JPS6346499A - Big vocaburary word voice recognition system - Google Patents

Abstract

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

Description

TECHNICAL FIELD The present invention relates to a three-word speech recognition system, and more particularly to a pre-selection method for words in a nine-word speech recognition device, and a classification method for large-word speech.

BACKGROUND TECHNOLOGY Conventionally, as a method for recognizing large word speech, a series of consonants and vowels in input speech is recognized.

It is common practice to match the label series of consonants and vowels with the label series of large words, and select the one closest to the label series of the input speech from a word dictionary as the recognition result. How?

Segmentation and recognition of phonemes in input speech is extremely difficult, and recognition errors are likely to occur.

Therefore, since multiple phoneme recognition results are prepared and label matching is performed between the so-called "phoneme lattice" and the word GF, there is a drawback that it is difficult to identify a single recognition result. .

Purpose The present invention was made in view of the above-mentioned circumstances.
In particular, in order to recognize large word-expense word speech at high speed, we have developed a method that broadly classifies the input speech in advance based on the phonological characteristics to limit the words to be recognized, and a method that limits the recognition target words. When recognizing word sounds, in order to shorten recognition processing time and improve recognition accuracy, Daigobo words are classified in advance in conjunction with a word pre-selection method that limits candidate words from recognition targets. This was done with the purpose of providing a method for efficient dictionary searches.

In order to achieve the above-mentioned object, the present invention provides a microphone for inputting speech, a feature analysis section for obtaining a characteristic time series in speech, and a feature analysis section for obtaining a characteristic time sequence in speech, prior to recognition of large word speech. Preliminary selection section 7 that selects word candidates using the preliminary selection section 7 A preliminary selection dictionary that is checked when performing preliminary selection, a recognition processing section that recognizes the candidate words narrowed down by the preliminary selection section, and words that are referred to during recognition processing. It consists of a standard pattern storage section and a recognition result output terminal that outputs the recognition result, which is the first consonant in the input speech.

To sequentially narrow down and make preliminary selections of candidate words based on the classification or recognition results of consonants, vowels, etc. in two vowel words, a microphone for inputting speech, and obtaining a characteristic time sequence in speech. a feature analysis unit, a preliminary selection unit that selects word candidates prior to recognition of major word sounds, a dictionary for preliminary selection that is checked when performing preliminary selection, and two candidate words narrowed down by the preliminary selection unit. It consists of a recognition processing unit that recognizes word standard patterns, a word tank zero pattern storage unit for reference during recognition processing, and a recognition result output terminal that outputs the recognition results.Category names that group the classifications of word standard patterns. and fixed category names, and the input unknown speech is classified according to the two category names to limit candidate word names.

The following is a description based on two embodiments of the present invention.

FIG. 1 is an electrical block diagram for explaining one embodiment of the present invention. In the figure, 1 is a microphone for audio input;
2 is a speech feature extraction section, 3 is a word preliminary selection section, 4 is a preliminary selection dictionary storage section, 5 is a recognition processing section, and 6 is a standard pattern (dictionary) storage section.

Reference numeral 7 denotes a recognition result output unit, in which the voice input through the microphone 1 is converted into a time series of characteristic parameters specific to the voice in a feature extraction unit 2. In the preliminary selection section 3, candidate words are refined by performing matching with a preliminary selection dictionary 4 created in advance from the major word dictionary using this series of feature parameters. In the recognition processing section 5, the words narrowed down by the preliminary selection section 3 are matched with the word standard patterns in the dictionary storage section 6, and the pattern name closest to the input cover pattern is outputted as a recognition result to the recognition result output section 7. .

FIG. 2 is an electrical block diagram for explaining the preliminary selection process of word standard patterns. In the figure, 31 is a processing start terminal, 32 is a silent section detection unit in an input speech pattern, and 33 is a duration length. Detection unit.

34 is a word-initial Cv classification unit (where C represents a consonant and ■ represents a vowel), and 35 is a word-internal vCv classification unit.

The silent section detecting unit 32 detects silent sections in the input speech, and excludes patterns that are significantly different from standard patterns that have been analyzed in advance mainly based on the number and position of silences from recognition targets. Similarly, the duration detection unit 33 excludes input voices whose lengths are significantly different from the standard pattern (usually by ±30% or more). In addition, in parallel with the processing of the silent interval detection section 32 and the duration detection section 33, the word-initial CV classification section 34 analyzes the beginning CV (or V) in the input speech.
The classification and intra-word vCv classification unit 35 sequentially classifies the intra-word VCVs and narrows down the standard pattern candidates. 07 classification by the word-initial Cv classification section 34 and the word-initial VC■ minute order section 3
Vowel recognition at 767 minutes m in No. 5 is performed prior to consonant recognition, and since the recognition rate of consonants is extremely low compared to vowels, only rough classification is performed.

