KR950011485B1 - Sounding managenent system - Google Patents

Abstract

The device endows the unlimited speech synthesis function for the audiotex system to provide the information quickly to the end user. The device comprises: a CPU(10) which controls the memory, peripherals, and address mapping logic; a DSP(30) which processes the LSP parameter to produce natural synthesized speech and generates the synthesized speech into analog form; channel link(20) which converts the synthesized analog speech into PCM data format, transmits the PCM converted synthesized speech to the subscriber, and disconnects the channel by sensing the hook-on signal; and a DTMF generator(60) which generates the sixteen DTMF signals.

Description

Speech processing device and method of audiotex device for converting Korean characters into speech

1 is an overall configuration diagram of an audiotex device to which the present invention is applied.

2 is a hardware configuration diagram of the present invention.

3 is the overall functional diagram of the software.

4 is a process flowchart of converting Korean characters to speech according to the present invention.

5 is a display diagram showing the types of half-syllables.

6 is a diagram showing a smoothing rule required when combining syllables.

7 shows an example of a smoothing model required for half-syllable combination.

8 shows an example of a syllable database and an index table.

9 is a main control part flow chart.

10 is an LSP transmission flow diagram.

11 is a duration control flow chart.

12 is a half-syllable database access code generation flow chart.

FIG. 13 is a flow chart of connection of half syllable.

14 is a fundamental frequency control flow diagram.

* Explanation of symbols for main parts of the drawings

10: central processing unit 20: channel connection unit

30: DSP unit 40: serial input / output unit

50: memory

60: DTMF receiver

The present invention relates to a speech processing apparatus and method for a Korean text-to-speech audiotex system that provides a subscriber with the necessary information by converting into a natural voice using information composed of text (database).

The voice information system currently in service has a recording / playback method (ADPCM method) that simply converts a note into digital and then plays back digital signals as voice, and a limited voice word. In a large way.

However, since the current voice information system uses limited voice words, it provides various simple information such as traffic guide, tourism guide, agricultural and fish price guide, real estate guide, and test pass guide, but it is quite insufficient in the content and quantity of information. Not quick to update

Therefore, the present invention devised to solve the problems of the prior art, the audio processing system of the audiotex system that can deliver a considerable amount of information in the content and quantity quickly to the subscriber to have an unlimited voice synthesis function And to provide a method.

In order to achieve the above object, the voice processing apparatus according to the present invention comprises a central processing means, a channel connecting means, a DSP means, a serial input / output means, a memory means, and a DTMF receiving means.

In addition, the speech processing method applied to the speech processing apparatus is generated by receiving Korean text data, classifying Korean letters, alphabets, abbreviations, and numbers, generating internal codes including phonological symbols, and generating information for intonation and breathing. A text code processing step of converting an internal code into a phonetic code sequence using Korean phonological rules, and reading the phonetic code string, syntax information, and input text number generated in the text internal code processing step from an memory stored in an array form, and each syllable Code parameter processing to generate the semi-syllable synthetic parameter string and store it in memory after applying the phoneme duration control rule, semi-syllable connection rule, and basic frequency control rule by using phonetic symbols and syntax information data And the syllable LSP parameter generated in the chord parameter processing step. And a parameter speech processing step of converting the analog speech to the analog speech through the digital analog converter while producing the digital speech.

The new services that the Audiotex system can provide include an unparalleled number of information retrieval services for subscribers by phone number, all information retrieval services for vehicle owners by car number, newspaper reading service, and stock market breaking service. There are many.

1 is a basic configuration of an audiotex system to which the present invention is applied, in which 1 is a subscriber block, 2 is a voice processing block, 3 is a service control block, 4 is a data network connection block, and 5 is an audiotex database. Represent each.

The subscriber block 1 of FIG. 1 is in charge of the T1 trunk interface to form a call path between the subscriber and the audiotex system and to perform the calling number inquiry function. The trunk interface of the TDX-10 developed by the Electronics and Telecommunications Research Institute has been modified and implemented.

