speech recognition result when the acoustic feature sequence X is input is determined based on the following equation.

The model parameter θ is the learning data D = (W ₁ , X ₁ ),…, (W _{| D |} , X _{| D |} ), which consists of a set of multiple (two or more) word sequences and acoustic feature sequence sets. (However, | D | is determined by learning in advance based on the number of elements of the training data D). The parameter θ ^ optimized by D follows the following equation.

Various methods can be adopted for detailed modeling. For example, a method using a neural network is typical, and the methods of Non-Patent Document 1 and Non-Patent Document 2 can be used.

Jan Chorowski, Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio, “End-to-end continuous speech recognition using attention-based recurrent NN: first results,” in NIPS: Workshop Deep Learning and Representation Learning Workshop, 2014. Jan Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk, Kyunghyun Cho, and Yoshua Bengio, “Attention-based models for speech recognition,” in Advances in Neural Information Processing Systems (NIPS), 2015, pp. 577-585.

The above-mentioned conventional technique models the problem of recognizing the voice of a single utterance, and even when recognizing a voice series composed of a continuous multiple utterances, each voice in the multiple utterances is voiced. In recognizing, the relationship between utterances of multiple utterances cannot be used at all. That is, there is a problem that information such as what kind of word sequence has been output for the voice input of the past utterance cannot be taken into consideration when performing the voice recognition of the current utterance.

A specific example will be described. For example, in a scene where a lecture voice of about 10 minutes is recognized by voice, it is assumed that the lecture voice is divided every 0.5 seconds of silence and a total of 200 speech voices are included. These 200 utterances are a continuous series, and it is highly probable that the continuous utterances are utterances about information related to each other. However, when the conventional technique is applied, 200 utterances are recognized independently for each utterance, and context information cannot be used for voice recognition. For example, if the 100th utterance is the utterance "The achievements of this term are wonderful" and the 101st utterance is "Is it because it is wonderful", if the 100th utterance can be considered as the context, 101 There is a high possibility that the utterance can be voice-recognized as "a wonderful achievement", but if the 100th utterance cannot be considered as a context, the 101st utterance can be mistakenly recognized as "a wonderful confectionery" or "a wonderful holy fire". There is sex.

For example, it is assumed that all the utterances (200 utterances in the above example) are collectively treated as one utterance with a long utterance length to solve the above problem. In this case, since the end-to-end speech recognition algorithm (device) is a mechanism that converts the entire speech into a vector and handles it, a problem that the speech with a long speech length does not work well arises. It is unrealistic to treat all utterances as one utterance by the end-to-end speech recognition algorithm (device). Therefore, in the past, there was a problem that end-to-end speech recognition in consideration of context could not be realized.

Therefore, an object of the present invention is to provide a speech recognition device capable of realizing end-to-end speech recognition in consideration of context.

The voice recognition device of the present invention includes an utterance voice recognition unit.

The spoken voice recognition unit uses the word sequence to be recognized as an observation value based on the recognition data consisting of a set of acoustic feature quantity sequences acquired in chronological order, and has already recognized the word sequence to be recognized in the past. The most probable standard for the likelihood function of the probability that an observed value will occur under the parameters, using the already-learned word sequence, the acoustic feature quantity sequence corresponding to the word sequence to be recognized, and the pre-learned model parameter θ as parameters. The process of recognizing the word sequence to be recognized is repeated in chronological order.

According to the speech recognition device of the present invention, end-to-end speech recognition can be realized in consideration of the context.

The block diagram which shows the structure of the voice recognition apparatus of Example 1. FIG. The flowchart which shows the operation of the voice recognition apparatus of Example 1. FIG. The block diagram which shows the structure of the utterance voice recognition part of the voice recognition apparatus of Example 1. FIG. The flowchart which shows the operation of the utterance voice recognition part of the voice recognition apparatus of Example 1. FIG.

Hereinafter, embodiments of the present invention will be described in detail. The components having the same function are given the same number, and duplicate explanations are omitted.

