JP5975938B2 - Speech recognition apparatus, speech recognition method and program - Google Patents

Description

  The present invention relates to a speech recognition technology, and more particularly to a technology for performing speech recognition at high speed using a latent word language model.

  Speech recognition technology requires a language model for linguistic prediction. In the field of speech recognition, an N-gram language model is generally used as a language model. The N-gram language model is very compatible with speech recognition decoding. The N-gram language model can be easily learned if there is a learning text. A method for learning an N-gram language model is described in Non-Patent Document 1 and the like.

  Various language models other than the N-gram language model have been proposed. However, considering compatibility with speech recognition decoding, language models other than the N-gram language model are rarely used in speech recognition. This is because it is difficult for a language model other than the N-gram language model to perform the search process executed at the time of speech recognition in a realistic time.

  Although the N-gram language model is generally used in speech recognition, there is a problem. For example, the N-gram language model can only learn language phenomena that exist in the learning text. An example will be described. If there is "eating apples" in the learning text and "eating tangerines" does not exist, the probability of "eating tangerines" is basically calculated using the N-gram language model constructed with the learning text. Only uses "eating" information. However, “mandarin orange” and “apple” are usually similar words, and the probability of “eating an apple” is often considered to be better to use the information of “eating an orange”.

  As a language model that uses similar word information, there is a latent language model described in Non-Patent Document 2. The latent language model is a model that considers latent words for each word in the learning text. As the latent language model, a probability model can be constructed in consideration of, for example, that “apple” and “mandarin” are similar words. Non-Patent Document 2 reports that the latent language model has higher language prediction performance than the N-gram language model.

  The structure of the latent language model is the same as that of the class N-gram language model described in Non-Patent Document 3. The class N-gram language model is not a general N-gram language model, but is expressed as an N-gram language model related to the class sequence by dropping words into a form called a class. Here, the class is a concept in which a set of words is classified according to some standard, and for example, a fruit name class, a surname class, or the like can be considered. The class N-gram language model further has a distribution of intra-class word occurrence probabilities. For example, in the case of a fruit name class, the probability of occurrence of an in-class word is a value such as 0.2 for the occurrence of “apple” and 0.1 for the occurrence of “mandarin”.

Kenji Kita, “Language and Computation-4 Stochastic Language Model”, The University of Tokyo Press, November 1999, pp. 57-62 K. Deschacht, J. D. Belder, M-F. Moens, “The latent words language model”, Computer Speech and Language, vol.26, pp.384-409, 2012. Kenji Kita, “Language and Computation-4 Stochastic Language Model”, The University of Tokyo Press, November 1999, pp. 72-74

  The latent language model is not used for speech recognition or machine translation. It is unrealistic to directly use a latent language model as a substitute for a conventional general N-gram language model in speech recognition. The reason is that the calculation time for decoding is significantly increased. This is because the latent language model has a structure as a soft clustering class N-gram language model.

  A general class N-gram language model also used for speech recognition has a hard clustering structure. A hard clustering structure is a structure in which a word belongs to only one class. On the other hand, the structure of soft clustering refers to a structure in which one word can belong to a plurality of classes. In the case of a latent language model, it is a structure in which one word belongs to all classes. In the case of hard clustering, the class sequence can be uniquely determined by looking at a certain word string. However, in the case of soft clustering, there are a huge number of classes for a certain word string. In speech recognition, a process for obtaining a probability value for an arbitrary word string using a language model is required. However, in a latent language model, an enormous amount of calculation is required, and use for speech recognition is unrealistic. .

  When a latent language model is used for speech recognition, the problem of finding a class is very difficult because one word belongs to all classes. The method of performing speech recognition while finding a class is called Viterbi search. Viterbi search is a well-known technique that realizes simultaneous decoding of word sequences and latent word sequences, and is used in various decoding problems.

  An object of the present invention is to practically use a latent language model having excellent language prediction performance for speech recognition. More specifically, the problem of the enormous amount of calculation required for the Viterbi search is greatly solved, and the latent language model is configured to be used for the Viterbi search.

