JP2757356B2 - Word speech recognition method and apparatus - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and an apparatus for recognizing words and words, and more particularly to a method of feeding back information obtained from a language processing unit to a sound processing unit to obtain a more reliable recognition result. The present invention relates to a word speech recognition method and apparatus suitable for obtaining [Prior Art] A conventional general word speech recognition apparatus is disclosed in, for example,
As disclosed in Japanese Unexamined Patent Publication No. 61-175858, an input voice is converted into a digital signal, and a break between voices is detected and converted into a voice pattern in which features are extracted. This sound pattern is compared with a standard pattern of the dictionary to specify a sound. The above device can be said to be a device in which information of a sound processing result is reflected in language processing. On the contrary, there is also known a speech recognition device that reflects linguistic information in an acoustic processing unit. For example, IEICE Transactions '84 /6vol.J67-D.No.6, pp.693-700, “Top
Devices discussed in "Word speech recognition based on down-like phoneme recognition". In this device, first, as many word hypotheses as possible are established in advance, and phonological recognition processing is performed only on acoustic hypotheses obtained by dividing the word hypotheses. According to this device, acoustic recognition processing is not performed for phonemes not included in the word hypothesis, that is, phonemes that are not originally possible, which is somewhat advantageous in terms of both processing amount and recognition accuracy. . [Problems to be Solved by the Invention] However, even with this device, all possible candidates are considered for a preset word hypothesis, and precise acoustic recognition processing is performed on the candidates. There are still many problems, and there is a problem that the possibility of erroneous recognition is high. This is because the above device limits the range of possibilities as a phoneme by using linguistic information, but does not consider the use of acoustic information. The present invention has been made in view of the above circumstances, and its purpose is to solve the above-described problem in the conventional word speech recognition device, effectively use acoustic information to narrow down the range of word candidates, It is an object of the present invention to provide a word-speech recognition method and apparatus capable of improving the recognition accuracy and reducing the processing amount by performing precise acoustic recognition processing only in this range. [Means for Solving the Problems] An object of the present invention is to provide a word speech recognition apparatus for outputting a plurality of candidate words for a word input by voice, wherein each combination of the plurality of candidate words is A pair of single syllables of the same type among the single syllables constituting the candidate word is obtained, and pair determination is performed for each pair of the single syllables that do not match according to a pair determination rule prepared in advance. A word-speech recognition device having means for selecting from the plurality of candidate words, wherein the pair determination rule is to determine a single syllable pair by a voice feature parameter. Achieved by the device. [Operation] In the present invention, with respect to combinations of words left as candidates in the word dictionary matching unit, only pairs of syllable pairs that need to be identified in order to identify these words are paired. A determination is made, and a selection is made from the remaining candidate words based on the result. Words can be represented as a sequence of monosyllables, but the number of combinations of sequences of monosyllables is enormous, whereas the number of words that actually exist is much smaller. Therefore, by limiting the range of candidate words by collating with the word dictionary, the range of searching for possibilities is greatly reduced. In addition, as long as the correct syllable is included in the higher-order single syllable recognition candidate by the single-syllable unit pattern matching, the correct word is surely included in the candidate words. That is, the search range is narrowed under the guarantee that the correct word is contained, and the effect of the search is improved and the recognition accuracy is improved. Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram of a word speech recognition apparatus showing one embodiment of the present invention. In the figure, reference numeral 1 denotes an input unit of a word voice, 2 denotes an analysis unit of the input word voice, and 3 denotes a feature parameter outputted from the analysis unit 2 and a single syllable standard pattern stored in a single syllable standard pattern storage unit 4. A single syllable matching unit that performs matching and outputs a single syllable candidate;
The word dictionary matching unit 7 generates a concatenation of the monosyllable candidates output from the word dictionary 6 and checks it against the word dictionary 6, and the word candidate output from the word dictionary matching unit 5 and the word dictionary output from the analysis unit 2. Based on the characteristic parameters, the pair determination unit 9 performs pair determination according to the pair determination rule stored in the pair determination rule storage unit 8. The pair determination unit 9 aggregates the pair determination results output from the pair determination unit 7 and determines the final output. 4 shows a decision unit. Also, 10
Denotes an input word voice, 11 denotes the above-mentioned feature parameter, 12 denotes a monosyllable candidate sequence, and 13 denotes a word candidate. The outline of the operation of the present embodiment is as follows. The input word voice 10 input from the input unit 1 is
Is converted into a sequence of feature parameters representing the features of the speech, and is compared in the single-syllable matching unit 3 with all single-syllable standard patterns stored in the single-syllable standard pattern storage unit 4 in advance. The similarity is calculated for each syllable. As a result of the above calculation, a certain number of single syllable candidates are obtained from the single syllable matching unit 3 in the order of the similarity in units of single syllable, and the single syllable candidate sequence 12
It is output in a format as shown in FIG. The example shown in FIG. 2 is an example in which a word of four syllables “Yokohama” is input, and (a) shows single syllable candidates to be output for each single syllable limited to the top three places. It is a thing. Next, the word dictionary matching unit 5 first generates a connection of possible single syllable candidates from the single syllable candidate sequence obtained in the format shown in FIG. 2A. In the above example, 3 ⁴ = 81 cases will be generated. Word Dictionary 6 for all of this
Is performed, and only the concatenation of the monosyllable candidates existing in the word dictionary 6 is sent to the pair determination unit 7 as a word candidate. In the above example, the two are "Yokohama" and "Yokokawa" shown in FIG. 2 (b). The pair determination unit 7 obtains a pair of different types of single syllables constituting the word for each combination of the input word candidates, and according to the pair determination rule stored in the pair determination rule storage unit 8 in advance for the single syllable pair. , The characteristic parameter sequence 11 is investigated, and a determination result as to which of the two monosyllables this part is is obtained, and this is sent to the determination unit 9. In the above example, two pairs of “C” and “F” and “M” and “W” shown in FIG. 2C are obtained, and a pair determination is performed for each pair. The deciding unit 9 sums up the paired judgment results for each single syllable pair obtained from the paired judgment unit 7 and ranks all the word candidates output from the word dictionary matching unit 5 based on the summed up results. Then, a fixed number of word candidates from the top are output as final output. According to the present embodiment, the range of monosyllables required to be acoustically reliably identified by the limiting effect of word candidates is significantly limited as compared with the range of all monosyllables to be recognized,
Therefore, the processing amount is reduced, and the recognition accuracy of a single syllable is improved. It is needless to say that the configurations and the like shown in the above embodiments are only examples, and the present invention should not be limited to these. [Effects of the Invention] As described above, according to the present invention, in a word speech recognition device that outputs a plurality of candidate words, for each combination of the plurality of candidate words, a monosyllable constituting the candidate word Means for determining a pair of single syllables whose types do not match, performing pair determination according to a pair determination rule prepared in advance for each single syllable pair, and selecting from the plurality of candidate words based on the result. Because of this, the range of word candidates can be narrowed by effectively using acoustic information, and precise acoustic recognition processing can be performed only in this range, thereby improving recognition accuracy and reducing the amount of processing. It has a remarkable effect that the word speech recognition device described above can be realized.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram of a word speech recognition apparatus showing one embodiment of the present invention, and FIGS. 2 (a) to (c) are diagrams for explaining intermediate results obtained in the course of processing. It is. 1: input unit, 2: analysis unit, 3: single syllable matching unit, 4: single syllable standard pattern storage case, 5: word dictionary matching unit, 6: word dictionary, 7: pair judgment unit, 8: pair judgment rule Storage case, 9: decision unit, 10: input word voice, 11: feature parameter, 12: monosyllable candidate sequence, 13: word candidate.