1) 96 channels separated from four T1 trunks, 2) Signal control with the exchanger by R2 signal, 3) Calling party number inquiry, 4) Ring and hook signals are sent to voice processing unit 2 for each channel 5) Receive the hook signal sent from the voice processing unit 2, 6) Send a codec initialization command to the voice processing unit 2.

The present invention speech processing unit 2 receives the desired sentence information from the database 5 of the information provider and then performs a speech synthesis algorithm to generate a synthesized sound.

LSP rule synthesis using half-syllable is used as the compositing technique, and the real-time operating system is always operating to process all tasks in real time. The speech synthesis algorithm, the control program and the half-syllable data are fixed in the ROM and perform the speech synthesis function upon receiving the ring signal from the subscriber block. The functions are listed as follows.

1) Connect subscriber block (1) with Cortec, 2) Receive and translate DTMF signal, 3) Connect central processing unit and DSP using FIFO, 4) Semi-syllable data and exception data retrieval and speech synthesis algorithm Performance, 5) modifying and monitoring the program's variables and parameters in the console, and 6) changing the sentence to analog voice.

The service control block 3 is responsible for providing and managing services for 96 channels, processing input commands of telephone subscribers, and managing input / output devices of the information provider's database 5. The service provision and management function interconnects telephone and packet calls, monitors the status of each channel, performs function tests and handling of errors, and generates and manages billing data. The subscriber command processing function sends the initial guidance message and the corresponding message in case of error to the voice processing block (2), manages the location table of the subscriber's identity, password, and information providing database, and records the data entered by the subscriber. Export to

The input / output device function connects the voice processing block through the serial port and the data network connection block through the rule buffer.

The data network access block (4) performs X.25 call setup, connection and disconnection functions to connect the information database through the public data network and manages the X.25 protocol. The block consists of a monitoring unit that monitors the actual line, a timer unit that manages a freely running clock, an exception unit that handles interrupts, and a transmission unit that handles packet data streams with packet links. The maximum transmission rate is 64Kbps.

The audiotex database (5) should be designed to store, edit, and modify text information quickly and easily, and the communication protocol for communicating with audiotex devices and the general protocol for rapid speech synthesis of audiotex devices. This must be observed.

The general protocol is as follows.

When editing a database, do not exceed 80 characters per line.

Each line ends with a composition point or ends with a word unit.

-Use short sentences whenever possible for clarity.

-Design the tree structure to have 4-5 menus for each stage so that the user can easily select the menu.

The communication protocol consists of ENQ (Ox05), XON (Ox11), XOFF (Ox13) and CAN (Ox18) signals, which are simply one byte characters. When the audiotex device sends out the ENQ signal, the database 5 starts the operation, sends out welcome and guide messages, receives the data pressed by the user, and sends the data to the audiotex device.

Like other serial communication, XON and XOFF signals are used as control signals to solve the problem of sending and receiving data. When the audiotex device detects a hook-on signal or detects an error signal and sends a CAN signal to the database (5), the database (5) stops all operations and returns to the initial state. If it is determined that the user no longer uses the database (5), the operation is stopped and the CAN signal is sent so that the audiotex device cuts the channel.

2 is a block diagram of the audio processing unit 2 according to the present invention. As shown in the drawing, the central processing unit 10, the channel connection unit 20, the DSP unit 30, the serial input / output unit 40, the memory unit 50, It is comprised by the DTMF receiver 60.

The central processing unit 10 controls memory, peripheral devices, and address mapping logic. A 32-bit data bus can drive all devices having 8-bit and 16-bit data sizes. The central processing unit 10 used the MC68020 MPU and operates at a speed of 20 MHz (MH _Z ).

The channel connection unit 20 converts the synthesized analog voice into a PCM (2,048MH _Z subhighway) and transmits it to the subscriber block, detects a ring signal and a hook-on signal, and sends a hck-on signal to the subscriber block. Induce cleavage. The channel connection unit 20 is configured using a TP3070 codec chip, a shift register, and a decoder.

The DSP unit 30 has a 25 ns ROM (4K long word) and a 25 ns ROM (32 K long word). The DSP unit 30 reads the LSP parameter sent from the central processing unit 10 to the FIFO from time to time to process the 12th LSP synthesis filter algorithm in real time. To produce a natural synthesized sound synchronized with the pitch. This synthesized waveform is sent to a 12-bit D / A converter by an internal timer interrupt that occurs every 0.1 ms to convert to analog voice.