  In order to solve the above problems, a speech recognition apparatus according to the present invention includes a latent language model storage unit, a baseline language model storage unit, a multiple hypothesis generation unit, a latent word sequence determination unit, a score recalculation unit, and a first hypothesis. Includes a decision part. The latent word language model storage unit includes a probability distribution of learning latent word sequences that are latent word sequences corresponding to observed word sequences included in the learning text, and probability distributions of observed words in the observed word sequences and latent words in the latent word sequences. The latent language model that learned The baseline language model storage unit stores a baseline language model in which the probability distribution of the latent word sequence and the probability distribution of the observed word sequence included in the latent word language model are mixed. The multiple hypothesis generation unit recognizes input speech using the baseline language model, and generates a plurality of speech recognition result hypotheses and a provisional speech recognition score for each hypothesis. The latent word series determination unit determines a hypothetical latent word series that is a latent word series corresponding to a hypothesis using a latent word language model, and obtains a joint probability of the hypothesis and the hypothetical latent word series. The score recalculation unit obtains a speech recognition score using the provisional speech recognition score and the joint probability. The first hypothesis determination unit determines a speech recognition result for the input speech based on a speech recognition score from a plurality of hypotheses.

  Since the speech recognition technology of the present invention performs speech recognition using a latent language model having excellent language prediction performance, the speech recognition technology uses speech recognition using an N-gram language model constructed by a conventional general learning method. High language prediction performance can be obtained. Moreover, at the time of speech recognition, by first narrowing down speech recognition result candidates using a baseline language model having a general N-gram language model structure, and then performing a Viterbi search using a latent language model, Compared with the case where the Viterbi search is simply performed, the amount of calculation can be greatly reduced. This makes it possible to realize speech recognition using a latent language model even with a normal computer. Therefore, according to the speech recognition technology of the present invention, speech recognition using a latent language model having excellent language prediction performance can be performed at high speed.

FIG. 1 is a diagram for explaining a latent word series and an observed word series. FIG. 2 is a diagram illustrating a functional configuration of the speech recognition apparatus. FIG. 3 is a diagram illustrating a processing flow of the speech recognition method. FIG. 4 is a diagram for explaining the data flow of the speech recognition search.

[Points of Invention]
Prior to the description of the embodiments, the points of the present invention will be described.

  In this invention, focusing on the fact that the latent language model is a probabilistic generation model, a general word N-gram language model is constructed from the pseudo-learning text generated according to the probability process of the latent language model. That is, the properties of the latent language model are approximately expressed as an N-gram language model. In addition, by combining the constructed N-gram language model with the N-gram language model learned directly from the learning text by a general method, a language model utilizing both properties is constructed. For example, even if “eating apples” exists in the learning text but “eating tangerines” does not exist, once the latent language model is generated and the text is generated, the text “eating tangerines” Is a framework that can be generated. Therefore, if an N-gram language model is learned from a learning text including “eating mandarin oranges”, an N-gram language model can be constructed in consideration of the similarity between “mandarin oranges” and “apples”. If it can be expressed as a simple N-gram, it can be mixed with an N-gram language model constructed by a general method and expressed as a single language model. This mixed language model can be expected to be a language model that complements the characteristics of the latent language model and the characteristics of a simple N-gram language model.

  In the present invention, in order to perform a Viterbi search using a latent language model at high speed, a candidate word sequence to be output is narrowed down, and a latent word sequence hidden behind each word sequence is estimated. In the stage of narrowing down word sequence candidates, a general N-gram language model is used. At the stage of estimating the latent word sequence hidden behind each word sequence, a latent language model is used. Finally, the speech recognition result is determined with reference to the simultaneous probability of the hidden latent word sequence and the actual word sequence.

[Latent language model]
The latent word corresponds to a class in the class N-gram language model, and represents a representative word when words having similar meanings and syntactic roles are grouped in a certain context. FIG. 1 illustrates the relationship between latent words and observed words. An observation word is a word that appears in the learning text, and is an individual word that constitutes a sentence such as “Today is a good weather” as shown in FIG. In the following, a sequence of observation words is called an observation word sequence. For this observed word, the latent word is represented using a representative word of a word similar to the observed word. The observation word “today” corresponds to the latent word “tomorrow” and is represented as a representative of a plurality of similar words such as “tomorrow”, “yesterday”, and “today”.

