TW569182B - Voice tone changing recognition method in Chinese voice recognition system - Google Patents

Abstract

The invention provides a voice tone changing recognition method in Chinese voice recognition system targeting at automatically recognizing voice tone changing sentence in terms of text string relation and sentence boundary to increase accuracy of system's recognition when continuously generating with level-3 tone characters or repeated characters during Chinese voice (pronunciation).

Description

569182 V. Description of the invention (i) &lt; Field of invention &gt; The present invention is a method for identifying, broadcasting in Chinese, especially in Chinese; The language of 1i &lt; Background of the Invention &gt; According to the current speech recognition at home and abroad = Yes: in the languages of various countries = in Chinese and Chinese methods to improve the positive degree of system identification; however, the success of :::; There are sounds, "two sounds m ::: reading the single word reading * (Note;)" 1 "2: in the ordinary Chinese sentence, if there is no problem with the sentence according to the text, at this time; Li = sound or soft Laiha contains the continuous "two sounds of σ, beijing, the same as the king, but when the sentence, the change in the pronunciation and pronunciation of the same superposition and pronunciation used in the appellation (pronunciation 7 丄 :: == ί) and the difference between the pronunciation" There are cases of inflection that make the sound off: clever examples explain the following; the situation is also related to the number of words in the word, then the -dn :: string in the voice if: each word is "three" you "etc. Two-character string, and "i 你: 原 ί 音 ::;! :: V—think page 4 569182 V. Description of the invention (2) V) The tone, but the voice (when pronounced) is changed to (Ding Yi You / Yunyi V) Tone. (2) In the three-character string, if there are two consecutive "three" sounds, the first "three" sounds must be changed to "two" sounds; If All three words are "three" sounds, so the first and second words must be changed to "two" sounds. For example, in the three-character string of "submarine", the original note is (&lt; Yi bow / corpse乂 \ V 六 一 厶 V) tone, but it changes to (<一 巧 / (乂 \ / 六 一 厶 v) tone when you pronounce it; for example: the three-character string of "Presidential Award", the original note is (卩 乂 厶 v 六 乂 厶 v tj — You v) tone, but it changes to (改变 / 六 乂 厶 / Μ —You v) tone when you pronounce. (3) Among the four-character strings, if the four characters are all “three” sounds, then the first and third characters must be changed to “two” sounds; for example, “rarely countable”, etc. In the word string, the original Zhuyin is (Li Yi 幺 ν Li Yi 万 万 Vango V corpse V), but when it is pronounced, it changes to (Li Yi 幺 / 力 一 幺 V i / corpse V). (4) Among the five-character words, if the five characters are all “three” sounds, then the first character and the third and fourth characters must be changed to “two” sounds, such as: "Nine-nine" five-character string (including the digital form of this example), the original Zhuyin is W — and v W — I) — v Μ a and V), but when it is pronounced, it changes to (Η — and / Μ — 又 V Μ—The tone is again and again. (5) If the number of consecutive "three" sounds is even and there are six or more words, the first two words are a group, according to two consecutive "three"

569182 V. Explanation of the invention (3) What is the sound? For example: the six-character string of "99,999,99", the original phonetic is (ij-一 VL) ^ 一 v Lj 一 V Lj '一'% V ij a

Lj and V) tones, but change to (I) and / q — ^ v Μ — and / I] — ^ ν% y — tone 0 (6), if the number of words in three consecutive sounds is When the number is odd and there are seven or more characters, the first two characters are a group, according to two consecutive “three” sounds, but the last group is three characters, according to three consecutive “three” Pronunciation: For example: "Five Five Five Five Five Five Five" the original note is (乂 v 乂 v 乂 V 乂 V 乂 V 乂 V 乂 V 乂), but when it is pronounced, it changes to (乂 / 乂 v 乂〆 乂 V 乂 〆 乂 〆 乂 V) tones. It can be seen from the above that the above-mentioned tone change rules are applicable to general words and proper nouns. When there is a "word boundary" distinction in a three-character string, it must be distinguished according to the "word boundary" before applying the above criteria. Applicable to the combination of numbers and general words, such as when used for "name", and its tone change law is based on the following methods: (1) In the application of "name" identification, the "last name" and the Separate the "name", such as "Yo You Ma" or "Liu Shuibian". Although the three words are all "three" sounds, the original Zhuyin is (门 丫 v — 又 V —5 &lt; v) and (力 一 又 V deadlock \ V 勹 一 弓 V), but "horse" and "liu" are "surnames", and they say the "three" sounds without changing the pitch. The words "youyou" or "shuibian" follow the two consecutive "three" The rule of "voice" reads the "two" sound in the first word of the two words, so it is changed to (门 丫 V — 又 / 一 又 V) and (力 一 dead 乂 \ / 勹 一 弓 V).