Hereinafter, the voice recognition device 1 (minimum configuration, refer to the configuration in the broken line frame of FIG. 1) of this embodiment will be described. Here, it is assumed that the model parameter θ is learned in advance by a device different from the speech recognition device 1.

In this specification, "^" may be added after the character for the convenience of document creation software, but this "^" is treated as being displayed above the character. For example, when writing W ^L ^

It shall mean.

<Overview of input, output, and operation of voice recognition device 1 (minimum configuration)>
Input 1: Series of acoustic features series of L consecutive utterances X ¹ ,…, X ^L
Input 2: Model parameter θ (learned by another device and input to this device)
Output: L consecutive word sequence series W ¹ ^,…, W ^L ^

In the speech recognition device 1 of the present embodiment, the sequence X ¹ , ..., X ^L of the acoustic feature quantity series of L consecutive utterances and the model parameter θ are input, and L is calculated by the probability calculation according to the model parameter θ. Outputs a series of consecutive word sequences W ¹ ^,…, W ^L ^. Here, let X ¹ , ..., X ^L be the sequence of the acoustic feature sequence that can be automatically extracted from the speech sequence of L consecutive utterances that are the inputs of end-to-end speech recognition. Here, X ^l is the acoustic feature series of the _lth utterance and is expressed as X ^l = (x ^l ₁ ,…, x ^l T l). The output word sequence is W ¹ ^,…, W ^L ^, where W ^l ^ is the lth utterance word sequence.

It is expressed as.

Here, as the acoustic feature quantity series, any feature quantity series that can be calculated from voice can be used, but for example, a feature quantity series such as a mel filter bank cepstrum coefficient or a logarithmic mel filter bank can be used. Melfilter bank The explanation of the cepstrum coefficient and the logarithmic melfilter bank is omitted.

In the case of English, the word sequence may be, for example, a space-separated expression, in the case of Japanese, for example, an expression automatically divided by morphological analysis, or an expression divided into character units may be used as the word sequence.

Next, a general configuration of the voice recognition device of the first embodiment will be described with reference to FIG. Here, it is assumed that the model parameter θ is learned in the speech recognition device 1. As shown in the figure, the voice recognition device 1 of this embodiment includes a model parameter learning unit 11, a model parameter storage unit 11a, an utterance voice recognition unit 12, and a word sequence storage unit 12a. However, as described above, the model parameter learning unit 11 and the model parameter storage unit 11a may be constituent requirements of separate devices. Hereinafter, the operation of each configuration requirement will be described with reference to FIG.

<Model parameter learning unit 11>
Input: Learning data D = (A ₁ , B ₁ ),…, (A _{| D |} , B _{| D |} )
Output: Model parameter θ

The model parameter learning unit 11 has training data D = (A ₁ , B ₁ ), ... Based on (A _{| D |} , B _{| D |} ), the word sequence of interest (W ^l of the following equation) is used as the observed value, and the word sequence of the past than the word sequence of interest (W ¹ of the following equation) , ..., W ^l-1 ), the acoustic feature sequence (X ^l in the following equation) corresponding to the word sequence of interest (W ^l in the following equation), and the model parameter θ as parameters ( A model by performing maximum likelihood estimation for the likelihood function of the probability that an observed value (W ^l in the following equation) will occur under W ¹ , ..., W ^l-1 , X ^l , θ in the following equation. The parameter θ is learned (S11). Note that (A _m , B _m ) = {(W ¹ , X ¹ ),…, (W ^Lm , X ^Lm )}. The parameter θ ^ optimized by D follows the following equation.

The θ ^ learned here is used as θ in the spoken voice recognition unit 12.

<Model parameter storage unit 11a>
The model parameter storage unit 11a stores the learned θ ^.