  For example, when a technical word and a general word have roles similar in context, the latent word language model can be learned to have a probability that is close to one latent word. Therefore, the latent language model can avoid the problem of data sparseness as compared with the normal N-gram language model, and can be expected to operate robustly for various tasks.

  However, the latent language model has a problem that the calculation cost greatly increases when decoding speech recognition as described above. In the present invention, a pseudo-learning text is once created from the latent language model, and an N-gram language model is created from the pseudo-learning text, so that the latent language model is approximated as a simple N-gram language model. .

Embodiment
Hereinafter, embodiments of the present invention will be described in detail. In addition, the same number is attached | subjected to the component which has the same function in drawing, and duplication description is abbreviate | omitted.

  With reference to FIG. 2, an example of a functional configuration of the speech recognition apparatus according to the embodiment will be described. The speech recognition device 1 includes a learning text storage unit 10, a latent language model learning unit 12, a latent language model storage unit 14, a baseline language model learning unit 16, a baseline language model storage unit 18, a plurality of temporary generation units 20, A latent word series determination unit 22, a score recalculation unit 24, and a first hypothesis determination unit 26 are included.

  The speech recognition device 1 is a special configuration in which a special program is read into a known or dedicated computer having a central processing unit (CPU), a main storage device (Random Access Memory, RAM), and the like. Device. For example, the speech recognition apparatus 1 executes each process under the control of the central processing unit. The data input to the speech recognition device 1 and the data obtained in each process are stored in, for example, a main storage device, and the data stored in the main storage device is read out as needed and used for other processing. The Each storage unit included in the speech recognition device 1 includes, for example, a main storage device such as a RAM (Random Access Memory), an auxiliary storage device configured by a semiconductor memory element such as a hard disk, an optical disk, or a flash memory, or It can be configured with middleware such as a relational database or key-value store. Each storage unit included in the speech recognition device 1 may be logically divided, and may be stored in one physical storage device.

  With reference to FIG. 3, an example of the processing flow of the speech recognition method executed by the speech recognition apparatus 1 will be described according to the order of procedures actually performed.

  The learning text storage unit 10 stores learning texts prepared in advance. The learning text is a text used for learning a language model such as a latent language model or a normal N-gram language model. In the present invention, it is assumed that the learning text is divided in units of words.

  In step S12, the latent language model learning unit 12 reads the learning text stored in the learning text storage unit 10, and the latent word word sequence corresponding to the observed word sequence that is the observed word sequence included in the learned text. A latent word language model is generated by learning the probability distribution of the learning latent word sequence and the probability distribution of the observed word and the latent word. The generated latent language model is stored in the latent language model storage unit 14.

The latent language model has _two probability distributions, a probability distribution of P (h _k | h _k−2 , h _k−1 ) and a probability distribution of P (w _k | h _k ). Here, h is a latent word and w is an observed word. The latent word h corresponds to a latent variable in the latent language model, and the observed word w represents a word that actually appears in the learning text. The probability distribution P (h _k | h _k-2 , h _k-1 ) is in the form of a general word N-gram language model, and the probability distribution P (w _k | h _k ) is in the form of a unigram language model. ing. This is the same form as the general class N-gram language model, and the latent word corresponds to a class in the class N-gram language model.

The learning of the latent language model is a problem of estimating the allocation of latent words for each observed word of the input learning text. In other words, as (the L total number of words contained in a learning text) "w ₁ w ₂ ...... w _L" there is a learning text (series of observation words), "w _1", "w _2" ... of "w _L" It can be said that the latent words “h ₁ ”, “h ₂ ”... “H _L ” corresponding to each observation word are estimated. If this allocation can be estimated, the probability distribution P (h _k | h _k-2 , h _k-1 ) can be constructed by learning an N-gram language model for the latent word sequence “h ₁ h ₂ ... H _L ”, A probability distribution P (w _k | h _k ) can be constructed by learning a unigram language model for “h ₁ → w ₁ ”, “h ₂ → w ₂ ”... “H _L → w _L ”.

