JP2000075885A - Voice recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice recognition device in which voice recognition is conducted with a high recognition rate and a high speed employing a highly precisely approximate language likelihood. SOLUTION: A word dictionary initialization processing section 10 generates a tree structure word dictionary based on memory learning text data, computes a look-ahead probability, which is an approximate language likelihood, and adds the probability to each node of the tree structure. A word collating section 6 computes the look-ahead probability, which is the approximate language likelihood given to the non-terminating state of the words in a tree structure word dictionary of a memory 22, for every word hypothesis inputted from a phoneme collating section 4 based on the probability data of the statistical language model N-gram in a memory 23, updates the dictionary in the memory 22 and voice recognizes the inputted voice signals employing the updated tree structure word dictionary.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus for performing speech recognition using a tree-structured word dictionary.

[0002]

2. Description of the Related Art In recent years, in a continuous speech recognition apparatus, a method using a statistical language model has been studied in order to improve the performance thereof. This aims at improving the recognition rate and reducing the calculation time by predicting the next word and reducing the search space using a statistical language model.
As a statistical language model that has been widely used recently, N-
gram (N-gram; N is a natural number of 2 or more). It learns large-scale text data and statistically gives the transition probability from the previous N-1 words to the next word. Multiple L word strings w ₁ ^L
= W ₁ , w ₂ ,..., W _L The generation probability P (w ₁ ^L ) is represented by the following equation.

[0003]

(Equation 1)

[0004] Here, w _t represents a t-th one word of the word string w ₁ ^L, w _i ^j represents the j-th word string from the i-th. In the above _equation 1, the probability P (w _t |
wt _{+ 1-} ^Nt-1 ) is a word sequence wt _{+ 1-} ^Nt-1 composed of N words.
Is the probability that the word w _t will be uttered after is uttered, and similarly, the probability P (A | B) means the probability that the word A will be uttered after the word or word string B has been uttered. Also,
“Π” in Equation 1 represents the probability P (w _t from t = 1 to L
| W _{t + 1−N} ^t−1 ), and so on.

In recent years, the statistical language model N
Techniques for improving the performance of continuous speech recognition using -gram have been actively proposed.
"LR Bahl et al.," A Maxim
um LikelihoodApproach to
Continuous Speech Recogni
Tion ", IEEE Transactions o
n PatternAnalysis and Mac
hine Intelligence, pp. 179-
190, 1983 "and prior art document 2" Shimizu et al.,
"Free speech recognition using word graphs", IEICE Technical Report, SP95-88, pp. 146-64. 49-54,
1995 ". ).

[0006] However, N-gram has a large number of parameters, and an enormous amount of text data is required to accurately obtain each value. As a method for solving this problem, a smoothing method for giving a transition probability even to a word transition that does not appear in the learning text data (for example, see the related art document 3 “F. Jelinek et a
l. , “Interpolated estimati
on of Markov Source Param
eters from Sparse Data ”, P
rosedings of Workshop Pa
ttern Recognition in Prac
tice, pp. 381-387, 1980 "and prior art document 4" SM Katz et al., "Es
timing of Probabilities
from Sparse Data for the
Language model Component
of a Speech Recognizer ",
IEEE Transactions on Acou
stics, Speech, and SignalPr
ossing, pp. 400-401, 1987
Year "and Prior Art Document 5" Kawabata et al., "Back-of N-gram Statistical Language Model Based on Binomial Posterior Distribution"
f Smoothing ", IEICE Technical Report, SP95-9
3, pp1-6, 1995. " ), Classification,
Techniques for reducing the number of parameters such as variable-length N-grams (for example, see Prior Art Document 6 “PF Brownet
al. , “Class-Based n-gram
models of natural language
e ", Computational Linguist
ics, Vol. 18, No. 4, pp. 467-47
9, 1992 "and prior art document 7" TR Nies "
ler et al. , "A Variable-Le
Nth Category-Based N-gra
m Language Model ”, Proceed
ins of ICASSP '96, Vol. 1,
pp. 164-167, 1996 "and prior art document 8" Masataki et al., "Variable-length chain statistical statistical language model for continuous speech recognition", IEICE technical report, SP
95-73, p. 1-6, 1995 ". ) Have been proposed. However, even with these methods, it is considered necessary to use a considerable amount of data in order to construct an accurate statistical language model.