The serial input / output unit 40 implements three serial input / output ports using the MFP (MC68901) and the SCC (Z8530), and the MFP monitors a watchdog timer and a console for detecting software tick timers and software bugs. Used as a port The status of tasks, buffer status and service progress is displayed on the console through the console port at 9600bps. The two serial ports of the SCC also provide a data path for text data commands to and from the service control block.

The memory unit 50 is composed of a nonvolatile ROM module and a volatile RAM module. The ROM module is composed of four banks in a 16-bit bus structure, and each bank is arranged in two ROM chips, 512 kilobytes, 1 megabyte, It is designed to have a capacity of 2 megabytes. Control programs, speech processing algorithms and LSP parameter data are stored in this ROM.

The RAM module is composed of two banks with 32-bit bus structure, and each bank is arranged with four RAM chips. Like RAM, the RAM module is designed to have a capacity of 512 kilobytes, 1 megabyte, and 2 megabytes. It is used for storage of variables, queues, flags and counters. The address of ROM module is from OFFE80000H to OFF7FFFH and the address of RAM module is assigned from 0 to OFFFFFH.

The DTMF receiver 60 uses 75T2090, a DTMF transmitter and receiver chip, detects all 16 DTMF signals generated from the telephone, and sends out only 4 bits of data to the 8-bit data bus. And 16 generate DTMF. Currently, it is used for DTMF reception. When the DTMF signal is detected and the automatic interrupt signal is generated at level 4, the interrupt routine of the central processing unit reads 4 bits of data and replaces it with the corresponding ASCII data. Are stored in.

3 is a diagram of the overall functional configuration of the software. The four main sections of the main control unit 70, the text code unit (TTC) 80, the code parameter unit (CTP) 90, and the parameter voice unit (PTS) 100 are shown. It consists of a module and nine submodules.

The main control unit 70 initializes a Mail box, generates a transmission queue, a reception queue, and a DTMF according to the main control flow of FIG. 9, and generates a ring flag and a hook-on flag. If the reactivation flag is reactivated by checking whether it is on, the port for the audiotex database is initialized and all variables used are initialized. In addition, the console displays a message indicating that it is starting, and then generates the specified priority to the text receiving task, the text sending task, the ring task, the DTMF task, the hook-on, the LSP task, and the text code task. Turn Rotate the loop to perform a fault check and display the results and the results from other tasks on the console.

The text code unit 80 converts the text data transmitted from the service control block 3 of FIG. 1 into an internal sounding code and produces phonological information.

The code parameter unit 90 first calculates the duration and duration of the syllable in the sentence with reference to the code sent from the text code processing unit 80. Pitch line lines are formed by rule by using boundary information of sentence, phrase and clause. After this process, consecutive sentences are made by combining syllables in syllable units. The parameter voice unit 100 of FIG. 3 performs a function of converting the parameter data sent from the text code unit 80 into a voice that can be actually heard, and employs a 12th order all-pole filter to make a high quality voice in real time. It is programmed in assembly language using DSP chips.

The text data receiving unit 71 receives the character information sent from the audiotex database and stores the character information in two receiving buffers alternately and receives a period (.,!,?) Finally, the EOS character is inserted to send the text of the reception buffer to the text code unit 80 while sending a command to operate the text code unit 80. If too much data is received and the receiving buffer is full, the XOFF character is sent to the audiotex database (5) to stop the data transmission. When the receiving buffer is empty, the XON character is transmitted to resume the data transmission.

The DTMF receiver 72 receives the DTMF data pressed by the user, converts the ASCII characters into ASCII characters, and then informs the corresponding key value by voice to confirm whether the key pressed by the user is correct. In addition, if DTMF data is checked, if it is a number (0-9), it is exported as it is to the audiotex database (5). If it is an asterisk (*), then DTMF data is classified and used for sentence repeat function, paragraph repeat function, sentence skipping, and high speed processing. Perform. If the DTMF key is pressed while the voice is flowing in the middle, the voice stops immediately and the above function is performed.