A specific latent word assignment can be estimated by a method called Gibbs sampling described in Non-Patent Document 2 and the like. If the set of observed words included in the learning text is W, each latent word is assigned one of the words included in W. For example, if the learning text (observed word sequence) is a sentence consisting of four words “apple, orange, pine,” W is a set of four words, “apple,” “mandarin,” “pine,” and “is”. is there. In this case, the latent word sequence of the learning text is “apple apple apple orange”, “mandarin orange apple orange” or “pine pine pine”. In other words, the latent word sequence candidates are all combinations of latent words. For example, if there are J types of latent words in M word sequences, M ^J (M to the power of J) latent word sequence candidates are obtained. Will exist. Each latent word sequence candidate has a probability that the latent word sequence is actually a correct answer sequence. The process of determining a more probable one from the plurality of candidates can be performed by Gibbs sampling. For example, in Non-Patent Document 2, a single latent word sequence is finally selected. The latent word sequence to be selected may be a latent word sequence finally obtained when Gibbs sampling is repeated.

In the above example, the meaning of the two probability distributions will be explained using the case where “apple apple apple is finally” selected as a latent word sequence as an example. As described above, if a latent word sequence is determined, a probability distribution P (h _k | h _k−2 , h _k−1 ) and a probability distribution P (w _k | h _k ) can be obtained. In this example, since the latent word series is “apple apple apple”, the word “apple” is represented as a fruit name. In other words, the probability distribution P (h _k | h _k−2 , h _k−1 ) has a distribution in which the fruit name often appears after the fruit name. The probability distribution P (w _k | h _k ) indicates that the fruit name entity is actually “apple”, “mandarin orange”, and “pine”.

In the Gibbs sampling described in Non-Patent Document 2, the assignment of latent word sequences is uniquely determined. However, it is possible to have multiple latent words to assign. That is, a plurality of latent word series candidates can be used. For example, when using B (≧ 2) candidates, there are B sets of probability distribution P (h _k | h _k-2 , h _k-1 ) and probability distribution P (w _k | h _k ). Will do. For example, if B = 2, a set of probability distribution P ₁ (h _k | h _k-2 , h _k-1 ) and probability distribution P ₁ (w _k | h _k ) and probability distribution P ₂ (h _k | h _k-2 , h _k-1 ) and probability distribution P ₂ (w _k | h _k ). The value B is given manually in advance. To make the language model more detailed, increase the value of B. If the value of B is about 10, it can be said that it is somewhat detailed. When a plurality of latent word sequence candidates are used, the last B assignments in Gibbs sampling may be used as latent word sequence candidates.

  In step S16, the baseline language model learning unit 16 reads the learning text stored in the learning text storage unit 10 and the latent language model stored in the latent language model storage unit 14, and is included in the latent language model. The probability distribution of the latent word sequence is learned, the probability distribution of the observed word sequence included in the learning text is learned, and these probability distributions are mixed to generate a baseline language model.

First, the baseline language model learning unit 16 creates an N-gram language model by counting frequencies for all combinations of N sets of words in the read learning text. Refer to Non-Patent Document 1 for the learning method of the N-gram language model. For example, a baseline language model having a trigram structure gives a probability distribution of P _base (w _k | w _k−2 , w _k−1 ). This represents the probability of appearance of the word w _k after the word strings w _k-2 and w _k-1 appear.

Next, the baseline language model learning unit 16 generates a latent word sequence from the probability distribution P (h _k | h _k−2 , h _k−1 ) using the read latent word language model, and the latent word A pseudo learning text is created from the sequence and the probability distribution P (w _k | h _k ).

