TW449734B - Keyword spotting method for mandarin speech without using filler models - Google Patents

Abstract

The invention is a keyword spotting method for mandarin speech without using filler models that by means of the connection structure for the regularity for sound origin in connection with tune origin in pronunciation of mandarin character, first identify the possible locations of the end points of character pronunciation signal in the input speech signal and then form a set of character sound link map with the connection states of speech paragraph generated by the end points of the speech signals, and then recognize and verify the keyword from the character sound link map to spot the keyword in the speech signal.

Description

449734 ----------- ----------- V. Description of the invention (1) [Field of the invention] The present invention relates to the technology of speech recognition. Chinese keyword extraction method using redundant word model. [Background of the present invention] According to the general speech recognition system, the range that can be recognized is a limited vocabulary. The speech input by the speaker and a certain sound in the vocabulary must completely match 'cannot be mixed with other speech, otherwise recognition It is easy to make mistakes. For example, in the voice ticket booking system, 'assuming that the vocabulary range that can be recognized is the station name, when the user is asked at the end station, he can only answer "Taipei" verbatim instead of saying "I want to go to Taipei." "Or you wo n’t get the correct identification result. Therefore, 'in order to improve the aforementioned deficiency, there is the development of keyword extraction technology to allow users to enter any words freely, and the possible keywords are automatically extracted by the speech recognition system, thus eliminating many inconvenient restrictions on speech recognition' and In the conventional method of keyword extraction, as shown in the fifth figure, a keyword model (51) and a redundant word model (52) are first used to construct a speech model (53) of a connection relationship as shown in the figure, and then used. The Viterbi algorithm matches the speech signal (55) with the possible connection path of the speech model (53) to obtain a matching path map as shown in the sixth figure and obtain a maximum approximation (Maximum Like 1 ihood (ML) = Matching path (6 1), you can know that in the signal, those paragraphs are f sounds of redundant words, and those paragraphs are the pronunciation of that keyword. For the extracted key β, you can send θ I VeriHCau〇n and then make a further decision to accept (Accept) or reject (Rejecf), for the accepted paragraph

Page 4 449734

V. Description of the invention: 2) and its identification results are the key words that have been extracted. However, the aforementioned conventional keyword extraction method involves the establishment of a redundant word model (Garbage Mode 1 or F i 1 1 er Model) (52), and in order to obtain a better extraction effect > in European Patent EP0800158 A1 971008 '' words Extraction (Word spo 11 ing) In case B, it has been pointed out that the superfluous word model must be established based on the superfluous words commonly used by users in the application, and the Chinese speech recognition __ US Patent Case USP56490 57 " Using Keyword Model And non-keyword model speech recognition (Speech recognition employ i ng ke y word mode 1 i ng and non-keyword mode 1 i ng) " also illustrates the same phenomenon because the scope of redundant words often varies with the application area ( Task Domai η) and the user ’s terminology have changed. Therefore, in order to obtain a better extraction effect, the verbal vocabulary that often appears during use must be used as the basis for building a model, that is, the verbal model must be based on different Application-specific adjustments (Task Dependent ), And the time spent adjusting all involves the reconstruction of superfluous words. Therefore, the disadvantage of the conventional keyword extraction method is that whenever the application city is changed, the superfluous words that may be generated by users in this application are collected. It is extremely tedious to build an appropriate redundant word model and collect vocabulary based on this information. Therefore, it is necessary to improve the aforementioned keyword extraction method. Because of this, the spirit of active invention was' immediately thinking about a Chinese keyword extraction method that can solve the above problem without using redundant word models. After several research experiments, this novel and progressive invention has been completed. Overview] 4 497 3 4 V. Description of the invention (3) The purpose of the present invention is to provide a Chinese keyword extraction method to extract keywords in a voice message without the need to establish a redundant word model for the application field. ^ In order to achieve the aforesaid objective, the Chinese keyword extraction method of the unnecessary word model proposed by the present invention first uses the consonant structure of consonants and vowels in the initials of Chinese characters, and finds its possible from the input voice signal The end points of the zigzag signal and the syllabic paragraphs are used to construct a zigzag connection diagram which is a continuation of a plurality of syllabic paragraphs; and then the keywords are identified based on the word length of the keywords to avoid the redundant Word model and correct Chinese keywords. Since the novel design of the present invention can provide industrial utilization and indeed enhance the efficacy, a patent is applied in accordance with the law. In order to enable the review committee to further understand the structure, characteristics and purpose of the present invention, detailed descriptions of the drawings and preferred embodiments are attached as [Simplified Description of the Drawings]. Word model Chinese key method processing flow. μ flat drawing Figure 2: This is the general sound and vowel model used in the Chinese key method of the unnecessary word model of the present invention. The third figure is the maximum approximation obtained by matching calculations based on the assumption of different numbers of sounds. a The fourth picture: it shows a voice signal " from Hsinchu to Taoyuan "