The command data transmitter 73 transmits the state or commands of the voice processing block 2, billing information, and other guide texts to the audiotex database 5. First, it checks whether there is a message to send and if there is a message, the first character is sent out and the remaining characters are stored in the transmission queue. Then, a transmission interrupt occurs automatically and the message is transmitted to the audiotex database (5).

The fault management unit 74 always monitors the state of the voice processing block 2 and collects the generated faults and informs the console.

The interrupt control unit 75 manages all interrupts generated by the voice processing block 2, and a text receiving interrupt routine generated when a text is received, a text sending interrupt routine generated when a text is sent, and the user is DTMF. It consists of a DTMF interrupt routine that occurs when a key is pressed, a hook-on interrupt routine that occurs when the user hangs up, a ring interrupt routine that occurs when the user makes a call, and a fault interrupt routine that occurs when a fault occurs.

The timer controller 76 generates a timer interrupt every 10ms so that all tasks can check the timeout time.

The hook-on control unit 77 detects the hook-on flag generated when the user hangs up the phone, turns off the operation indicator lamp, and sends out CAN characters to the audiotex database 5 to stop the operation of the database.

In addition, it removes all the tasks currently in operation, initializes the parameter voice unit, and makes the main control unit 70 the initial state.

The ring signal detection unit 78 of FIG. 3 detects a ring flag generated when a user makes a phone call, turns on an operation indicator lamp, and sends out a welcome message stored in the voice processing block 2 as a voice. Export the ENQ character to (5) to start the database operation.

The LSP transmitter 101 of FIG. 3 performs a function of transmitting LSP data to the parameter voice unit 100 according to the LSP transmission flowchart of FIG. 10. First, it waits for LSP data, and when data comes, it divides LSP data into frame units (16) and checks the FIFO status. If it is empty, it sends LSP data to LFI frame as FIFO.

Repeat the above process until all LSP data has been transmitted. After all the LSP data has been transmitted, the voice synthesis status is checked and if it is still in progress, it indicates that the LSP transmission is completed. If it is in the state of notifying the end of the voice synthesis, the end-of-transmission command is sent to FIFO No. 15, and the LSP transmission is completed. If it is, the contents of FIFO are erased and the end command is sent to FIFO No. 15 to inform the completion of LSP transmission.

4 is a flowchart illustrating a process of generating a synthesized sound by speech processing a text sentence. The process of generating the synthesized sound is executed by a program stored in the memory unit 50 of FIG. 2, the first step 110 of text code processing, the second step 120 of code parameter processing, and the third of parameter voice processing. It consists of step 13.

In the first step 110 of text code processing, the Hangul process 401 alphabet / abbreviation process 402 receives Korean text data and classifies Hangul, alphabets, abbreviations, and numbers to generate internal codes including phonological symbols. Morphological and syntax analysis (403) and rhyme and boundary analysis (404) generates information for intonation and breathing, and in Korean phonological rules only (405), the function of converting the generated internal codes into phonetic symbols using Korean phonological rules. In charge of.

The second step 120 of processing the code parameter reads the phonetic code sequence, the syntax information and the input text number generated in the text code processing step 110 from the memory stored in an array form, and the data size is the phonetic code sequence. Therefore, the syllables of each syllable occupy one byte each of the first consonant, the neutral, and the final consonant, and the syntactic information additionally occupies one byte. The variable that stores the number of input texts takes two bytes. The variable that stores the number of input tests takes two bytes. Using these data, we apply the semi-syllable synthesized parameter string that can be used in the third step (130) to apply the phoneme duration control rule, half syllable unit connection rule, and basic frequency control rule, and finally process the parameters. It is responsible for creating and storing it in memory. Finally, the data size stored in the memory occupies 2 bytes of pitch, 2 bytes of energy, and 24 bytes of LSP parameter.

In the second step 120 of processing the code parameter, the main control unit is defined and operated in the form of one task under the control. The mailbox is used to synchronize data input / output with the main control unit. It also uses a mailbox to synchronize with the task responsible for the LSP forwarding function.