The pseudo-learning text is created as follows, for example. First, the baseline language model learning unit 16 generates a uniform random number rand _h1 and determines the first latent word h ₁ according to P (h ₁ | −, −). The latent word h ₁ is determined by the relationship between the uniform random number and the probability of the latent word h ₁ . For example, assuming P (apple |-,-) = 0.3, P (mandarin |-,-) = 0.3, P (pine |-,-) = 0.3, P (is |-,-) = 0.1, When rand _h1 = 0.1, the first latent word h ₁ = “apple” is determined. Similarly, if rand _h1 = 0.4, the latent word h ₁ = “Mikan”, if rand _h1 = 0.7, the latent word h ₁ = “pine”, and if rand _h1 = 0.95, the latent word h ₁ = “is”. To decide. Next, the baseline language model learning unit 16 generates a uniform random number rand _w1 and determines the observation word w ₁ with reference to P (w ₁ | h ₁ ). The observation word w ₁ is determined by the relationship between the uniform random number and the probability P (w ₁ | h ₁ ). For example, assuming P (apple | h ₁ ) = 0.3, P (mandarin | h ₁ ) = 0.3, P (pine | h ₁ ) = 0.3, P (is | h ₁ ) = 0.1, rand _w1 = 0.1 In this case, the observation word w ₁ = “apple” is determined. Similarly, when rand _w1 = 0.4, the observation word w ₁ = “Mikan”, when rand _w1 = 0.7, the observation word w ₁ = “pine”, and when rand _w1 = 0.95, the observation word w ₁ = “is”. To decide. The determined observation word w ₁ is output to the pseudo learning text. By repeating this series of processing a predetermined number of times, a pseudo learning text is generated. The larger the number of pseudo-learning texts to be output, the more pseudo-learning texts that better represent the properties of the latent language model.

In order to use a plurality of latent word sequence candidates, a plurality of probability distributions P _b (h _k | h _k-2 , h _k-1 ) and probability distributions P _b (w _k | h _k ) are set. If (1 ≦ b ≦ B), Upon generation of the learning text, further uniform random number rand _p is generated, to select a set of probability distributions utilized in relation to the uniform random number rand _p and the inverse of the value B What should I do? For example, if B = 2 and rand _p = 0.3, then using probability distribution P ₁ (h _k | h _k-2 , h _k-1 ) and probability distribution P ₁ (w _k | h _k ) The latent word h _k and the observation word w _k may be determined.

  Subsequently, the baseline language model learning unit 16 creates an N-gram language model by counting frequencies for all combinations of N sets of words in the generated pseudo learning text. Refer to Non-Patent Document 1 for the learning method of the N-gram language model.

  Then, the baseline language model learning unit 16 generates a baseline language model by weighted sum of the N-gram language model generated based on the learning text and the N-gram language model generated based on the latent language model. . The generated baseline language model is stored in the baseline language model storage unit 18.

  Hereinafter, details of processing for performing speech recognition using the latent language model and the baseline language model generated as described above will be described. In the following description, description will be made with reference to the example shown in FIG. 4 as appropriate.

  In step S <b> 20, an audio signal is input to the multiple hypothesis generation unit 20. The multiple hypothesis generation unit 20 reads the baseline language model from the baseline language model storage unit 18, recognizes the input speech using the baseline language model, and provides a plurality of speech recognition result hypotheses and a speech recognition score for each hypothesis. A temporary speech recognition score is generated. The generated hypothesis is input to the latent word sequence determination unit 22. The generated hypothesis and provisional speech recognition score are input to the score recalculation unit 26.

  Since the baseline language model is a normal N-gram language model, an arbitrary speech recognition decoding algorithm can be used. For example, a one-pass frame synchronization beam search can be used. Details on 1-pass frame-synchronized beam search include “Kiyohiro Shikano, Katsunobu Ito, Tatsuya Kawahara, Kazuya Takeda, Mikio Yamamoto,“ IT Text Speech Recognition System ”, Ohm Publishing Co., Ltd., pp.110. Please refer to.