<Speech voice recognition unit 12>
Input 1: Acoustic feature series of l-th utterance X ^l
Input 2: The word sequence from the first to the l-1st utterance already obtained as a speech recognition result W ¹ ^,…, W ^l-1 ^
Input 3: Model parameter θ
Output: lth utterance word sequence W ^l ^

The spoken voice recognition unit 12 is based on the recognition data consisting of a set of acoustic feature quantity series (X ¹ , ..., ^{XL) acquired in chronological order, and the word sequence to be recognized (W l of} the following equation). ) Is the observed value, and the already recognized word sequence (W ¹ ^,…, W ^l-1 ^ in the following equation) prior to the word sequence to be recognized (W ^l in the following equation), and the recognition target Using the acoustic feature sequence (X ^l in the following equation) corresponding to the word sequence (W ^l in the following equation) and the trained model parameter θ as parameters, the parameters (W ¹ ^,…, W ^l in the following equation) For the likelihood function of the probability that the observed value (W ^l in the following equation) occurs under ^-1 ^, X ^l , θ), the word sequence to be recognized (W ^l ^ in the following equation) according to the most likely criterion. The process of recognizing the above is repeated in chronological order (S12).

That is, the utterance voice recognition unit 12 has the acoustic feature sequence X ^l of the l-th utterance and the recognized word sequence W ¹ ^, ..., From the first to the l-1th utterance obtained as the voice recognition result. Posterior probability distribution for the lth utterance by probabilistic calculation according to the model parameter θ when W ^l-1 ^ is input

And determine the word sequence W ^l ^ of the speech recognition result of the lth utterance by the maximum likelihood criterion. That is, the determination based on the maximum likelihood criterion follows the following equation.

As described above, the spoken voice recognition unit 12 recursively executes step S12 in chronological order. For example, by setting the word sequence W ^l ^ of the speech recognition result of the l-th utterance as a known recognition result, the posterior probability distribution for the l + 1-th utterance

And similarly, the word sequence W ^{l + 1} ^ of the speech recognition result of the l + 1st utterance is determined as follows.

note that,

The detailed formulation and detailed calculation method of are described later.

<Word sequence storage unit 12a>
The word sequence storage unit 12a stores the word sequence recursively used by the utterance speech recognition unit 12. For example, in step S12, when the word sequence W ¹ ^ is recognized, the word sequence storage unit 12a stores the word sequence W ¹ ^, and when the word sequence W ^l ^ is recognized, the word sequence storage unit 12a Stores the word sequence W ^l ^, and when the word sequence W ^L ^ is recognized, the word sequence storage unit 12a stores the word sequence W ^L ^.

<Detailed configuration of the spoken voice recognition unit 12>
As shown in FIG. 3, the utterance speech recognition unit 12 includes an utterance vector calculation unit 121, an utterance series embedded vector calculation unit 122, a context vector calculation unit 123, and a posterior probability calculation unit 124.

As described above, the utterance voice recognition unit 12 has the utterance voice recognition unit 12.

To calculate. This detailed formulation is expressed by the following equation.

note that,

Is realized by the utterance vector calculation unit 121 in the utterance voice recognition unit 12, the utterance series embedded vector calculation unit 122, the context vector calculation unit 123, and the posterior probability calculation unit 124. In the following, referring to FIG. 4, the probability for the nth word of the lth utterance

Represents the detailed processing for calculating.

<Utterance vector calculation unit 121>
Input 1: l-1 word sequence of the first utterance W ^l-1 ^
Input 2: Model parameter θ
Output: l-1 utterance vector of the first utterance u ^l-1

The utterance vector calculation unit 121 converts the already recognized l-1st utterance word sequence W ^l-1 ^, which is earlier than the lth utterance word sequence W ^l to be recognized, based on the model parameter θ. The function converts the utterance vector u ^l-1 of the l-1st utterance (S121). At this time, the word sequence W ^l-1 ^ of the l-1st utterance contains one or more words. The utterance vector represents a vector in which information contained in the word sequence is embedded, and semantic information of the utterance necessary for voice recognition of the next utterance is embedded. The larger the number of dimensions of a vector, the more information can be embedded. For example, the number of dimensions is manually determined as a 512-dimensional vector. At this time, any conversion function can be used as long as it is a function that converts a variable-length symbol string into a single vector. For example, a function that constitutes a frequency vector of a spoken word should be used. A recurrent neural network, a bidirectional recurrent neural network, or the like can also be used.