FIG. 3 is a diagram showing the classification of word-initial CVs.

This is the case where the vowel V is /1/. Similar classifications apply to other vowels. The classification of the initial consonant C focuses on the fact that characteristics appear in the low and high ranges of the bandpass filter group output.
When the continuous frame length Fc of a consonant is larger than a certain threshold FL in the low range, it is 7 m.

rl+ br d2g+ Zt yT Wr r/, and if it is smaller than another certain threshold FH in the high range,
/S+Pr'j+k+h/If r is larger, it is classified as /S/. In this way, words to be recognized are limited by classifying the initial consonants.

Next, FIG. 4 shows an example of classification based on word middle consonants.

Here, the word middle consonant refers to the consonant in the vCv syllable that follows the first CV syllable in the input speech. As shown in Figure 1, this classification is those that produce a silent interval before a consonant / p r
tHk + 8 + )1 /, which causes a power dip in the consonant part /b+drg+z, r/, which is a group of other consonants. The vowels on both sides of a vCV syllable are recognized before the consonant between them, and the word candidates narrowed down by the initial 07 classification are classified by vCv syllables to further narrow down the word candidates.

Figure 5 shows the state of phoneme classification when the input speech is uttered as "Chiba". In the figure, the vowel candidates in the first Cv syllable are / i / and / u /. However, it is determined as / i / by matching with the vowel weight pattern cut out from the word in advance. Also, the initial consonant is determined by the duration of the high frequency power /S+ P+ t+
Either k+h/ is determined, and in the same way, the vowels and consonants of the vCv syllable in the word are also determined. In this way, in the word preliminary selection method based on phoneme classification, candidate words can be narrowed down in parallel with the input of speech, so high-speed preliminary selection processing is possible. The recognition processing unit 5 at the final stage performs pattern matching on a word-by-word basis with reference to the standard 1 special pattern 6 for the narrowed down candidate words, and outputs the word name with the minimum distance as a recognition result.

FIG. 6 is a diagram showing how the large word gold word set is classified into a tree structure based on the consonant classification and vowel classification in FIGS. 3 and 4. In the diagram, Gla, G, b.

Glc is the consonant classification group in Figure 3, and G, a.

G2b and G2c are the consonant classification groups in Figure 4, a,
i.

u, e, and O represent five vowels. If the input voice is "Chiba" mentioned in Figure 5, Gib-)i→a2b in Figure 6,
→ Word group 4 containing rTIBAJ following route a
This indicates that 0 is to be searched from among the major words.

By classifying word cents into a tree structure in this way, word searches can be performed at high speed.

FIG. 7 is a diagram for explaining the classification of the word dictionary in mark 7 (Q pattern storage unit 6) shown in FIG. , 63a to 63e are fixed category names in the second hierarchy, 62a to 62c are groups different from the power category in the third hierarchy, 64a
~64e indicates a fixed category name in the fourth hierarchy.

Note that the following hierarchies are also alternately composed of one grouped hierarchy and fixed category hierarchies. 65□ to 65n are standard pattern names determined by tracing the roots of the hierarchically structured categories. Therefore, information on the chain of categories is recorded along with the standard pattern names (word names) W1 to Wn. for example,
As shown in Fig. 8, when unknown input voice is input,
Appropriate segmentation is performed from the beginning of the audio, and G-engineering is performed.

Fl, G2. F, and four categories are determined.

Group G is classified as Glb in the 9th U'A initial consonant classification.

1'1, and the fixed category F1 is determined to be the vowel /1/. Similarly, it is assumed that G2 belongs to G2b of the middle consonant classification in FIG. 10, and F2 is identified as the vowel / a /. Therefore, the word name 651 in FIG. 7 includes the place name ``rtiba (Chiba)'', for example. In this example, G, , G2 represent the initial and middle consonant groups, and F□I F2 represents the 5
represents a vowel. Alternatively, as another classification method, the unknown input speech shown in FIG. 8 may be categorized for each segment based on the reliability of recognition (for example, the comparison distance with the phoneme standard pattern). That is, phonemes with high reliability are set to a fixed category, and phonemes with low reliability are set to a group category. For that reason, the seventh
It goes without saying that the dictionary classification shown in the figure also needs to be structured into layers based on this reliability.

In this way, by classifying the input speech into categories in order from the beginning, it is possible to quickly limit candidate words from the major word dictionary. In addition, parts of the input audio that can be relatively reliably recognized and parts that are uncertain are classified as separate layers.
This has the effect of making it difficult for cross-rejects (exclusion of correct candidates) to occur when the candidates are limited.

