JPS6173199A - Voice preselection system for large-vocabulary word - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a pre-selection method for words in large word speech recognition.

Regarding the preliminary word phonetic selection method, which clusters a set of words into several to several dozen clusters in advance based on the distances between the entries in the rice and blood word dictionary, and matches them with the central word of each cluster. , for example, "Japanese π Society Lecture Proceedings 1-1
-2. PP3～・↓.

It is described in “Study of preliminary word selection method for speech recognition of big words” by Arata Matsui, Shozo Sakano, and Kenji Kido.
In the above method, the amount of calculation is 1 compared to the case of direct recognition.
The reduction is only 5 to 23%, and the recognition rate is 86 to 95% for 212 words, which is not a sufficient value.

The present invention was made in view of the above-mentioned circumstances.
Similar words are roughly classified into groups (called clusters) based on the distance between words during DP matching, the central word of each cluster is registered, and DP matching is performed between unknown input words and each of these central words. After determining the cluster to which the input word belongs, DP is applied only to the words within the cluster.
This is an attempt to speed up the recognition process by performing matching again.

C. In order to achieve the above object, the present invention performs DP matching between dictionary words, calculates the distance between each word, and stores a table of calculation results in a memory in a speech recognition device that performs speech recognition of large words. and a word clustering means for grouping similar words in the dictionary words based on this inter-word distance table, registering the central word in each cluster, and is input, DP matching is performed to match the central word of the cluster registered in advance, and after preliminary selection of the cluster in which the recognized word exists, DP matching is performed again only for the words within the cluster. It is characterized by the following. Hereinafter, the present invention will be explained based on examples.

Figure 1 is a system configuration diagram when registering dictionary words.
In the figure, 1 indicates N dictionary word sounds for registration n1=1.2.・
..., N input terminal section, 2 is a spectrum analysis section, 3 is a DP matching section between dictionary words ni, 4 is an inter-word distance table creation section, 5 is a word clustering section, 6 is a cluster center word registration section First, N registration word sounds are input from the input unit 1, frequency analysis is performed in the spectrum analysis 81;2, and the registration word N is input in the DP matching unit 3.
DP matching between individuals is performed. In reality, the distance between words and itself is O, and due to the symmetry of combinations, the total number of unique combinations (N' - N) is 72 D
P matching can be performed. Therefore, the registered words ni and n
An inter-word distance table D (ni, nj) between j (ni, nj-1, 2, . . . , N) is created by an inter-word distance table creation unit 4.

FIG. 2 is a diagram showing an example of an inter-word distance table created as described above, and the table is a symmetric matrix with O diagonal elements. That is, D= (ni, nj) =O: ni =njD=
(ni, nj) = D (nj, ni) ≠O:n
i≠njni, nj =1. 2. -, N.

Next, inter-word distance table 4 created as described above.
Based on this, the clustering unit 5 performs word clustering to group similar words, and at this time,
A central word for each cluster is registered in the word registration section 6.

FIG. 3 is a conceptual diagram of word clustering. In the figure, 7 is a set of all human word sounds, 7a each represents one of the word clusters, 7b is the central word of cluster 7a, and 7C is Represents one of the dictionary words belonging to cluster 7a. Note that details of the clustering algorithm will be described later.

FIG. 4 is a system configuration diagram when recognizing unknown input voice DP matching unit with the center word of each cluster, 10 is a cluster selection unit, 11 is D with the dictionary word in the selected cluster.
P matching unit, 12 is a word identification unit, and 13 is a recognition result output unit.The unknown human voice 3) DP matching with Figure 7b) is DP
This is performed by the matching unit 9, and the cluster to which the central word with the shortest distance belongs is determined by the cluster selection unit IO (in FIG. 3, 7a is the determined cluster). Next, the DP matching unit 11 performs DP matching with the unknown input speech X for the words belonging to the cluster determined by the cluster selection unit 10, and the word selection unit 12 determines the word with the minimum distance, and outputs the recognition result. The unit 13 outputs the result as a recognition result.