Detailed description of the second step 120 of processing the code parameter of FIG. 4 is as follows.

The duration control function of the phoneme is performed according to the duration control flowchart of FIG.

First, the duration of syllable is increased by using boundary point information of words, phrases, phrases, and sentences for each phonetic symbol string. After the duration of each syllable is determined, the duration percentage of each phoneme is calculated by applying the rules according to phonological phenomenon, and finally the duration of the phoneme is finally determined using the minimum duration of the phoneme and the inherent duration of the phoneme. Calculate.

After the duration control of the phoneme is completed, a parameter generation function 407 for generating a code for accessing a half-syllable database is performed according to the flowchart of generating a half-syllable database approach to FIG. At this time, the half syllables beginning with / a, b, ㅂ, ㅈ, ㅎ / should be voiced mute if the preceding phoneme is not a non-negative voice or a vowel sound. do. In addition, a half-syllable that starts with / ㄹ / generates an access code using / la / mutated sounds or / ra / mutated sounds if the leading phone is / ㄹ /.

When the access code is made in this way, according to the connection diagram of the unit of syllable of Fig. 13, if the syllable is a left half syllable, it is classified whether there is a consonant first. Choose from 4 and 5 If the syllable's right syllable is the syllable's, it's classified whether or not there is a finality. In this way, when the syllable shape is determined, the synthesized parameter string is taken from the half-syllable database by the length of each phoneme, and then between vowels or voiced sounds, between unvoiced and unvoiced sounds, between vowels or voiced sounds and unvoiced sounds according to the half-syllable shape. The smoothing process is performed at a certain interval of.

When the parameter connection in the unit of the half syllable is completed, the basic frequency control 408 is performed according to the basic frequency control flowchart of FIG.

First, the basic outline of the fundamental frequency is created according to the sentence type (e.g. sentence, question and exclamation sentence), and the rising and falling values are determined using the boundary information of the phrase and the clause. And the final fundamental frequency contour is generated by calculating the change of fundamental frequency according to phonological environment and modulation method.

The third step 130 of processing the parametric voice reads the half-syllable LSP parameter generated in the second step 120 of processing the code parameter to produce 100 samples (10 ms) of digital voice and converts it into analog voice through a digital analog converter. A note performs a function.

The sound source signal generator 409 uses impulse strings for voiced sounds and white noises for unvoiced sounds from the pitch information of the LSP parameter, and connects the synthesized sounds by new half-syllable parameters every voice synthesis section (10ms). Linear interpolation is performed to control the discontinuities that occur. The LSP speech synthesis unit 410 uses an all-polling LSP filter algorithm, implemented using a 32-bit DSP chip for real-time implementation, and outputs a speech by outputting a digital speech sample every 100 microseconds by an internal timer. Make it up

5 is a display diagram showing the types of half-syllables, and the first type (TYPE) is not a half-syllable. Indicates a section where a person speaks and rests.

Type 2 is the first half syllable, which consists only of vowels without a leading. Type 3 is the first half syllable consisting of vowels and vowels. Type 4 A first syllable consisting of vocal cordless consonants and vowels. Type 5 is a first half syllable consisting of ruptured adolescent consonants and vowels. Type 6 is a posterior half syllable consisting only of vowels without bells. Type 7 is a posterior half syllable consisting of vowels and meteoric consonants. Type 8 is a posterior half syllable consisting of vowels and asexual consonants.

Type 9 is a posterior half syllable consisting of vowels and aplastic vocal consonants. FIG. 6 shows the smoothing rules required for combining half syllables, type P in the form of leading syllables, type in the form of gynecological syllables, and the values in the table shown in FIG. 7 according to the combination of leading syllables and current syllables. The form is shown.

In the smoothing mode of the half-syllable unit of FIG. 7, Form 1 is a method of linearly interpolating the interval between the rear Ta of the front half syllable and the front Tb of the rear half syllable. Linear interpolation between the beginning and the interval of the shape, Form 3 is the method of linear interpolation between Ta of the last half syllable at the end of the front half syllable, and Form 4 is the pure interpolation between the front half syllable and back syllable without linear interpolation. It shows how to connect.