  The multiple hypothesis generation unit 20 outputs various speech recognition results with scores. As shown in FIG. 4, multiple hypotheses for a certain input speech are “Today is sunny: −15000”, “Today is sunny: −18000”, “Today is rare: −20000”, and so on. A character string in front of “: (colon)” is a hypothesis of a speech recognition result, and a numerical value behind is a provisional speech recognition score. Here, the reason why the provisional speech recognition score is negative will be described. In speech recognition, the logarithmic value of probability is generally used as a speech recognition score. A logarithmic value greater than 0 and less than 1 is negative. Therefore, the provisional speech recognition score is a negative value.

  Multiple hypotheses are called N-best. N best represents a hypothesis of a speech recognition result in which N hypotheses are selected from those having a good speech recognition score. The larger the value of the speech recognition score, the better the hypothesis. For example, the 10 best represents the top 10 hypothesis score. The value of N is set manually. For example, 1000 is set. This value is ideally very large, but is generally from 1000 to 5000 in consideration of the calculation amount. Finally, the multiple hypothesis generation unit 20 obtains and outputs a set of N hypotheses and provisional speech recognition scores, such as “Today is sunny: −15000” in the above example.

  In step S22, the latent word series determination unit 22 reads the latent word language model from the latent word language model storage unit 14, and uses the latent word language model to determine a hypothetical latent word series that is a latent word series corresponding to the hypothesis. The simultaneous probability of the hypothesis and hypothesis latent word series is obtained. The obtained joint probability is input to the score recalculation unit 24 in combination with the latent word sequence.

  Specifically, the latent word sequence determination unit 22 can use a Viterbi algorithm. The Viterbi algorithm is an algorithm that can determine a latent word sequence hidden behind a certain hypothesis and calculate a joint probability of a certain hypothesis and latent word sequence in parallel. For details on the Viterbi algorithm, see Reference 1 above. If the Viterbi algorithm is used, the latent word sequence and the joint probability can be obtained. Here, the joint probability is called the Viterbi probability. For example, when the input hypothesis is “mandarin orange”, the determined latent word sequence is “apple”, and the Viterbi probability of the input hypothesis and latent word sequence can be uniquely determined as 0.0005. Viterbi probability is logarithmized and output. The reason is that the original score is also based on the logarithm of the probability, so that the meaning of the score is the same. If the Viterbi probability is 0.0005, the log probability value is −3.301. As shown in FIG. 4, the output for the hypothesis “Today is sunny: −15000” is, for example, “Today is the weather: -8”. Here, the character string in front of “: (colon)” is a latent word sequence, and the numerical value behind is a joint probability.

  In step S24, the score recalculation unit 24 calculates a speech recognition score for each of a plurality of hypotheses using the temporary speech recognition score input from the multiple hypothesis generation unit 20 and the joint probability input from the latent word sequence determination unit 22. Ask. The speech recognition score is paired with the hypothesis and input to the first hypothesis determination unit 26.

The voice recognition score can be obtained by adding the provisional voice recognition score and the joint probability that is the Viterbi logarithmic probability with weights. That is, the speech recognition score is obtained by calculating the following equation.
(Speech recognition score) = (temporary speech recognition score) + (weight) × (simultaneous probability)

  At this time, the weight is manually adjusted. For example, voice data for weight adjustment may be prepared and adjusted to a value that improves voice recognition performance. For example, 5000 is set. As shown in FIG. 4, if the joint probability is “−8” for the hypothesis “Today is sunny: −15000”, the speech recognition score is −15000 + 5000 * (− 8) = − 55000. Become.

  In step S <b> 26, the first hypothesis determination unit 26 determines a speech recognition result for the input speech based on the speech recognition score from a plurality of hypotheses input from the score recalculation unit 24. Specifically, the speech recognition scores are sorted in descending order, and the hypothesis having the largest speech recognition score is determined as the first hypothesis. The first hypothesis determined in this way is output as a speech recognition result. In the example of Fig. 4, the speech recognition score for "Today is sunny" is -55000, the speech recognition score for "Today is sunny" is -40500, and the speech recognition score for "Today is sunny" is -85000. Therefore, the hypothesis “Today is sunny” is output as a speech recognition result.