FIG. 5 is a general flowchart showing an example of a word clustering algorithm. In the figure, 14 is an initialization block, 15 is a word clustering section, 16 is a calculation section for the average value of distance between words, and 17 is a clustering block. 18 is a central word registration section; 19 is a recalculation section for the central word within each cluster; 20 is an inter-word distance updating section and a counter section for the number of repetitions m.

FIG. 6 is a flowchart showing another example of the word clustering algorithm, in which 1 is selected as the most important word in the large word group in the initialization block 21, the clustering level is countered up in 22, and the clustering level is countered in 23. In step 24, select the word π2 with the greatest distance from the beach 1, and in step 24, calculate the distance from ni (i=1.2...,k) using Table 4 shown in Figure 2. Find the word with the maximum distance among the minimum distances found in step 25 and n, +.

Then, 26τL ya 1 and lower 1 (i=1.2. . .
, k) is smaller than a certain threshold V2D (Ei, ne ), the central word nk (k
=1゜2,...,K), otherwise 2
At 8, raise the clustering level k by one and return it to 24.

Note that in the above embodiment, there was no overlap between clusters when clustering words;
As shown in Figure (a), clusters may overlap. Furthermore, by forming the clusters into a tree structure as shown in FIG. 2(b), it is possible to reduce the amount of calculation required for determining candidate clusters. In addition, the darkness value is set based on the distance between the words from the cluster center of the first candidate,
It is also possible to select clusters within a certain dark value other than the first candidates to improve the preliminary selection rate. In addition, apart from the method of setting the darkness value, from the 11th complement to the k (ffl complement (k<
It is also possible to select clusters up to <K). Also,
By selecting the clusters up to the first candidate among the clusters within a certain dark value, it is also possible to improve the performance of the preliminary selection rate.

As is clear from the above explanation, according to the present invention, major words are roughly classified into word clusters based on the distance between words during DP matching, and the words at the center of each cluster are registered. When an unknown input voice is input, a candidate cluster is determined by DP matching with the foreign language at the center of the cluster, and then DP matching is performed only with words within this cluster to recognize it. It becomes possible to recognize word sounds quickly and accurately.

(Figure 2, Figure 3, Figure 5, Figure 7, Figure 7, Figure 7, Figure 7, Figure 7, Figure 7, Cb)) Procedural amendments imported November 22, 1981 Manabu Shiga, Commissioner of the Japan Patent Office 2, Name of invention large search word audio preliminary selection method 3 , Relationship with the case of the person making the amendment Patent applicant Ota Ri Nakamagome Address 1-3-6 Nakamagome, Ota-ku, Tokyo Name (674) Shares Daisha Rico - Representative
Hama 1) Hiro 4, Agent Address 6, 807-807, 1-2-7, Furo-cho, Naka-ku, Hama-shi, ItItItItItItItItI-Hama-shi, Naka-ku, Hama-shi, 231 Japan, Subject of amendment 7, Contents of amendment (1), Specification page 2 “Matsui Arata” written in the 9th line.

"Shozo Sakano," is corrected to "Arata Murai, Shozo Makino,".

(2) "In the word selection section 12" written in lines 10 to 11 on page 6 is corrected to "in the word identification section 12."

(3), "Table 4" written on page 7, line 8 of the same page is "
Correct the table.

(4), “threshold 1/2D (n i, n e) J” described in page 7, line 11 of the same (5),
Figures 5 and 6 are corrected as shown in the attached sheet.

Claims (1)

[Claims] In a speech recognition device that performs speech recognition of large vocabulary words, there is a means for performing DP matching between dictionary words to calculate the distance between each word and storing a table of calculation results in a memory, and a means for storing a table of calculation results in a memory, and It has a word clustering means for grouping similar words in the dictionary words, registers the central word in each cluster, and when an unknown input word is input, the word clustering means groups similar words in the dictionary words. Preliminary speech selection of large vocabulary words characterized by performing matching with the center word of a cluster by DP matching, preliminary selecting a cluster in which a recognized word exists, and then performing DP matching anew only for the words in the cluster. method.