Page 6 449734 V. Description of the invention (4) The character-to-speech connection diagram formed by the Chinese keyword extraction method of the shape-free and redundant word model according to the present invention. Figure 5: Shows a conventional speech model constructed by a keyword model and a redundant word model. Figure 6: Shows the matching path map obtained by matching the voice signal with the voice model using the Viterbi algorithm. [Illustration of drawing number] (2 1) Initial model (22) Final model (31) (42) (43) (44) Word audio signal endpoint group (4 1) (5 5) Voice signal pronunciation waveform (4 5) Word sound Connection diagram (5 1) Keyword model (52) Redundant model (5 3) Speech model (61) Matching path [Detailed description of the preferred embodiment] Chinese keyword extraction related to the free redundant model of the present invention A preferred embodiment of the method, please refer to the first figure, which shows that the method of the present invention mainly includes a two-stage processing step (SI, S2) of searching for a voice signal, a paragraph of words, and a keyword search. The processing step (S1) of the voice signal word and sound paragraph is used to find the word sound in the voice signal to be recognized

4497 3 4 V. Explanation of the invention (5) The possible positions of the signal endpoints, and according to (Training-Graph), and the key = sound connection circle looks for keywords from the word sound connection map. According to U = reason: Suddenly extract correctly The purpose of Chinese keywords. According to the processing steps that can be used to avoid the redundant word model, the consonant sound of the block (for example, the word "" "in the" electricity "vowel sound is blocked by the vowel organ (eg" & 子 ～ 幻”) And vocal vocal tremor, and according to the“ -bow ”of the Chinese character voicing”) can use the voice model and the signal to compare the regularity of the vowel, the vowel and the vowel paragraph, and then find the word sound; the number matches The processing steps first need to establish the same / failure. Therefore, at this stage, the vowel model (22) is followed by: the f mother model f) can identify whether the voice signal is: I, the mother model, so as to be known It can be distinguished by radiant θ · or Mingmu, but it does not need to be a sound pattern such as a batch of Xi J symbol -¾ 疋. The establishment of the universal voice and final model = training using the training data of the voice and final model only = Collecting all initials and finals is obtained by combining the existing general models of special initials, initials, or finals, or by merging the ϋΞΪ = # model. The rule of signal comparison is in the words 3 = type, and Han can use the voice model and the output of the voice signal; 1 :: =: = vowel paragraph, and then the rule of sign comparison is Viterbi t ^ during the speech model When matching with the signal = method, = the public second 'not shown in the first picture, assuming that the input signal has a maximum of L characters, five, the invention is called ^' 1 The matching calculation process of the t-round voice signal at ^ can be obtained The number of words is the maximum approximation ML and MLl of the number of presuppositions L, each of which is the maximum approximation: 'Yingyu signal signal endpoint group (3 1)' where the L maximum approximations are made by using the Viterbi algorithm after cutting The result may have a slight interpolation = two division (lnserti〇n / Deleti〇n) error, that is, the number of positive consonants and the number of phonetic sounds corresponding to the maximum approximation are 1, 1, 2 In order to allow the existence of such errors, when selecting the matching path, the top} ^ highest approximate value is selected from ML and MLl instead of taking only the 1 highest maximum approximate value. There is a high probability that the corresponding zigzag j can include the number of zigzags of the input voice. In addition, in the f 貭 calculation process, in addition to the information of the number of phonetic numbers, the position of the second paragraph of the signal can be obtained. The continuation of these phonetic paragraphs can form a phonetic connection. Similarly, there is a high probability in the connection diagram that there is a correct connection. The fourth picture of the word segment shows a practical example, which uses a Viterbi algorithm to match the pronunciation waveform of a speech signal from Taoyuan " (41) to the maximum approximation obtained from the matching calculation. Choose the first three highest maximum approximations, which correspond to these three Hobinxi counties 间 _ ^. Β 7. ^ ^ The number of vowels with the maximum approximate value of dare is 6, 7, 8, among which the correct number of vowels 6 is included, and the endpoints of the voicing signal corresponding to the three highest maximum approximations are, respectively, ^ 42, 43 J4) combined, you can get all possible phonics & points 0 to 9 and sub-syllables ⑨ (indicated by the start point to the end of the arrow). ). After the construction of the connection map of the input voice is completed, the first step can be performed.