569182 V. Description of the invention (4) (2) When there is a "word boundary" distinction in other three-character strings, such as: "President Jiang" or "Presidential Award", among which, the three words "Presidential Award" are "three ", The original Zhuyin is (V 六 乂 厶 v I] — 尤 V), but because it does not include the" surname ", so still using the above rule, the first two words pronounce" two ", so Do "卩 乂 厶 / 六 乂 厶 / L] 一 尤 V". And the words "President Jiang" have "three" sounds. The original note is (M—You VV 乂 厶 V 六 乂 厶 V), but the word "Jiang" is "surname" and is not part of the word. "The second two words" President "read the first word" two "according to the above two rules of three consecutive tones, so say" 1] Yiyou V 卩 乂 厶 / 六 乂 厶 V ". In another type of title with a superimposed form, that is, if there are two identical words in the title of a Chinese phrase, the superimposed words are even if the two words are "three" sounds, and the pronunciation (pronunciation) is not Similarly, the first word of the superimposed word is pronounced according to the original sound, and the second word should be soft. For example: "Grandpa", "Grandma", "Dad", "Mom", "Brother", "Sister", "Brother", "Sister", "Wolf" and so on. In summary, tonal changes in Chinese words and phrases are inevitable, and the differences in pronunciation and pronunciation formed by the tonal changes will obviously increase the difficulty and incorrectness of speech recognition, causing practical problems such as: In a Chinese speech recognition system, a Chinese database is first input by using a pronunciation method, and then a Chinese word (or string) is to be input into the speech recognition system by using a speech method, and the recognition by the speech recognition system is expected. Processing, and can extract the correct Chinese words or signals from the Chinese database and output them for use in order to achieve the convenience of voice control; however, in actual use, go to

569182 V. Description of the invention (5) Going to the same—Chinese ^ is different from 'zhiwu mr string), the sergeant's voice due to the inflection relationship and the pronunciation example description is wrong also affects the follow-up operation, try to be realistic '' Telephone system installation ^ :: The company telephone system of the voice extension extension: The above-mentioned voice name already has the identification system and system to input the incoming call; and the company's Ren Litao: Mike Hunt enters the voice recognition system -Chinese data; :: Identification systems generally use the pronunciation method to enter the number ", then use two m: i: name" and its extension "extension data to discriminate," θ identification system must input the sentence or 'To extract the correct Chinese words from the Chinese database', "People use them, ㈤: The original data of the caller and input person is" Ma Youyou "—the Chinese database of the identification system is based on pronunciation) (Door ^ ^ V I and V) storage, so that the input voice data (gate I / I and V) could not be easily identified and caused the subsequent system to work, and the staff could not quickly transfer automatically because they could not recognize the correct name to Indeed extension, reducing the effectiveness of the use of the voice switching system. ^ 2) If using a voice registration system in a hospital: All of the above voice registration system devices already have a voice recognition system to input incoming calls σ quot; k material like name netting 'into the voice through a microphone In the recognition system, the speech recognition system in a hospital usually uses a pronunciation method to input a Chinese database. For example, the name of a person in the hospital and the registration form to which they belong. In use, the speech recognition system must respond to the voice input by a caller. Data plus $ 8 pages, 569182, and 简 j brief description of invention (6) identification, in order to extract the correct from the Chinese database for use in rotation, such as ·· Incoming call rotation data was originally "horse identification number (such as Registered order number), but the input language η 丫 v — / / 一 又 \ /), and the “Ma Youyou” in the speech recognition system uses pronunciation (门 丫 V — and V to make it easy to input the voice data (门 丫 V —Also, it is easy to identify and output correct Chinese words or signals, which is troublesome in the manufacturing industry. For example, the hospital cannot recognize the correct name and the long string of registered slip numbers may also occur. In the system of data—and V—Κ ν, it is impossible to change the tone or signal in the subsequent voice, and it is stored together with the document data)) The system cannot register the voice of the system: 3), such as the voice dialing of the hotel room extension system All have a set of stand-up: 隹 上 隹 voice data, such as name or 4 ::; cognition = come = incoming call recognition system; and the restaurant's access to the voice is 4 μ &amp; ~ &lt; ° 口Generally speaking, the identification system usually uses the room reading database first. ㈤: When a hotel room-room number and live electricity D are used, the speech recognition system must be targeted at: t data f to identify, and Chinese data The data in the library is taken out as "Min ef ° * 5fl" and used in turns. For example, the original data of the incoming call = Friends and the same identification number (such as the room number), but the input m: (（门 丫 v… …), And the "Ma Youyou" in the Chinese database of the speech recognition system is stored in pronunciation (门 丫 v one after another ^), which makes it easy to input the voice data (ηγν one after another and w cannot be easily identified to output Correct Chinese words or signals, resulting in Industry in trouble, such as the inability to identify the caller · Hotel phrase V /