Note that when l = 1, the input word sequence W ⁰ does not exist, so u ⁰ in the output may be a vector with all elements 0.0.

Note that step S121 is performed for each of W ¹ ^, ..., W ^l-1 ^. Therefore, the utterance vector calculation unit 121 outputs u ¹ , ..., U ^l-1 respectively.

<Utterance series embedded vector calculation unit 122>
Input 1: A series of utterance vectors for past utterances u ¹ ,…, u ^l-1
Input 2: Model parameter θ
Output: l-1st utterance series embedded vector v ^l-1

The utterance series embedded vector calculation unit 122 uses a transformation function based on the model parameter θ to convert the utterance vector series u ¹ , ..., u ^l-1 for past utterances into the l-1st utterance series embedded vector v ^l-1 . Is converted to (S122). This utterance series embedded vector is a single vector, and semantic information necessary for speech recognition of the next utterance is embedded. The larger the number of dimensions of a vector, the more information can be embedded. For example, the number of dimensions is manually determined as a 512-dimensional vector. At this time, any conversion function can be used as long as it is a function that converts a variable-length vector sequence into a single vector, such as a recurrent neural network or averaging each vector of an utterance vector series. Functions can be used. In the case of averaging, the number of dimensions of the utterance series embedded vector depends on the number of dimensions of the utterance vector series.

When l = 1, there is no utterance vector series for the past utterance series that is the input, so v ⁰ of the output should be a vector with all elements 0.0.

<Context vector calculation unit 123>
Input 1: The nth word in the l-th utterance word sequence W ^l w ^l A word sequence older than _n w ^l ₁ ,…, w ^l _n-1
Input 2: Acoustic feature series of l-th utterance X ^l
Input 3: Model parameter θ
Output: Context vector for the nth word of the lth utterance s ^l _n

The context vector calculation unit 123 is a word sequence w ^l ₁ ,…, w ^l _n-1 (word sequence) that is older than the nth word w ^l _n in the word sequence W ^l of the lth utterance to be recognized. The l-th acoustic feature sequence X ^l corresponding to the l-th word sequence W ^l to be recognized is uttered by the conversion function based on the model parameter θ. Convert to the context vector s ^l _n for the _nth word w ^l n in the word sequence W ^l of (S123). This context vector is embedded with information that integrates semantic information and phonological information necessary for speech recognition of the next word. At this time, any conversion function can be used as long as it is a function that converts two types of variable-length vector sequences into a single vector. For example, as in Non-Patent Document 2, an acoustic feature sequence and a word sequence can be used. It is also possible to use a function expressed as a single context vector by providing a recurrent neural network for each of the above and adding an attention mechanism. Also, if it is the simplest, construct a combined vector of the frequency vector of the word sequence past the nth word of the lth utterance and the acoustic feature sequence of the lth utterance averaged. Functions can also be used.

<Posterior probability calculation unit 124>
Input 1: l-1st utterance series embedded vector v ^l-1
Input 2: Context vector for the nth word of the lth utterance s ^l _n
Input 3: Model parameter θ
Output: Posterior probabilities for the nth word of the lth utterance

The posterior probability calculation unit 124 converts the utterance vector sequence u ¹ , ..., u ^l-1 up to the past one from the word sequence W ^l to be recognized, and l-1st utterance sequence embedded vector v ^l . ^-1 and the nth word of the lth word sequence W ^l from the context vector s ^l _n for the nth word of the lth word sequence W ^l to be recognized by the conversion function based on the model parameter θ Posterior probability about

Is calculated (S124). The posterior probability can be expressed as a vector with each word as an element, and the posterior probability distribution can be expressed by vector transformation. At this time, any conversion function can be used as long as it is a function that converts two types of vectors into a posterior probability distribution. It can be realized by the function to be performed. Other than that, a function that can convert the sum of the elements of the output vector corresponding to the posterior probability distribution to 1.0 is applicable.