[0007] In order to solve the above-mentioned problems, the related art 9 [Volker Steinbiss et al.
l. , “Improvements in beam
search ”, ICLSP 94, Yokoha
ma, Japan, pp. 2143-2146 "and prior art document 10" Stephan Ortmanns "
et al. , "A word graph arg
orient for large vocabulary
y continuous speechrecogn
ition ”, Computer Speech &
Langage, 1997, 11, pp. 4
3-72 "discloses a speech recognition method using a tree-structured word dictionary (hereinafter, referred to as a conventional example). In this conventional example, the non-terminal state (non-terminal node) of the tree structure dictionary
As the approximate language likelihood for, the largest of the unigram probabilities of all the words to which the node belongs is used. Here, the unigram probability of a word refers to the appearance probability of one word.

The processing of the unigram prefetching method based on the statistical language model used in this conventional example will be described. P _lookahead of each node in the tree structure word dictionary
Is set as follows. (1) For each leaf node in the tree-structured word dictionary, as shown in the following equation, the maximum value of all unigram (word set displayed as W _leafnode ) probabilities P (w) of words ending at this leaf node Is calculated and set to the look-ahead probability p _lookahead (leafnode) at each leaf node. Due to homonyms and multiple pronunciations, a leaf node may end with several words.

[0009]

## EQU2 ## p _lookahead (leafnode) = MAX {P (w)} where wｗW _leafnode

(2) For the look-ahead probabilities p _lookahead in all non-leaf nodes, the look-ahead probabilities p for all child nodes branching from the look-ahead to leaf nodes
Set the maximum value of _lookahead (child-node).

[0011]

## EQU3 ## p _lookahead (non-leafnode) = MAX {p
_lookahead (child-node)}

Note that the prior art unigram look-ahead method does not depend on the currently expanded word hypothesis, and is therefore a statistical procedure, which should normally be calculated only once in advance. . Here, an embodiment according to a conventional method will be described below. The unig used in this example
An example of a statistical language model of ram is shown in the following table. FIG. 4 shows a tree-structured word dictionary obtained by the above processing.

[0013]

[Table 1]

[0014]

However, since the approximate language likelihood used in the conventional example is based on the unigram probability of a word, the approximation accuracy is low, and the effect of reducing the calculation time required for recognition is not sufficient. Therefore, there is a problem that the calculation cost is high and the capacity of the memory for storing the tree structure word dictionary is relatively large.

SUMMARY OF THE INVENTION An object of the present invention is to solve the above-mentioned problems, and to realize a speech recognition apparatus capable of recognizing speech at a higher recognition rate and at a higher speed by using a language likelihood approximated with higher precision as compared with the conventional example. Is to provide.

[0016]

According to a first aspect of the present invention, there is provided a speech recognition apparatus which generates a tree-structured word dictionary based on learning text data and generates an approximate language likelihood for each node of the tree structure. A speech recognition device provided with speech recognition means for calculating and adding a look-ahead probability as a degree and recognizing a speech signal input using the tree-structured word dictionary, wherein N is a natural number of 2 or more. Storage means for storing a statistical language model including N-gram probability data, wherein the speech recognition means provides, for each generated word hypothesis, an approximate language likelihood given to a non-terminal state of a word in the tree-structured word dictionary Is updated based on the N-gram probability data of the statistical language model stored in the storage means, thereby updating the tree-structured word dictionary. Using a word dictionary, and recognizes speech inputted audio signal.

According to a second aspect of the present invention, in the voice recognition apparatus of the first aspect, the voice recognition unit generates a tree structure word dictionary based on learning text data; First assigning means for calculating and assigning, as a look-ahead probability, the maximum probability of all unigrams of words ending at a leaf node to each leaf node in the tree-structured word dictionary; A second assigning means for storing the tree-structured word dictionary in another storage means by setting and giving the prefetch probability to a node which is not a node and the maximum probability of all child nodes branching to a leaf node; And for each generated word hypothesis, a statistical language model stored in the storage means except for a word unigram for each set of word hypotheses. All of the N-gr present in
a third adding means for calculating the look-ahead probabilities of each leaf node by expanding to the maximum N-gram probability of the input data of am and adding the calculated prefetch probability to the tree-structured word dictionary stored in the another storage means; By setting and giving the prefetch probability to all the nodes that are not leaf nodes in the structured word dictionary and the maximum probabilities of all child nodes branching to the leaf node, Using a fourth assigning means for updating the tree structure word dictionary, the updated tree structure word dictionary, and the statistical language model stored in the storage means, the maximum likelihood of the input speech signal is obtained. Search recognition means for searching for and determining a word hypothesis and outputting the result as a recognition result.