569182 V. Description of the invention (7) The correct name of the tenant in the middle cannot be automatically transferred to the call in the store's telephone system that requires the transfer of the guest room-~ and the name of the official salute of the meal, if it matches, it will be transferred in the future ^ = Extensions and rooms should be promoted. J got the guest room just in case <Objective of the Invention> One of the objects of the present invention is to provide a neutral sound-changing ancient week of Nokushi ^ ^ T again. Oral 9 recognition system Chinese ii: = i It is in Chinese phonetics. When there are “three” in succession = sub-connected 赉 sounds will naturally produce phonetic transliteration ϋ = change according to the substring relationship and the word boundary, and identify the already ... Adjust the § sentence to increase the accuracy of the system identification. . For the purpose of the present invention to achieve the above-mentioned objectives and its effects, the technical means for better implementation # 丨 # j 征 所 十 木 and the "Explanation of the Invention" "exemplary and illustrated with drawings are described below: The present invention is a" Chinese "Five-voice method", please refer to the identification of the tone change in the first music system. $ # Ί ί The Chinese database of the s recognition system: 俜 Using ordinary or other input devices (,, Ding inch thick, and using the keyboard (2 ) And # by 1 h ρ, input Chinese data into the data buffer]] and decompose and process it by clever sentence judgment; r Q ,. Chinese caution in buffer (2) * I 早 疋 (3) The processing data is judged to be "called fold bangbao ,,,,,," to judge, if the Chinese data does not work, it will generate storage media such as hard disk memory, etc. # f • 于 贝 予(4), and then stored in "continuous" three "sounds" will produce: 浐; if this: the text data is judged to be stored in the storage medium such as hard disk memory ^ "sound # material (5), and then stored in the media (6 ) In the "J H", so that hard disk memory, etc. to store the Chinese shell database: and hard disk memory chat 18 2

Other storage media (6) are related to the industry system. The signal comparison processor (7) is connected into one

When in use, a Chinese voice data is input through the microphone, and the Chinese language is then compared with the signal processor (round out (10)) for subsequent systems to refer to as shown in the second figure, which is connected with pronunciation or the same stack. When the comparison data is entered, it is divided into the general comparison process (丨 3) and the titled general comparison process (1 2). For continuous "three" sounds, the number of consecutive "three" sounds is divided into two consecutive "three" sounds (17, 4 consecutive "three" sounds (19), and six consecutive "three" sounds (21). The number of words, and according to the different number of words for two consecutive "three" sounds (17) sound comparison data (23), for the first one, two words to establish "two consecutive" three "sounds (19) Acoustic comparison data (2 5). For the first, third, and fourth words in 5 or other speech input devices (8), the input speech data is identified by signal encoding (9) 7), and the comparison result is then compared. Receiving the correct signal for follow-up operations. Schematic diagram of the changing pattern of the continuous three-voice continuous tone in the mantra Enter the decomposition procedure according to the word boundary rules (j-matching procedure (12), continuous "three" sound comparison procedure (14), where if the data is directly built in the database (i 5 comparison procedure (13), The data program (16) will compare the data area), 3 consecutive "three" sounds (丨 8), 5 consecutive "three" sounds (20), and 7 consecutive "three" sounds ( 22) The construction of different comparison data, including:) For the first word, "two 3 consecutive" three "sounds (18) then &amp;" sound comparison data (24), for 4 of which The first and third words establish "two consecutive" two "sounds (20), and then compare the two" sound comparison data (2 6).