According to the speech recognition device 1 of the present embodiment, since the modeling of end-to-end speech recognition that handles the utterance sequence is introduced instead of the conventional end-to-end speech recognition that handles single utterances, speech. Context-aware end-to-end speech recognition can be achieved when the input is represented as an utterance sequence. That is, when recognizing a certain utterance in the utterance series by voice, it is possible to use the information from the first utterance of the utterance series to the utterance immediately before the target utterance as a context. For example, as described above, it is assumed that the lecture voice of about 10 minutes is recognized by voice, and if this lecture voice is divided every 0.5 seconds of silence, it is assumed that 200 speech voices are included. In this case, according to the speech recognition device 1 of the present embodiment, all the relevant context information before a certain utterance in 200 consecutive utterances can be used for the current speech recognition. For example, the voice recognition device 1 can use the voice recognition results from the first utterance to the 99th utterance as a context when recognizing the 100th utterance.

The voice recognition device 1 of the present embodiment can enhance the recognition performance of voice recognition such as lectures, telephone calls, and conferences.

<Supplementary note>
The device of the present invention is, for example, as a single hardware entity, an input unit to which a keyboard or the like can be connected, an output unit to which a liquid crystal display or the like can be connected, and a communication device (for example, a communication cable) capable of communicating outside the hardware entity. Communication unit, CPU (Central Processing Unit, cache memory, registers, etc.) to which can be connected, RAM and ROM as memory, external storage device as hard hardware, and input, output, and communication units of these. , CPU, RAM, ROM, has a bus connecting so that data can be exchanged between external storage devices. Further, if necessary, a device (drive) or the like capable of reading and writing a recording medium such as a CD-ROM may be provided in the hardware entity. As a physical entity equipped with such hardware resources, there is a general-purpose computer or the like.

The external storage device of the hardware entity stores a program required to realize the above-mentioned functions and data required for processing of this program (not limited to the external storage device, for example, reading a program). It may be stored in a ROM, which is a dedicated storage device). Further, the data obtained by the processing of these programs is appropriately stored in a RAM, an external storage device, or the like.

In the hardware entity, each program stored in the external storage device (or ROM, etc.) and the data required for processing of each program are read into the memory as needed, and are appropriately interpreted and executed and processed by the CPU. .. As a result, the CPU realizes a predetermined function (each configuration requirement represented by the above, ... Department, ... means, etc.).

The present invention is not limited to the above-described embodiment, and can be appropriately modified without departing from the spirit of the present invention. Further, the processes described in the above-described embodiment are not only executed in chronological order according to the order described, but may also be executed in parallel or individually as required by the processing capacity of the device that executes the processes. ..

As described above, when the processing function in the hardware entity (device of the present invention) described in the above embodiment is realized by the computer, the processing content of the function that the hardware entity should have is described by the program. Then, by executing this program on the computer, the processing function in the above hardware entity is realized on the computer.

The program describing the processing content can be recorded on a computer-readable recording medium. The recording medium that can be read by a computer may be, for example, a magnetic recording device, an optical disk, a photomagnetic recording medium, a semiconductor memory, or the like. Specifically, for example, a hard disk device, a flexible disk, a magnetic tape or the like as a magnetic recording device, and a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only) as an optical disk. Memory), CD-R (Recordable) / RW (ReWritable), etc., MO (Magneto-Optical disc), etc. as a magneto-optical recording medium, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. as a semiconductor memory. Can be used.

Further, the distribution of this program is performed, for example, by selling, transferring, renting, or the like a portable recording medium such as a DVD or a CD-ROM in which the program is recorded. Further, the program may be stored in the storage device of the server computer, and the program may be distributed by transferring the program from the server computer to another computer via the network.