[0018]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram of a continuous speech recognition apparatus according to one embodiment of the present invention. In the continuous speech recognition device of this embodiment, the word dictionary initialization processing unit 10
Generates a tree-structured word dictionary of a binary tree format based on the text data for memory learning, and calculates and adds a look-ahead probability, which is an approximate linguistic likelihood, to each node of the tree-structured word. A statistical language model memory 23 that is stored in the memories 21 and 22 and stores a statistical language model including N-gram probability data of a word in which N is a natural number of 2 or more.
Is provided. Here, the word collating unit 6 includes the phoneme collating unit 4
For each word hypothesis input from, the look-ahead probability, which is the approximate language likelihood given to the non-terminal state of the word in the tree-structured word dictionary in the memory 22, is stored in the N-gram probability data of the statistical language model in the memory 23. Based on the calculation, the tree-structured word dictionary in the memory 22 is updated, and the input speech signal is speech-recognized using the updated tree-structured word dictionary.

The process of searching for the most likely hypothesis in the speech recognition apparatus is performed based on a word dictionary that includes all recognizable words. In a conventional speech recognition apparatus, a tree structure word dictionary (recognizable words are displayed in a memory as a tree structure instead of a simple linear list) is usually used. When a tree-structured word dictionary is used, a procedure called a pre-reading method of a statistical language model is used in order to incorporate the statistical language model probability as soon as possible in a search using the tree-structured word dictionary. One of the frequently used procedures
One is a unigram prefetching method, which has been described in the conventional example. On the other hand, in the present embodiment, an on-demand N-gram look-ahead method that can extend the conventional method and improve the search speed by about 20% is used.

First, a method of prefetching a statistical language model will be described. The look-ahead method based on the statistical language model is
It is used in many speech recognition devices that use a tree-structured word dictionary. When you enter the tree-structured word dictionary during the search,
The identity of a word is unknown until it reaches a leaf node (the node where the word ends, which means the end state of the word).
Therefore, the exact language model probability in the tree structure word dictionary is also unknown. In order to obtain good high-speed search performance, it is necessary to incorporate the language model probability as soon as possible while passing through the tree-structured word dictionary. In many speech recognition systems that use a tree-structured dictionary, a procedure called a statistical language model look-ahead method is used to incorporate language model probability estimates into the tree-structured dictionary. The look-ahead probability (p _lookahead ) of the statistical language model belongs to every node of the tree-structured dictionary. Assuming they have already been set, it is used during the search as follows.

(A) When entering a node, add p _lookahead (node) to the current overall score. (B) After leaving the node, p
Subtract _lookahead (node).

This method promotes pruning of nodes with weak language model probabilities faster than if no statistical language model look-ahead method is used, resulting in an increase in search speed. P of each node in the dictionary
A method for setting _lookahead will be described.

The word dictionary initialization processing unit 1 according to this embodiment
0 is a binary tree-type tree-structured word based on text data (corpus) of a plurality of uttered speech sentences stored in the learning text memory 2 and unigram probability data of the word in the statistical language model memory 23. A dictionary is generated, and a look-ahead probability p _lo is calculated for each node using a conventional method.
_By calculating and adding _okahead , an initial value tree structure word dictionary is generated and stored in the memory 21, and then copied to the memory 22. By the following processing, each time a word hypothesis is input from the phoneme matching unit 4 to the word matching unit 6 via the buffer memory 5 (on-demand), the tree-structured word dictionary in the memory 22 is updated. The maximum likelihood word hypothesis is searched and determined using the tree-structured word dictionary in the memory 22 and the statistical language model in the memory 23, and is output as a recognition result.

On-demand N-gr according to this embodiment
The am look-ahead method is a look-ahead procedure for a new statistical language model, which incorporates the hypothesis constraints developed at that processing point. As a result, the estimated value of the actual language model probability is improved as compared with the procedure of the conventional unigram look-ahead method, which leads to an improvement in the pruning accuracy and hence a high-speed search. The speed increase is about 20%.