──────────────────────────────────────────────────続 き Continuation of front page (56) References JP-A-57-128400 (JP, A) JP-A-57-102689 (JP, A) JP-A-60-121499 (JP, A) JP-A 54-128 148305 (JP, A) Proceedings of the Acoustical Society of Japan (October 1986) 2-3-16, p. 83-84 (58) Field surveyed (Int. Cl. ⁶ , DB name) G10L 3/00 531

Claims (1)

(57) [Claims] A word-speech recognition method for analyzing an input word-speech to output a feature parameter, performing a word recognition and matching process on the feature parameter, and outputting a plurality of top candidate words. Decomposing a word into syllables; obtaining pairs of syllables of different types for syllable parts corresponding to the decomposed words; and a pair determination rule prepared in advance for the obtained syllable pairs and the feature. A word speech recognition method, comprising: determining a syllable pair according to a parameter; and determining a final candidate from candidate words by combining results of the pair determination. 2. Input means for inputting a word as a word voice, analysis means for converting the word voice input from the input means into a sequence of feature parameters, single syllable standard pattern storage means for storing a single syllable standard pattern, and the analysis means A single-syllable matching means for comparing single-syllable feature patterns from the single-syllable standard pattern from the single-syllable standard pattern storage means to extract single-syllable candidates for each single syllable; and a word dictionary for storing standard single-syllable syllables Word dictionary matching means for comparing the connection of single syllable candidates extracted from the single syllable matching means and the word dictionary, and extracting the connection of single syllable candidates present in the word dictionary as word candidates, A pair determination rule storage means for storing a syllable pair determination rule, and a different type of syllable pair for the single syllable candidate,
A pair determination means for performing a pair determination from the single syllable pair determination rule stored in the pair determination rule storage means and the characteristic parameter of the single syllable from the analysis means; and a pair determination result output from the pair determination means. Determining means for counting and determining a final output.