A computer that executes such a program first temporarily stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. Then, when the process is executed, the computer reads the program stored in its own recording medium and executes the process according to the read program. Further, as another execution form of this program, a computer may read the program directly from a portable recording medium and execute processing according to the program, and further, the program is transferred from the server computer to this computer. You may execute the process according to the received program one by one each time. In addition, the above-mentioned processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and the result acquisition without transferring the program from the server computer to this computer. May be. The program in this embodiment includes information to be used for processing by a computer and equivalent to the program (data that is not a direct command to the computer but has a property that regulates the processing of the computer, etc.).

Further, in this form, the hardware entity is configured by executing a predetermined program on the computer, but at least a part of these processing contents may be realized in terms of hardware.

Claims (7)

Based on the recognition data consisting of a set of acoustic feature quantity series acquired in chronological order, the word sequence to be recognized is used as the observed value, and the already recognized word sequence and the word sequence to be recognized are older than the word sequence to be recognized. The acoustic feature quantity series corresponding to the word sequence to be recognized and the model parameter θ learned in advance are used as parameters, and the probability function of the probability that the observed value occurs under the parameters is recognized by the most likely criterion. A voice recognition device including a spoken voice recognition unit that repeats the process of recognizing a word sequence in chronological order. The voice recognition device according to claim 1.
The utterance voice recognition unit is
An utterance that converts an already recognized word sequence past the recognition target word sequence into an utterance vector, which is a vector containing semantic information necessary for speech recognition of the next utterance, by a conversion function based on the model parameter θ. Vector calculator and
An utterance series embedded vector calculation unit that converts a series of utterance vectors into a utterance series embedded vector containing semantic information necessary for speech recognition of the next utterance by a conversion function based on the model parameter θ.
The model parameter is a word sequence in the word sequence to be recognized and an acoustic feature quantity sequence corresponding to the word sequence to be recognized, which is older than the word of interest in the word sequence to be recognized. A context vector calculation unit that converts semantic information and phonological information necessary for speech recognition of words in the word sequence to be recognized into a context vector containing integrated information by a conversion function based on θ.
Conversion based on the model parameter θ from the utterance sequence embedding vector formed by converting the utterance vector sequence up to the past from the word sequence to be recognized and the context vector for words in the word sequence to be recognized. A speech recognition device that includes a post-probability calculator that calculates the post-probability of a word in a word series to be recognized by a function. Based on the learning data consisting of the set of the word sequence acquired in chronological order and the set of the corresponding acoustic feature quantity series, the word sequence of interest is used as the observed value, and the word sequence past the word sequence of interest is used. , And the acoustic feature quantity sequence corresponding to the word sequence of interest, and the model parameter θ as parameters, and the maximum likelihood estimation is performed for the likelihood function of the probability that the observed value will occur under the parameters. A voice recognition learning device including a model parameter learning unit that learns the model parameter θ. Based on the recognition data consisting of a set of acoustic feature quantity series acquired in chronological order, the word sequence to be recognized is used as the observed value, and the already recognized word sequence and the word sequence to be recognized are older than the word sequence to be recognized. The acoustic feature quantity series corresponding to the word sequence to be recognized and the model parameter θ learned in advance are used as parameters, and the probability function of the probability that the observed value occurs under the parameters is recognized by the most likely criterion. A voice recognition method including a spoken voice recognition step in which the process of recognizing a word sequence is repeated in chronological order. Based on the learning data consisting of the set of the word sequence acquired in chronological order and the set of the corresponding acoustic feature quantity series, the word sequence of interest is used as the observed value, and the word sequence past the word sequence of interest is used. , And the acoustic feature quantity sequence corresponding to the word sequence of interest, and the model parameter θ as parameters, and the maximum likelihood estimation is performed for the likelihood function of the probability that the observed value will occur under the parameters. A speech recognition learning method that includes a model parameter learning step to learn the model parameter θ. A program that causes a computer to function as the voice recognition device according to claim 1 or 2. A program that causes a computer to function as the voice recognition learning device according to claim 3.

Images