Next, the processing of the on-demand N-gram prefetching method is as follows. (1) Prior to the search, the prefetch probability p of all nodes is determined by the procedure of the above-mentioned unigram prefetch method.
Initialize _lookahead . (2) The hypothesis H _i is calculated for each word set, and in the statistical language model, all N-gram data (H) existing in the statistical language model except for the unigram already set during the initialization of the unigram. _i ,
w) to the maximum N-gram probability P (w | H _i ). Identify related leaf nodes belonging to word w (there may be several due to homonyms and multiple pronunciations), and as the look-ahead probability p _lookahead , the calculated probabilities and The maximum value of the pre-reading probabilities p _lookahead of the set unigram is set.

[0027]

Equation 4] _{p lookahead (leafnode) = MAX {} P (w | H i)} ∀H i and ∀W∈ {present in _{N-gram (H i, w} )}

(3) For all non-leaf nodes (that is, nodes that are not leaf nodes and refer to non-terminal states of words), their look-ahead probabilities p _lookahead
Is set to the maximum value of the look-ahead probabilities p _{lookah ead} (child-node) of all the child nodes branching to the leaf node as in the following equation.

[0029]

## EQU5 ## p _lookahead (non-leafnode) = MAX {p
_lookahead (child-node)}

This procedure must be executed every time a new word hypothesis set that cannot be executed in advance is developed as in the case of the normal unigram look-ahead method. Regardless of this additional procedure, the more accurate the language model probability, the more accurate the pruning, which leads to a faster full search.

Next, an example of an on-demand N-gram prefetching method will be described. Note that all N-gram probabilities are used, depending on the hypothesis list to be expanded. An example of a statistical language model is shown in the following table, and an example of a hypothesis list to be expanded stored in the memory 7 is shown in the following table. FIG. 5 shows a tree structure word dictionary created by using these. The words w ₁ , w ₂ , w ₃ , w ₄ ,... In the following table are words expressed by phoneme strings, for example.

[0032]

[Table 2]

[0033]

[Table 3]

The hypothesis list memory 7 to be developed temporarily stores the history of the word hypotheses generated by the processing of the word collating unit 6. As shown in FIG. 5, in the tree-structured word dictionary, a tree grows in a binary tree form from the root node RN to the leaf node LN, a look-ahead probability p _lookahead is given to each node, and the word matching unit 6 Each time a word hypothesis is input by the processing according to the above, the look-ahead probabilities p _{lookah ead} of the _{assigned nodes} are updated and word matching is performed. Here, the direction from the root node RN toward the leaf node LN is the direction toward the child node.

FIG. 2 shows the word dictionary initialization processing unit 10 shown in FIG.
6 is a flowchart showing a word dictionary initialization process executed by the CPU. The statistical language model memory 23 stores N or more words of “trigram” or more based on learning text data which is a corpus including a plurality of uttered voice sentences.
-Gram connection probability data is stored in advance.

In FIG. 2, in step S1, a tree-structured word dictionary in a binary tree format is generated based on the learning text data in the memory 20 and the probability data of the unigram of the statistical language model in the memory 23. I do.
Next, in step S2, for each leaf node LN in the tree structure, the maximum probabilities of all unigrams of words ending at the leaf node LN are determined by a look-ahead probability p.
Calculated and given as _lookahead (leafnode). Further, in step S3, the look-ahead probabilities p _lookahead (non-l
eafnode) with the maximum probability of all child nodes branching to the leaf node. Finally, the tree-structured word dictionary with probability generated in step S4 is stored in the memory 21 and copied and stored in the memory 22, and the word dictionary initialization process ends.

FIG. 3 is a flowchart showing a word matching process executed by the word matching unit 6 of FIG. FIG.
First, in step S11, it is determined whether or not a word hypothesis has been input, and the process waits until the word hypothesis is input. Each time the word hypothesis is input, the processes of the following steps S12 to S14 are executed. Next, in step S12, the word hypothesis H _i
, The maximum N-gram probability p (w) of all N-gram input data (H _i , w) present in the statistical language model that does not include the word unigram
│H _i ) to expand the look-ahead probability p of each leaf node
Calculate and add _lookahead (leafnode). further,
In step S13, the look-ahead probability p _lookahead (non-leafnod
The maximum probability of all the child nodes branching to the leaf node is set and assigned to e), and the tree structure word dictionary in the memory 22 is updated. Finally, in step S14, the updated tree-structured word dictionary in the memory 22 and the memory 23
The maximum likelihood word hypothesis is searched for and determined using the statistical language model in, and is output as a recognition result.