  With this configuration, the speech recognition apparatus 1 according to the embodiment can perform speech recognition using a latent language model at high speed. That is, since speech recognition is performed using an N-gram language model that has excellent language prediction performance of a latent language model, it is more effective than speech recognition using an N-gram language model constructed by a conventional general learning method. High language prediction performance can be obtained. In addition, when searching for a latent word language model after narrowing candidates for speech recognition results using an N-gram language model of a latent word sequence generated based on the latent word language model, a Viterbi search is simply performed. Compared with, the amount of calculation can be greatly reduced. This makes it possible to realize speech recognition using a latent language model even with a normal computer.

[Program, recording medium]
The present invention is not limited to the above-described embodiment, and it goes without saying that modifications can be made as appropriate without departing from the spirit of the present invention. The various processes described in the above-described embodiments are not only executed in time series according to the order described, but may be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes.

  When various processing functions in each device described in the above embodiment are realized by a computer, the processing contents of the functions that each device should have are described by a program. Then, by executing this program on a computer, various processing functions in each of the above devices are realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, the present apparatus is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

1 speech recognition device 10 learning text storage unit 12 latent language model learning unit 14 latent language model storage unit 16 baseline language model learning unit 18 baseline language model storage unit 20 multiple hypothesis generation unit 22 latent word sequence determination unit 24 score Recalculation unit 26 First hypothesis generation unit

Claims (6)

A latent language that has learned the probability distribution of the learning latent word sequence that is a latent word sequence corresponding to the observed word sequence included in the learning text, and the probability distribution of the observed word in the observed word sequence and the latent word in the latent word sequence A latent language model storage unit for storing a model;
A baseline language model storage unit for storing a baseline language model in which the probability distribution of the latent word sequence included in the latent word language model and the probability distribution of the observed word sequence are mixed;
A plurality of hypothesis generators for recognizing input speech using the baseline language model and generating hypotheses of a plurality of speech recognition results and provisional speech recognition scores for each hypothesis;
A latent word sequence determining unit that determines a hypothetical latent word sequence that is a latent word sequence corresponding to the hypothesis using the latent word language model, and obtains a joint probability of the hypothesis and the hypothetical latent word sequence;
A score recalculation unit for obtaining a speech recognition score using the provisional speech recognition score and the joint probability;
A first hypothesis determination unit that determines a speech recognition result for the input speech based on the speech recognition score from the plurality of hypotheses;
A speech recognition device. The speech recognition device according to claim 1,
The latent word language model storage unit stores the latent word language model obtained by learning the N-gram probabilities of the plurality of learning latent word sequences corresponding to the observed word sequences and the unigram probabilities of the observed words and the latent words. A voice recognition device. The speech recognition device according to claim 1 or 2,
The latent word sequence determination unit is configured to perform the determination of the hypothetical latent word sequence and the calculation of the joint probability of the hypothesis and the hypothetical latent word sequence in parallel. The speech recognition apparatus according to any one of claims 1 to 3,
The said score recalculation part calculates | requires the said speech recognition score by weighting and adding the logarithm value of the said joint probability with the predetermined weight to the said temporary speech recognition score. The latent language model includes a probability distribution of a learning latent word sequence corresponding to an observed word sequence included in a learning text, and a probability distribution of observed words in the observed word sequence and latent words in the latent word sequence. Has learned
The baseline language model is a mixture of the probability distribution of the latent word sequence included in the latent word language model and the probability distribution of the observed word sequence.
A plurality of hypothesis generation units that recognize the input speech using the baseline language model, and generate a plurality of hypothesis generation hypotheses and a provisional speech recognition score for each hypothesis;
A latent word sequence determination unit determines a hypothetical latent word sequence that is a latent word sequence corresponding to the hypothesis using the latent word language model, and determines a simultaneous probability of the hypothesis and the hypothetical latent word sequence Steps,
A score recalculation unit that calculates a speech recognition score using the provisional speech recognition score and the joint probability, and a score recalculation step;
A first hypothesis determination unit that determines a speech recognition result for the input speech based on the speech recognition score from the plurality of hypotheses;
A speech recognition method including:   A program for causing a computer to function as the voice recognition device according to claim 1.

Images