569182 V. Description of the invention (9) Six consecutive "three" sounds (21) For the first, third, and fifth words, "two" sound comparison data (27) is created, and for seven consecutive "three" sounds ( 22) then establish "two" sound comparison data (2 8) for the first, 3, 5, and 6 words, and compare all the converted data (2 3), (2 4), (25 ), (26), (27), (28) are built in the database (15). If it is the title overlapping program (1 4), a soft comparison data (2 9) is established for the second word, and the variable comparison data (2 9) is built in the database ( 1 5). Then the database (丨 5) has completed the complete comparison data for the general comparison program (12), the continuous "three" sound comparison program (13) and the title superposition comparison program (14), especially The most probable part of voice change between speech and pronunciation; when the voice signal (30) is entered and you want to identify, you can use the complete comparison data stored in the database (15) to perform a voice comparison (3 1 ), And then output the comparison result (3 2) for subsequent systems to receive the correct signal to perform subsequent operations. In order to let those who are familiar with the technology to better understand the features of the present invention, the embodiment of the "telephone system of special voice transfer" is used to explain the use of the present invention: Bei Qing refers to Figure 3A and uses it first. Caller (4 〇); proceed = automatically answer and announce the welcome speech (41), such as: "0〇 Hello company ^ tell me who are you looking for? ”; Then proceed until the user speaks (voice) the person C 42), such as:“ President Chen (test ~ / 卩 乂 厶 / 六 乂 乙 V) ”; then proceed to the general comparison (4 3) , That is, the comparison between President Chen ’s voice (test T7 乂 人 / 云 乂 厶 v) and pronunciation (彳 4 / 卩 乂 厶 v six; if the comparison fails (due to the difference between the pronunciation and the pronunciation)

Page 12 569182 V. Description of the invention (ίο), proceed to continuous ", (彳 ㈠〆T7 乂 厶 / 卩 乂 厶 / 六 乂 厶 v compared to 43 to increase), such as:" 去 乂 厶 v） cognition ( 46) 〇 ”sound comparison (4 4), that is, the speech of President Chen ’s speech six people v) and the converted speech (4 comparison; the comparison is successful (the correct rate has been more than normal)” and then proceed to the speech reading comparison Results and processing (4 5 What you are looking for is the voice of President Chen (4 4 / 卩 乂 厶 / 'I will transfer it for you right away'): the transfer is successful and the speech is over to ^ 照 According to the third picture Instructions, the user's incoming call (47); proceed, please listen and announce the welcome speech (48), such as: "Here is what the baby family is looking for \ \ 门 (b 幺 V pronunciation has sister (音 ((correct The result is (b), and the comprehensive speech recognition can be more accessible. \ Do you want me? Whom are you looking for? ”, And then proceed to the person who spoke (voice) and fell (49), such as" Chen Xiao Younger sister (^ / 丁 一 么 ν door pronunciation has been compared with ', ·,' y; if the comparison fails (due to the difference between the voice and the sister's rz), the process is complete Predicate superposition comparison (51), that is, the comparison of Chen Xiaoding Yi Mo v gate entry, gate entry ·) and the converted language Yi gate v gate \, gate \ ·); the comparison is successful, compared with 43)) Proceed to the voice reading comparison "Wait a minute, my name is Chen Xiaomei \ ·)": then proceed to the general comparison (50), that is, Chen Xiaomei / Ding Yima \, door ~ ·) and the pronunciation (cause —One ^ 7 / by, ^ rate has been compared ~ β π jealousy 7 'and processed (5 2), such as: "work": tingyin one: know η' ·) listen to the phone "; transfer into voice: It can be known that, for any Chinese speech input, the Chinese language of the present invention, i ;; 糸 system can perform general speech comparison such as ⑷ ... 50 ". When the comparison fails, it will automatically use the string relationship and word boundary, and then use the conversion.

569182

V. Description of the invention (11) Recognition of the changed speech (4 4) or the term “triple” sound word connection plus the system's identification of the correct tone change of the system (51), or the same superposition When produced. For example, the continuous "three" sound ratio can effectively reduce the difficulty of recognizing continuous speech in speech, and increase her arsenic: "The invention" recognition method of tone change in Chinese speech recognition system "according to the present invention can indeed be used by the above The method of disclosure achieves the above-mentioned ί ΐ ΐ The application of the present invention has not been seen in publications or used publicly, and it has met the requirements for novelty and progress of the invention patent.