Achievements As is clear from the above explanation, according to the present invention, the classification results and recognition results of phonemes are calculated from the initial part of the input speech, such as the initial consonant, vowel, middle consonant, and vowel. Since candidate words are narrowed down sequentially from the Daigodo words, pre-a1u S selection processing can be performed in parallel with the input of speech, enabling high-speed recognition processing.Also, starting from the beginning of the input speech By classifying the categories in order.

It becomes possible to quickly limit candidate words from the Inugado Word Dictionary-11. Furthermore, relatively L2 o in the input speech
Since the parts where a can be reliably achieved and the parts where it is uncertain are classified as separate layers, cross-rejects (exclusion of correct candidates) are less likely to occur when limiting the candidates. effective.

[Brief explanation of the drawing]

FIG. 1 is an electrical block diagram for explaining one embodiment of the present invention, FIG. 2 is an electrical block diagram for explaining the standard pattern pre-('1u selection process), is a diagram showing the classification of word-initial Cv, Figure 4 is a diagram showing the classification by 1.+11 middle consonants, Figure 5 is a diagram showing an example of sound si (1 minute bar 1), and Figure 6 is a diagram showing the classification by FIG. 7 is a diagram showing the tree structure of the word set; FIG. 7 is a diagram for explaining the classification of the word dictionary in the standard pattern storage unit; FIG. The figure and FIG. 10 are diagrams showing examples of grouping. 1... Voice input microphone, 2... Voice feature extraction section, 3... Preliminary word selection section, 4... Preliminary selection dictionary storage unit, 5... recognition processing unit, 6... standard pattern (dictionary) storage unit, 7... recognition result output unit, 31
. . . Processing start terminal, 32 . . . Silent interval detection unit in the input audio pattern, 33 . . . Continuation time length detection unit, 34.
... Word initial Cv distribution part 35 ... Word middle VC ■ Classification part, 4
0...Word group. Patent Applicant Ricoh Co., Ltd. Agent Taka Fang Akira Mi゛゛: )′ l〆′?・Figure I M2 Ward Figure 3 Figure 4 Figure 5 Figure #9. itJ /in'ul
t'ai,,'olma=,+umpx+ -nu
t -caAAAa-L+Rokutomo □ HI
,ζlSeisei Reiya 1s/IF! /1llk/l
h/-7bHdl/q/1xllri-Figure 6Figure 7

Claims (5)

[Claims] (1) A microphone for inputting speech, a feature analysis section for finding characteristic time series in speech, a preliminary selection section for selecting word candidates prior to recognition of large vocabulary word speech, and a preliminary selection A dictionary for preliminary selection to be compared during recognition processing, a recognition processing section for recognizing candidate words narrowed down by the preliminary selection section, a word standard pattern storage section for reference during recognition processing, and a recognition result output terminal for outputting recognition results. and a large vocabulary word speech characterized by having a preliminary selection section that sequentially narrows down candidate words based on classification or recognition results of initial consonants, vowels, middle consonants, vowels, etc. in input speech. Recognition method. (2) The preliminary selection is performed by detecting silent sections in the input speech and comparing the duration of the input speech pattern and the standard pattern to narrow down candidate words from the standard pattern. A large vocabulary word speech recognition method according to claim (1). (3) Claims characterized by having a dictionary in which large vocabulary words are grouped into a tree structure based on classifications such as initial consonant classification, initial vowel name, consonant classification in words, vowel name, etc. The large vocabulary word speech recognition method described in paragraph (1). (4) A microphone for inputting speech, a feature analysis section for finding characteristic time sequences in speech, a preliminary selection section for selecting word candidates prior to recognition of large vocabulary word speech, and a preliminary selection A dictionary for preliminary selection to be compared during recognition processing, a recognition processing section for recognizing candidate words narrowed down by the preliminary selection section, a word standard pattern storage section for reference during recognition processing, and a recognition result output terminal for outputting recognition results. The classification of standard word patterns is performed based on a hierarchical network structure of grouped category names and fixed category names, and the input unknown speech is classified according to the two category names. A large vocabulary word speech recognition method characterized by limiting candidate word names. (5) Classify categories such as phonemes for each segment of the unknown input speech, and according to the level of reliability of the phoneme classification of each segment, if the reliability is high, each segment is associated with a fixed category name. , the large vocabulary word speech according to claim (4), characterized in that, when reliability is low, classification of word standard patterns is performed based on a hierarchical network structure in correspondence with grouped category names. Recognition method.