At this time, the connection section Ta, Tb is a value determined by experiment, and each half-syllable is smoothed in any one of the above types 1,2,3,4 according to the connection type.

FIG. 8 is a block diagram of an index table having a half syllable database and a start point, a phoneme boundary point, and an end point of each half syllable.

The half-syllable database consists of 1 to 640 half-syllables in an internally determined half-syllable order. Each half-syllable consists of several frame units consisting of pitch, energy, and LSP parameter. The position of each half-syllable in the half-syllable database is adjusted to the number of contacts table. This table provides information about the starting point of each half-syllable, the boundary point of the phoneme in the half-syllable, and the end point of the half-syllable for each half-syllable. Have half-syllable order. For example, look at the picture:

If the half-syllable "A" is called half-syllable number 1, the index table has a position in the half-syllable half-syllable database of "A" half-syllable. As shown in Fig. 8, the starting point of the half-syllable number 1 is 0, The phoneme boundary a ', the end point a, can be seen that the position occupied by the syllable number 1 in the syllable database is 0 to a. The same rule applies to the remaining half syllables.

The present invention configured and performed as described above can obtain the following effects by applying to the audiotex service.

First, because it converts the text information to the phone so that all future information can be provided simply and quickly by phone.

Second, profitability of the fixed-line and telephony businesses will increase due to a surge in information retrieval via telephone.

Third, new job creation is expected due to the rapid increase in information retrieval by telephone and the increase of information providers.

Fourth, the establishment of the information society is naturally achieved through the diversification of information mind service and information mind recognition through information search and collection.

Fifth, welfare services for the blind can also be used.

Claims (3)

In a speech processing apparatus applied to an audiotex system for converting Korean characters into speech, a central processing means (10) for controlling a memory, peripheral elements, and address mapping logic, and the central processing means (10) are internal FIFO buffers. DSP means 30 which reads the sent LSP parameters from time to time and processes the 12th order LSP synthesis filter algorithm in real time to generate natural synthesized sound synchronized with the pitch, converts the synthesized waveform into analog voice by internal timer interrupt, and outputs it. The synthesized analog voice output from the DSP means 30 is converted into a PCM (2,048 MHz subhighway) and transmitted to the subscriber block, the ring signal and the hook-on signal are detected, and the hook-on signal is sent to the subscriber block. The channel connecting means 20 which induces the cutting of the Roy and all 16 DTMF signals generated by the telephone through the channel connecting means 20 are reduced. DTMF processing means 60 which sends only 4-bit data to the central processing means 10 on the 8-bit data bus and receives the 4-bit data exported by the central processing means 10 and generates 16 DTMF data. A nonvolatile ROM module for storing the serial I / O number 40 for relaying serial communication between the service control means and the central processing means 10 in the audiotex system, a control program, a voice processing algorithm, and LSP parameter data; And a memory means (50) comprising a volatile RAM module for storing buffers, variables, queues, flags, and counters used by the program. The voice processing device according to claim 1, wherein the serial input / output means (40) implements three serial input / output ports. An audiotex system including a central processing unit 10, a channel connection unit 20, a DSP unit 30, a serial input / output unit 40, a memory unit 50, and a DTMF receiving unit 60, and converts Korean characters into voice. In the speech processing method applied to the speech processing apparatus of the present invention, the Korean text data is received and classified into Korean, alphabet, abbreviation and numbers to generate an internal code including a phonological symbol, and to generate information for intonation and breathing. A text code processing step 110 for converting an internal code into a phonetic code sequence using a Korean phonological rule, and a phonetic code sequence generated in the text code processing step 110, syntax information, and input text heads are stored in an array form. After reading the syllables and the syntactic information data of each syllable, the phoneme duration control rule, half syllable unit connection rule, and basic frequency control rule are applied. A code parameter processing step 120 for generating a syllable unit synthesis parameter string and storing the result in a memory, and reads the half-syllable LSP parameter generated in the code parameter processing step 120 to generate a digital voice, and then performs an analog through a digital analog converter. Speech processing method characterized in that it is carried out with a parameter voice processing step (130) for converting to voice.