Next, the configuration and operation of the continuous speech recognition apparatus shown in FIG. 1 will be described. In FIG. 1, a phoneme HMM in a phoneme hidden Markov model (hereinafter referred to as HMM) memory 11 connected to the phoneme matching unit 4 is represented by including each state, and each state includes the following information. Having. (A) state number, (b) acceptable context class, (c) list of preceding and succeeding states, (d) parameters of output probability density distribution, and (e) self-transition probability and transition to succeeding state probability. Note that the phoneme HMM used in the present embodiment is generated by converting a predetermined speaker-mixed HMM because it is necessary to specify which speaker each distribution originates from. Here, the output probability density function is a Gaussian mixture distribution having a 34-dimensional diagonal covariance matrix.

In FIG. 1, a speaker's uttered voice is input to a microphone 1 and converted into a voice signal, and then input to a feature extracting unit 2. After performing A / D conversion on the input audio signal, the feature extraction unit 2 performs, for example, LPC analysis, and performs 34-dimensional feature parameters including logarithmic power, 16th-order cepstrum coefficient, Δlogarithmic power, and 16th-order Δcepstrum coefficient. Is extracted. The time series of the extracted feature parameters is input to the phoneme matching unit 4 via the buffer memory 3.
The phoneme matching unit 4 detects the word hypothesis of the phoneme string using the phoneme HMM 11 based on the feature parameter data input via the buffer memory 3 using the one-pass Viterbi decoding method, and detects the likelihood. Is calculated and output to the word collating unit 6 via the buffer memory 5. The word matching unit 6 is configured as shown in FIG.
Is executed, the tree structure word dictionary in the memory 22 is updated, and the maximum likelihood word hypothesis is searched for by referring to the statistical language model in the memory 23 and the hypothesis list to be expanded in the memory 7. Determined and output as recognition result.

In the above embodiment, the feature extracting unit 2
The phoneme matching unit 4, the word matching unit 6, and the word dictionary initialization processing unit 10 are composed of, for example, a computer such as a digital computer, and include buffer memories 3 and 5, and a hypothesis list memory 7 to be developed. And the phoneme HMM memory 11
The learning text data memory 20, the tree structure word dictionary memories 21 and 22, and the statistical language model memory 23 are configured by a storage device such as a hard disk memory. The statistical language model in the memory 23 is preferably an N-gram statistical language model in which N is a natural number of 2 or more, and more preferably trigram.
m is a statistical language model.

According to the embodiment of the present invention, by using the above-described on-demand N-gram look-ahead method, it is possible to calculate the approximate value of the language likelihood with a smaller storage area and higher accuracy than in the conventional example. As a result, speech recognition can be performed with a higher recognition rate than the conventional example, and the calculation time required for recognition can be significantly reduced.

In the above embodiment, the word dictionary initialization processing of FIG.
May be configured to be executed on demand each time the data is input to the word collating unit 6 via the buffer memory 5.

In the above embodiment, the tree-structured word dictionary of the binary tree format is generated. However, the present invention is not limited to this, and the tree-structured word dictionary of the plural N-tree format may be used.

[0044]

As described in detail above, according to the speech recognition apparatus of the first aspect of the present invention, a tree-structured word dictionary is generated based on the learning text data, and the tree-structured word dictionary is generated for each node of the tree structure. A speech recognition device provided with speech recognition means for recognizing a speech signal input using the tree-structured word dictionary by calculating and adding a look-ahead probability which is an approximate language likelihood. Storage means for storing a statistical language model including N-gram probability data of a word, wherein the speech recognition means gives, for each generated word hypothesis, a non-terminal state of the word in the tree-structured word dictionary The tree-structured word dictionary is updated by calculating a look-ahead probability, which is an approximate language likelihood, based on N-gram probability data of a statistical language model stored in the storage means. Using a tree word dictionary, the speech recognizing an input speech signal. Therefore, it is possible to calculate the approximate value of the linguistic likelihood with high accuracy in a smaller storage area than in the conventional example, to perform speech recognition with a higher recognition rate than in the conventional example, and to further reduce the calculation time required for recognition. Can be greatly reduced.