The figures and descriptions disclosed above are only examples of the present invention, and are not intended to limit the embodiments of the present invention. Anyone who is familiar with the technology may make other equivalents based on the features and scope of the present invention. Changes or modifications shall be covered by the scope of patent application in the following case. &lt;

Page 14 569182 Schematic description of the first diagram: The block diagram of the process of the present invention. The second picture: It is a schematic diagram of the way of changing the pronunciation of the "three" sounds in the Chinese phonetic sound. FIG. 3A is a schematic diagram of an actual control block according to an embodiment of the present invention. FIG. 3B is a schematic diagram of an actual control block according to another embodiment of the present invention. &lt; Explanation of component symbols &gt; (1) m input device (2) data buffering (3) sentence judgment (4) generating titles (5) generating continuous (6) storage media (7) signal comparison (8) voice input (9) Signal coding (10) Comparison result (11) According to word boundary rules (12) _ _ · General comparison (13) Continuous plant one by one (14) Appellation superposition (15) Database (16) According to continuous Factory (17) 2 continuous zone decomposition processing units superimposed data "three" sound data processor device output is decomposed program program "sound comparison program comparison program three" sound word number classification program "three" sound 569182 simple diagram Explanation (18) 3 consecutive "three" sounds (19) 4 consecutive "three" sounds (20) 5 consecutive "three" sounds (21) 6 consecutive "three" sounds (22) 7 consecutive factory three " (23), (24), (25), (26), (27), (29) Compare data softly Compare data (30) Voice signal (31) Voice comparison (32) The comparison result is output ( 40) &gt; (47) User calls (41), (48) Automatically answer And announce the welcome speech (42) &gt; (49) users say who they are looking for (43), (50) general comparison (44) continuous "three" sound comparison (45), (52) speech reading ratio The results are processed (46), (53) End (51) Appellation overlap (28)

Page 16

Claims (1)

569182 The method for identifying tone changes includes the following steps of inputting Chinese data into the data and a Chinese speech recognition system: using a keyboard or other input devices to punch the area; processing the Chinese sound of the data buffer or the title superposition; 'generating the title Reduplicated data, the vocalist 'produces continuous "three 乂 use sentence to determine the decomposition of processing units, materials, and distinguish the continuous according to the rules of the word field" Third, the Chinese word for the reduplication and stored in storage media; The continuous sound data in the data is stored in the storage medium so that the storage medium is connected to the signal comparison processor. The Chinese speech data of the signal-input speech input; Use k to compare the processor, according to the Chinese new data The media data is stored and the voice T is identified; the comparison result is output. 2. The method described in the first item of the patent scope, in which the word judgment second solution processing unit decomposes the input according to the word boundary rules. Chinese materials: for-general comparison, continuous "three" sound comparison and title overlap comparison 3 patent scope first The method of claim 2, wherein the processing solution words blade unit may be generally aligned in the build data to form the data storage media library. 4. ΐ :::! The method described in item 1 or 2 of the patent scope, wherein the word processing unit can classify the continuous "three" sounds into 2 569182 Six, the scope of patent application is a continuous "three" sound, three consecutive "" ears, five consecutive "three tones, two" chirps, four consecutive "three" three "tones, etc. different words =: three, 7 words to establish "two" sound comparison information [,; two "voices to it"] then to the first and second words, "to 3 consecutive" three "sounds 4 consecutive" three "sounds The sound comparison data of the “Fu order” ρ, the comparison data of f, and the five consecutive “three ::: words” “two” f, the four words of the “two” sound comparison. The "1" and "2" sounds are compared to the 2q, "quote", and for the 6 consecutive plant data, the "2nd" sounds are compared: each of the "2" sounds is compared to the 2nd, 3, and 5 ★ I :; :: Built in Jiancun Media to form the second; After the voice change declaration patent scope i or 2 item 2 decomposition processing unit can input the Chinese = 2 of the two words in the ::: The second word in the: to establish soft comparison data 2 Titles are built on storage media to form a database. For the ratio of patent application scope i 7 including hard disk memory. With COSCO storage media = the method described in item 丄 or 2 of the scope of the patent application, the H !! ί: device and the speech turn based on the information of the storage media to identify the data, you can perform general speech first By comparison, ί ί: When t loses, it is identified by the string relationship and word boundary, and the converted ^ is used for identification. 1 Watt Heart Page 18