Page 9 4497 3 4 V. Description of the invention (7) The processing step (S2) of the keyword search in the second stage, that is, searching for keywords from the phonetic connection diagram, using the example described above, if the candidate keyword is 29 This is the main place name in Taiwan, and assuming that the contents of the keyword collection are all 2 place names in Taiwan, the connection map (45) of the fourth circle contains it! j syllables with 2 syllables 0-2, 0-3, 1-3, 1_4, 2-4, 3_5, 4-6, 4-7, 5-8, 5-9, and 6-9, of which The phonetic paragraphs 1-4 and 5-9 are the pronunciations of "Hsinchu" and "Taoyuan". Therefore, when the Chinese speech recognizer based on the Hidden Markov Model (HMM) is used to recognize this paragraph, The correct result should be obtained. In addition, the remaining paragraphs will each get a place name through identification. Therefore, the result of each paragraph identification can only be regarded as a hypothesis of generating keywords (Hypotheses), so after identification, there will be 11 Keyword hypotheses. These hypotheses need to be further acknowledged to make an acceptance or rejection decision on the identification result, and the final accepted confirmation result is the extracted keywords. For this example, the results of its identification and confirmation are: Paragraph identification result confirmation result 0-2 Factory Zhong 坜 J Rejected 0-3 Factory Zhong1-3J Rejected 1-3 Factory Xinying J Refused 1-4 Factory Hsinchu J Accepted 2 -4 Factory Hsinchu J rejected 5-8 Factory Taoyuan J Rejected 5-9 Factory Taoyuan J Accepted

Page 10 449734

Among them, paragraphs 4 and 5-9 have higher confirmation scores, and the recognition results are accepted. The remaining paragraphs will be rejected due to low scores regardless of the recognition results. Therefore, the keywords corresponding to paragraphs 1-4 The key shrine corresponding to "Hsinchu" and paragraphs 5-9, "Taoyuan", is the key word from the extraction. In order to increase the speed of Chinese keyword extraction, another preferred embodiment of the present invention is to reduce the number of confirmations and recognitions by reducing the number of assumed keywords and candidate keywords. The reason is that it is based on the information provided by each phonetic paragraph to filter out paragraphs that are impossible to use as keywords. The more reliable information is the final result of each phonetic paragraph. C Chinese phonetic vowel recognition is more reliable.) Therefore, if the result of the first F name vowel recognition is taken for each syllable paragraph (general recognizer takes F = 5 to 10 to obtain a 98% accuracy rate) as the candidate vowel, then for the key temples containing k consonants, §, As long as you check whether the finals of each vowel contain the recognition result of the first F vowels, you can know whether the subsequent k-character paragraphs are likely to be distant keywords. And for this k-character paragraph, it was originally identified When comparing each candidate key word with k vowels one by one, 'but if you have the information of the first F vowels of each vowel', you only need to screen the candidate keywords that the remaining vowels match. Comparison, but if there are no candidate keywords left after screening, it means that the first F vowels of each syllable of this paragraph are not in the keywords, and no further processing is required, so the keyword confirmation and The number of identifications and the speed of effective extraction. : For the example of the previous embodiment, the original 〖丨 character segment = J7Jsolid ', that is, the confirmation action can be reduced from u times to 7 times, and in the time of recognition, the original, each character sound When identifying paragraphs, you need to pronounce 29 place nouns.

Page 11 449734 V. Description of the invention (9) Comparisons' The method used in this example is to select only one or two candidate words in the vowel section (such as ^) for comparison, so it can be significantly accelerated. The key 'extraction speed. The Chinese keyword extraction method of the free word model of the present invention has excellent performance under actual experiments. Under the condition that the application field is the query of the personal extension number in the telephone and the keywords are 2000 personal names For testing, the user only needs to say “Please transfer to XXX” or “Ask & phone number” on the phone, the system needs to extract the name of the person and correspond to the extension number. The keyword extraction using the redundant word model is known. The method failed to extract j 0 sentences in 5 of the 2 test sentences, and in the case of using the same recognizer, the method failed to extract 8 sentences. Therefore, it does not need to establish redundant words for the application field. Under the requirements of the model, it achieves an excellent keyword extraction effect. _In summary, the present invention, regardless of its purpose, means and efficacy, is different from the known technology in its characteristics. It is a breakthrough on the lambda meter of the Chinese keyword extraction method. Observing clearly, granting a quasi-patent as early as possible, and benefiting the society, I feel a sense of virtue. It should be noted that many of the above-mentioned embodiments are merely examples for the convenience of description. The rights claimed in the present invention should be based on the scope of the patent application, and not limited to the above-mentioned implementations.

Claims (1)

449734 6. The scope of patent application, 2 Bai 1 Yuzhong A Chinese keyword extraction method for seeding A, using \ hoo stroke to extract the keywords contained in the voice signal, the method mainly includes the following steps: (A) Using the Chinese Ning * 仏 ^, the consonant structure of the consonants and vowels in the son a to input the phonetic Indian girl a from 3, and construct a possible endpoint of the phonetic signal and Connection diagram; & a syllabary (B) concatenated from a plurality of syllabic paragraphs to confirm the keywords based on the key quotes E. a long to identify from the phonetic connection diagram and 2, 2: the neutral keyword extraction method described in the first patent application scope of the unnecessary redundant word model, where K Chinese. Danjia step (A) mainly includes the following Sub-steps: Shao (A1) ^ — universal initial model and a universal final model connected to the speech model 丄 sub-speech model and signal comparison rule: the input s signal to be matched with the speech model multiple times, ^ Yako finds the maximum approximation obtained by the four matches and finds the corresponding diacritic phonetic paragraph; and u ^ and (A2) finds the maximum approximation that makes the previous multiples the largest, and the corresponding multi-array word and audio signal endpoints. The majority ... And find a point to combine the phonetic connection diagram. , And give "signal terminal 3. As described in _ please patent scope of the useless unnecessary die extraction keyword extraction method, wherein step (B) is mainly from the red Chinese, step · · # the following sub-steps (B1) Find out all the words that meet the § Division 4497 3 4 4497 3 4 from the phonetic connection diagram. 6. The phonetic paragraph continuation of the patent application scope; and a tjv / ϊτ -rs: m, u, DO 2. Continue to identify and confirm (B2) For each character found in the paragraph, change only ^ recognize, ^ ^ ^ 中文 the Chinese without the use of redundant words model 4 · Shenzhou Patent Scope Item 1 ▲ also White pregnancy under γ evening; semi-keyword extraction method, in which step Lai (B) mainly includes the following sub-steps: D1, ^ _, identify the vowel in each pregnancy paragraph '(B1). Each = to find the highest number of identifying the highest number of social fruit k 0 '找出) to find the word length of all 2 keywords from the phonetic continuation, and continue to the # 字音 袄 found, According to its corresponding candidate finals, select keywords that have a match with one of the corresponding candidate finals for identification and Recognition. 5. The method for extracting Chinese keywords from the redundant word model described in item 2 of the scope of patent application, wherein the model and signal comparison rule in Yuzi Shaoxiang (Aug. 1) is a Viterbi algorithm. 6. The Chinese keyword extraction method of the superfluous word-free model described in item 2 of the scope of the patent application, wherein the universal initial molding in the substep (A1) is only used to identify whether the voice is a voice, The general finals model is only used to identify whether the speech is a final. 1. As described in item 2 of the patent claim, the method of taking out the Chinese key of the 5-bow model is to take the %% method, wherein, in the sub-step (A1), 'the number of vowels molded by the voice is set to 1 to L, in order to perform the matching of the input voice signal 1 times 'to find L maximum approximations', respectively, where L is the maximum number of words allowed in the input voice signal. 449734 VI. Scope of patent application 8 · Chinese keyword extraction method based on the unnecessary redundant word model described in item 3 of the scope of patent application, wherein in sub-step (B2), the Chinese speech based on the hidden Markov model is used The recognizer performs recognition. 9. The Chinese keyword extraction method of the redundant word model exemption model described in item 3 of the scope of the patent application, wherein in the sub-step (B2), a key word hypothesis can be obtained after each syllable paragraph is identified, the key Words are assumed to be extracted keywords after being confirmed and accepted. 10. The Chinese keyword extraction method of the unnecessary word model described in item 3 of the scope of patent application, wherein, in the sub-step (B2), a key word hypothesis can be obtained after each syllable paragraph is identified. The key Words are assumed to be extracted keywords after being confirmed and accepted. 11. The Chinese-language keyword extraction method as described in item 4 of the patent application scope, wherein in the sub-step (M), the top 5 to 10 candidate finals with the highest recognition result are found. 12. The Chinese-language keyword extraction method as described in item 4 of the patent application scope, wherein in the sub-step (B2), the Chinese speech recognizer based on the hidden Markov model is used for identification. Page 15