According to a second aspect of the present invention, in the first aspect, the speech recognition unit may include a generation unit that generates a tree-structured word dictionary based on the learning text data. A first assigning means for calculating and assigning, as a look-ahead probability, a maximum probability of all unigrams of words ending at a leaf node to each leaf node in the tree-structured word dictionary; Storing the tree-structured word dictionary in another storage means by setting and giving the prefetch probability to the node that is not a leaf node the maximum probability of all child nodes branching to the leaf node. Adding means and, for each word hypothesis to be generated, a statistical word stored in the storage means except for a word unigram for each set of word hypotheses All of the N present in the model
A third adding means for calculating the look-ahead probability of each leaf node by expanding to the maximum N-gram probability of the input data of -gram and adding it to the tree-structured word dictionary stored in the another storage means; By setting and giving the look-ahead probabilities to all nodes that are not leaf nodes in the tree-structured word dictionary and the maximum probabilities of all child nodes branching to the leaf nodes, the data are stored in the other storage means. A fourth assigning means for updating the tree-structured word dictionary,
Search recognition for searching and determining the maximum likelihood word hypothesis for the input speech signal using the updated tree-structured word dictionary and the statistical language model stored in the storage means, and outputting the result as a recognition result Means. Therefore, it is possible to calculate the approximate value of the linguistic likelihood with high accuracy in a smaller storage area than in the conventional example, to perform speech recognition with a higher recognition rate than in the conventional example, and to further reduce the calculation time required for recognition. Can be greatly reduced.

[Brief description of the drawings]

FIG. 1 is a block diagram of a continuous speech recognition apparatus according to an embodiment of the present invention.

FIG. 2 is a flowchart showing a word dictionary initialization process executed by a word dictionary initialization processing unit 10 of FIG.

FIG. 3 is a flowchart showing a word matching process performed by the word matching unit 6 of FIG. 1;

FIG. 4 is a structural diagram showing an example of a tree structure configuration of a conventional tree structure word dictionary.

FIG. 5 is a structural diagram illustrating an example of a tree structure configuration of a tree structure word dictionary according to the embodiment;

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Microphone, 2 ... Feature extraction part, 3, 5 ... Buffer memory, 4 ... Word collation part, 6 ... Word collation part, 7 ... Hypothesis list memory to be expanded, 10 ... Word dictionary initialization processing part, 11 ... Phoneme HMM memory, 21, 22 ... tree structure word dictionary memory, 23 ... statistical language model memory.

Continuation of the front page (72) Inventor Atsushi Nakamura 5 Sanraya, Seiya-cho, Soraku-cho, Kyoto, Japan F-term (reference) 5D015 BB01 GG01 GG05 HH11

Claims (2)

[Claims] 1. A tree structure word dictionary is generated based on learning text data, and a look-ahead probability, which is an approximate language likelihood, is calculated and assigned to each node of the tree structure. A speech recognition apparatus comprising speech recognition means for recognizing a speech signal input using a storage means for storing a statistical language model including N-gram probability data of a word in which N is a natural number of 2 or more. The speech recognition unit includes, for each word hypothesis to be generated, a look-ahead probability, which is an approximate language likelihood given to a non-terminal state of a word in the tree structure word dictionary, of a statistical language model stored in the storage unit. By updating the tree-structured word dictionary by calculating based on N-gram probability data,
A speech recognition apparatus, characterized in that an inputted speech signal is speech-recognized using the updated tree structure word dictionary. 2. The speech recognition device according to claim 1, wherein
The speech recognition unit includes: a generation unit configured to generate a tree-structured word dictionary based on the text data for learning;
All unigrams of words ending with leaf nodes
First assigning means for calculating and assigning the maximum probability of the child node as the look-ahead probability; and for all the non-leaf nodes in the tree-structured word dictionary, A second assigning means for storing the tree-structured word dictionary in another storage means by setting and assigning a maximum probability; and for each generated word hypothesis, a word unigram for each set of word hypotheses. Excluding the maximum N-gram probabilities of all the N-gram input data present in the statistical language model stored in the storage means except for the above, the prefetch probability of each leaf node is calculated and stored in the another storage means. Third assigning means for assigning to the stored tree-structured word dictionary, and a look-ahead checker for all nodes that are not leaf nodes in the tree-structured word dictionary. A fourth assigning means for updating the tree-structured word dictionary stored in the another storage means by setting and assigning the maximum probability of all the child nodes branching to the leaf node; And a search / recognition unit that searches for and determines the maximum likelihood word hypothesis for the input speech signal using the statistical tree model stored in the storage unit and outputs the result as a recognition result. A speech recognition device comprising: