TW526466B - Encoding and voice integration method of phoneme - Google Patents

Abstract

The invention provides an encoding and voice integration method of phoneme and employs an off-line method to sample voice. The sampled voice data can be classified into three phonemes in terms of voiced, unvoiced and mute phonemes that are further encoded in accordance with a pitch parameter, an amplitude parameter and a spectrum parameter. Those unvoiced phonemes are directly recorded, while recording the time period of mute phoneme, and then those encoded phoneme data is recorded in a voice database. Subsequently, as long as the encoded phoneme data in voice database is decoded and synthesized, the voice can be recovered. It is able to synthesize a voice close to the original one by synthesizing the retrieved voiced phoneme just via a voice synthesizer designed on the basis of the pitch parameter, amplitude parameter and spectrum parameter, directly receiving the unvoiced phoneme, and playing the mute length of the mute phoneme.

Description

526466

[Application Field of the Invention] [Background of the Invention]

At, in the mid-to-low-end electronic dictionary market, advertised the merits of being pronounced by real people, which has become the main feature of electronic dictionaries. In order to be competitive in the market, each manufacturer must: Reduce production costs. Some commercial firms; = ":: two 'Because of its large amount of data' and the type of system output is very expensive and expensive, so most manufacturers stand close to real people's pronunciation, which allows electronic dictionaries to This technique of speech analysis and synthesis is based on a certain processing method, which divides = speech signals and proposes necessary characteristic parameters, and uses these parameters to press "', ma to generate the branch Type synthesis technology for speech. Due to the combination of speech analysis, the process is to represent the original letter with the least digital data in the sound signal; Therefore, it is also commonly referred to as speech compression technology, which involves speech sampling and encoding and decoding technologies. For example, the adaptive differential pulse code modulation (AdaPtlve Delta pulse code; ADPCM) coding method in speech waveform coding focuses on making the reconstructed signal more similar to the original signal waveform. From a mathematical point of view, , Which adopts the Minimum Mean Square Error Criterion, but the bit rate of the ADPCM method is less than 24kbps (Kll〇Bit per Sec〇nd),

Page 4 526466 V. Description of the invention (2) There will be problems that the quality after reduction is reduced and the amount of calculation is large. The above-mentioned σ 3 knife analysis and synthesis has the characteristics that it can greatly reduce the amount of voice data, and can also have the advantage of confidential communication (using encryption technology). However, its shortcomings are the importance of speech synthesis, the partial separation, and the base period, which are often at a distance from natural speech, resulting in unnatural and even difficult to identify shortcomings. Even with compressed speech analysis and synthesis technology, there is still the possibility of saving memory space. In addition, '% of speech analysis and synthesis technologies mostly operate in a line f (on: line) mode, so' must be added to determine that the speech is a motion with sound 丄 'often in the process of judgment, will "have:" Misjudgment with the "No Tower" part resulted in hoarseness in speech synthesis. Therefore, how to make the analysis of speech close to natural speech, and also how to achieve the maximum compression range f; on the other hand, how to make the vocabulary; the above points have become important research [the purpose of the invention and Overview] The speech produced by the synthesis technology is improved in sound quality; the other is that the process of analyzing and synthesizing the least amount of memory is relatively simple. In view of the coding and speech synthesis methods of the above-mentioned conventional techniques, in the case of Ai bH-line), the processes are processed separately in advance and without vowels. The present invention provides a phoneme for speech. The purpose is to distinguish phonemes of speech into vowels offline, which can simplify speech synthesis during speech synthesis. Encode voiced phonemes to calculate amplitude, base period, and spectral parameters.

Page 5 526466

-I V. Description of the invention (3) The number of codes is encoded, in which the spectral parameters are encoded in the Lpc parameter; and the original phoneme file of the unvoiced (unvoiced) voice is not contracted; the mute part is Only the mute length is recorded. When decompressing, you only need to use the == part, and use the interpolation method to flatten the amplitude, base period, and spectrum parameters, and then use the speech synthesizer to restore the voiced voice; the silent part 8: It needs to be taken out based on the address The original voice is restored; and the mute part ',,': long silence time can be taken. According to the technology disclosed in the present invention, the present invention provides a method for encoding and synthesizing a speech, which includes :: * The establishment phase and the speech synthesis phase are explained as follows: ... 枓 Building a library to build a library, including the following steps: The & knife is a blocky, silent and silent ~ prime code 'and the voiceless element is-address code = = Store the silent and mute phonemes in the voice data. Once the user enters text data, Basu and read the phonetic data of the voice database, f: the sound phase of the text data. The receiver will enter the next stage of synthesizing the five vowels. According to the note of the text data of the ancient and ugly, the phoneme data on the next day will be synthesized into the phoneme code, the non-voice code, and the : Far have said the phoneme data, and synthesized according to the speech synthesizer - sound mute language according to the tone; mn: code generation - a silent voice, and Page 6526466

V. Description of the invention (4) During the establishment of the speech database, the amplitudes of the numbers are not equal to 夂 r, and the other numbers are encoded based on the base cycle number, unit number, and address parameters. Jing Li Li = Ear Day / , It is encoded according to the parameters between base cycles. Said prime is based on the base period parameter and the time sound $; 2: f: 'As long as the speech coding rules are used, the language: eight f ^ has a towering voice code, a silent voice code, and a mute voice code. Decode and synthesize to get a synthesized speech. In the bean, the speech is presented by a speech synthesizer, which is designed according to the three parameters of the base period parameter, the frequency and the amplitude parameter. 2. About the features and implementation of the present invention, the best embodiment will be described in detail with reference to the drawings. [Detailed description of the invention] If it is based on pronunciation, most languages are multi-syllable languages. Taking English as an example, if the English is subdivided into different monosyllables composed of different phonetic symbols, the English can be summarized into thousands of basic pronunciation units. These turtle a units are phonemes, and each different The phoneme itself contains its Pitch. Therefore, it is possible to use this phoneme-based pronunciation of 'languages', which in turn uses phonemes as the encoding and decoding of the speech, and the present invention is an application based on this concept. Secondly, because the electronic dictionary market has relatively regular speech processing and requires a large amount of data compression, the present invention uses Linear Predictive Coding (hereinafter referred to as LPC) as the encoding and decoding method of the present invention. the way. In addition, Linear Prediction Coding (LPC)

526466 V. Description of the invention (5) T-ct Pllter) ^, from the element rate to the purpose of compression, it can achieve a very low bit method. u Rate), so it is quite suitable as the coding code and phrase of the present invention: 成 成; :: Picture 1 "" The compilation of the phoneme of the present invention is silent and silent; i: (the second step of the picture into the package = of There are phoneme codes, no phonemes, and mute sounds: "a phoneme decoding and smoothing (step 40); and, synthesized speech = will. Among them, from the above encoding and decoding process, ^ including 4 05 4 The Vr stage (step 1.) and the decoding stage (step encoding stage focuses on the construction of the voice database, so it can also be called the voice database creation stage ~ :: In the electronic dictionary, the user presses what he wants To pronounce: text, :: The dictionary can be based on the five literary rectification j 乂 子 日 守 电子 素, and according to the encoding rules of the present invention to extract the place: "Decoding" and then restore and synthesize speech, so, cut :; two prime Re-strength: In the following stages, the individual steps will be explained one by one, which can also be referred to as the first step. In step 1Q, because of the seventy-five, the phoneme (ph0neme) is distinguished, and the _ ^ The pronunciation part of the text '· So' The present invention uses the phonemes Said "have also also: voiced", "unvoice" (unv〇jced) disc static stand method. Because "voiced" phonemes are cycles: the difference is to make the basic sub-parts', so you can go one step further. Compression; ❿ = 2 speech ..., "" prime is aperiodic

Page 8 526466 V. Description of the invention (6) —_ The speech part of nature (n〇nperi〇 (iic), so Γ; Ding, you can directly record its length. Compression; mute the English pronunciation in the electronic dictionary It is equipped with 彳 丨 标 (Phnetlc aiphabet), and the rules for distinguishing two syllables in the unit of each letter and syllable are syllables. That is, you can use the two …, The voiced or unvoiced voices of the ears. For example, "Jiebubei is added to areas S, t, etc. For example, the part of" f · Γ / "of the free sound production is ί, P, lreei9feLfri]. The same is true for speech processing in Mandarin and other languages :.]. As for the information in the language itself, you can use the voice I a μ + and 'under the condition of spring (of fl me)' to process it in advance, = Ear Axi towered before leaving the yard, and set all α α I% also 17. In the speech editing, the processing of "voice" phonemes: There are no strict ears. It only leaves the finals with sound .... Consonant vowel, consonant vowel sound of vowel and syllable phoneme Noise) are all set to zero, only 7 long recording silence. However, after this classifies the phonemes of the speech, you can enter the step code. Since the λ cow and donkey of the present invention λ 士 μμ 素 prime vowel editing "and" without standing! ", The phonemes of speech are classified into" there is a kind of pre-classified language stand Γ2 2 ". Therefore, the present invention will be directed to Three of the three voice coding :: husband;: 乂 coding. The coding method of the present invention is the average of the amplitude parameter = number of force two, the three parameters of 'code' are: (Pitch, that is n / 〇f mean _are ), Base period reflectl ^ 9 tuning) parameters and spectrum parameters (KC, S; reflection coefficient, reUection c〇efficients).

526466 V. Description of the invention (7) Among them, the amplitude parameter and the base period parameter of /,. Are obtained by a vertical handle (a sound frame frame 180 sampling points, 8kη, said busy cattle outer tube 屮 Α 夂) Product sampling rate / ^ kHz sampling rate) as the unit, and gradually read out the value of 50 pounds. And the frequency unit number (Rc, s) is obtained by calculating ^, that is, according to the following equation b · PC's A0 / (l + alZ-l + a2Z-2-+ alOZ-l〇) ^ ^ where , A0 is the amplitude parameter, z is' al ~ al〇 which is [π 夂 数 0 >

From the above three parameters, a "voice" speech frame (18〇SampleS) can be encoded as 54 bits, and the compressed bit rate is equivalent to 2 · 4 k b p s. The bit configuration of each parameter is as follows:

Pitch (6 bits), RMS (6 bits), RC, s (RC0 ~ RC9) 665555444433 As for the "silent" voice frame, since it is directly recorded by the present invention, the Pi tch parameter is defined The value is 1, which is encoded as follows:

Pitch (6 bits) Index_of_, υηνο iced__speech 6 8 (I d x)

6 8 (Idx) Among them, I dx is the index of actual speech (qi sound), that is, the address stored in it. "Mute" voice frame, set its base period parameter value to 0, and its encoding method is as follows:

Page 10 526466 V. Description of the invention (8)

Pitch (6 bits) Length_of_silence 68 (Ls) where Ls is the length of silence. 6 8 (Ls) Next, you can record the encoded audio data to Renlici's database, also π, step 30. The above step 10 has been said; the coding rules of Yuebei, that is, using the three parts of "voice &" one sound "and" mute "of the phoneme itself, are applied in different ways: code: if there is no such Can save considerable memory space. This established voice database can be used as the basis for five people. That is, when reading the voice data, start with the basic weekly reference :: 口 贝 取 ”, that is, if Pltch > 1, then read a total of 54 Original voiced speech; sPltch = 1, then read 8 bitsd back data I dx to load actual voice breath data, taking English as an example, f breath data account for about ^ lmbytes; state # ^ bA (Ls) , The decoding is restored to mute, length L㈣. The material is taken 8 "Right words, because the technical strategy adopted by the present invention is to separate the one, two, silent" and "silent" parts of the voice, so the data types of: and wide-coded are different. 'Various sounds as described above: 彳 ^ 彳 置. Therefore, when synthesizing speech, it is only necessary to compose i, i, i, and i according to the present invention. In the following, the selection method of the synthesis phase will be introduced, that is, steps 40 ~ 50.

Baixian, first introduce the part of phoneme decoding and smoothing, that is, I

Page 11 526466 V. Description of Invention (9) Step 40. In step 40, '' is also processed separately according to the three phonemes. For the first test of "voice" phoneme, please refer to "Figure 2". The clean sound synthesis block diagram of the present invention is 100 square meters. In speech synthesis, first, according to the text data typed by the user, and extract the appropriate phonemes according to the phoneme disassembly rule, the method is to first use a pulse sequence generator that can generate a "base period" with a phoneme ( Impulse Train; EKitation Signal) 101; Next, pass it through a channel filter (v〇cal Tract Fi Iter) 102, the frequency response of this channel filter i〇2 is determined by the value of RC, s; then, according to RMS The value is adjusted via the multiplier 103 to output the speech energy. Among them, the pulse sequence generator 1 〇1 is to simulate the vibration of the vocal vocal cord. Please refer to “Figure 3”. The method of generating the pulse sequence generator is p [25] = {8, -16, 26, -48,86, -162 , 294, -502,718, -728, 184, 672, -610, -672, 184, 728, 718, 502, 294, 162, 86, 48, 26, 1 6, 8} form a periodic sequence e (η ), The period is the pitch parameter. If Pitch > 25, then e (n) di {p [l], p [2], ..., p [25], 0, ..., 〇}; if Pitch < = 25, then e (n) di {p [1], P [2], ···, p [Pitch]}. Then e (n) passes a L0Wpass Filter (1 + 0 · 7 5 z-1 + 0 · 1 2 5 z-2) to obtain the input excitation signal (Excitation Signal) of the channel filter. As for the channel filter 102, which is the frequency response of the simulated oral channel, the filter parameters are the spectral parameters calculated according to the LPC method, and RC's' can be achieved through the channel filter. The input signal is e (n) 'and the output is speech s (n). Due to the LPC process, a pre-emphasis (1 — 0.9875z —D) is used during encoding, which is used to add

Page 12 526466 V. Description of the invention (ίο) Correct operation of strong high-frequency signals, so when decoding, a De-emphasis Filter W (; l-〇. 98 75z-l) needs to be added. In the multiplier of "Figure 2", the gain value (Gain) is added, that is, the RMS value of the decoded speech signal is multiplied by the decoded value of the channel filter 102, which is also the above-mentioned amplitude. Parameters, adjust it to AND, where: 14 "

Gam = PMS

In addition, in speech synthesis with vowels, it is necessary to add the base cycle: doubling. The method of synchronization is in speech synthesis, with one C ί: two: into several consecutive cycles', the length of the synthesized speech must be. Box f number of sample points and the length of the sound box (18.) The previous synthesized sound box is left with ‘strict: ^ left’ — the number of sample points that are less than the total number of sample points.

Said =. As shown in Figure 3 of the "Four Brothers", the length of a sound box at a sampling rate of 8 thousand per second is approximately 〖8 η. ^ ^ 々1 J is full of 180 points, and the remaining number Γ is not Γ. 〇 The next cycle continues, and so on. After it will be =, it enters the second stage of step 40, smoothing processing, that is, the base cycle and the amplitude seat p r. Will go to 1 ^ one RC number smoothing. The parameter is smoothed by the internal difference method. Its Φ person a, &: 50% of the number = the previous frame parameter -Prop) + head = said frame number 邛 Γ〇ρ. Where '〇 < = prOp (pr〇p〇rti⑽; h. ≪ = 1, Pαρ = number of sample points synthesized in the current frame / total number of sample points in the current frame.

526466 V. Description of the invention (11) For the encoding process of phonons, a clear description to describe the synthesis process. So, the above is a comprehensive introduction to the comparison of phoneme synthesis. In the following, it will be directed to three different processes. Please refer to "Figure 4". This issue: No. 2: The flow of synthesized speech passes through this flowchart. You will be able to change the flow chart of the 9-9 prime noon code, in order to tune the audio data: Steps 40 and 5. Specific operations. The code adopts the pitch parameter and is compiled in the resource f. “Because of the editing of the present invention, the“ sound ”base period parameter is calculated based on the calculation of the shell. , "Mute" base period parameter: "Ran Rong" base period data is judged to be "youshuang", "," can be base period parameter data, and processed separately. "Silent" or "mute" information First enter the 6th position (step 401) to judge whether it is silent or silent. If its $ ,, and its material are eared "and" voiced "phonemes, then take 1 (step 40 2) 'then it must be t8 bit data of amplitude parameter (m) and frequency, Also "voiced" voice: In steps of 4 ° 9), the coded must be "Standing, head, including λδ results, base period = 0 (step 4〇3), then its degree, and produce : L "S * 8, two digits: (step 404), to read the mute long 1 without ^^ treatment day. (Step 40 7); if the base period is not greater than 4 05), the templet parameter must be 1, then read in 8 bits (step 8 "(:, search for the storage address of Qiyin, According to the database, read the rest point of the qi sound (step 406). At last, the voice of Li Konglai @ 「有 声」, r ... output 5 " (step), restore the original sound, and "mute" Part of

526466

Please continue to refer to "Figure 5", the ^ mile chart of the voice combination of the present invention. Through this figure, the "prestige. The other ear" phonemes are combined in the "voice" data. It occupies about 54 bits, following the process of g, human. First, in step 4 1 1, first read in the first frame P parameter is a composite number '. Then, in step 4 1 2, let N: 0, L = 180, base cycle 0 = base cycle RMSO ^ O, RCOi two RCi, i two 0, 1, ..., 9 _ to read the RC parameters, and then, you can perform parameter smoothing to improve the sound quality, this is step 4 1 3, as shown below: prop Two N / L; base period, base period 0 * (l-prop) + base period * prop RMSj = RMS0 * (1-prop) + RMS * pr〇p; RCj (i) = RCO (i) * (l -prop) + RC (i) * pr op i = 0, 1, ···, 9, where prop is the proportion (P r ο porti ο n), and L is the size of the sound frame. At the beginning, L · = 1 8 0. _ Next, if N + base period j > L (step 41 4), that is, after taking the length greater than one frame, read the next frame again, that is, proceed to step 415: let L = L-NH80 N = 0

Page 15 526466

Page 16 526466 V. Description of the invention (14) The memory capacity of the prime file is repeated, # (~ 2.4kbPS), can be a large amount =, 2 "bytes (bytes) or less is sampled into 16 bits, decompressed ^ 1 5 Recall the body space and improve the poor quality of the connected speech. If you use smoothing, you can improve the individual processing of the speech. The person with sound a 'because the encoding method will produce the sound and sound without sound. General speech coding problems; silent parts, conditions, causing hoarseness and other fruits. Side tone original sounds to maintain the best tone. Although the present invention is used to limit the date and month #, a preferred embodiment is disclosed. As above, within the spirit and scope of iron, when it can be =, whichever does not depart from it. It is necessary to regard the scope of the patent application attached to this specification as the mother part of the sound and silence, the bad sound effect is not the present invention Defined by invention

526466 Brief description of the diagram. Figure 1 is a flowchart of the method for encoding and synthesizing speech phonemes of the present invention; Figure 2 is a block diagram of the speech synthesizer of the present invention; Figure 3 is a simulated vocal cord vibration diagram of the present invention; Figure 4 is the flowchart of the phoneme decoding of the present invention; Figure 5 is the flowchart of the signal processing of the speech synthesizer of the present invention; Figure 6A is the original voice waveform of the single word `` abbreviation "; Figure 6B is the single word nabbreviati ο ηπ speech waveform using encoding and speech synthesis method of the present invention; FIG. 6C is a single word nabbre ν iati ο ππ speech waveform of general encoding and speech synthesis method; FIG. 7A is a spectrum chart of FIG. 6A; The figure is a spectrum diagram of FIG. 6B; and FIG. 7C is a spectrum diagram of FIG. 6C. [Symbol description] 100 Speech synthesizer 101 Pulse sequence generator 102 Channel filter 103 Multiplier step 1 0 Differentiate voiced, silent and mute phonemes Step 2 0 Perform phoneme encoding Step 3 0 Store the encoded phoneme code Step 4 0 Decoding and Smoothing Phonemes

Page 18 526466 Schematic description step 5 0 step 401 step 4 0 2 step 403 step 4 0 4 step 4 0 5 step 4 0 6 step 4 0 7 step 4 0 8 step 4 9 step 4 10 step 41 1 step 4 1 2 Step 4 1 3 Synthetic speech reading 6-bit base period> 1 base period = 0 8-bit reading 8-bit reading Ls * 8 points based on database sound sample points 4 8-bit speech processed by the speech synthesizer. Read the first speech frame parameter. Let N = 0, L = 180, base cycle 〇 = base cycle; RMSO = 0, RCO i = RC i, i 20, 1 , ..., 9 prop = N / L; base period j = base period 〇 * (l-pr〇p) + base period 氺 prop RMSj = RMS0 * (1-pr〇p) + RMS * pr〇p; RCj ( i) = RC〇 (i) 氺 (l-pr〇p) + RC (i) 氺 prop i 2 0,1, ..., 9 N + basic period j > L Let 1 ^ 叶 -N + 180 N = 0 Base Week 0 = Base Week RMSO-RMS, «Step 41 4 Step 4 1 5

Page 526466

Page 20

Claims (1)

526466 6. Scope of patent application ---------- Coding and speech synthesis method, • Offline encoding of ^ and speech synthesis of speech phonemes in the language, including the following steps: Establish a speech database, including the following steps: the speech; the speech phoneme is divided into-voiced, silent and mute sound: the earphone compression coding, and the speechless phoneme: address coding and the mute phoneme A time-length encoding; static stand 压 3: press and% encode the phoneme and store the silent and comment phonemes into the voice database; when the user presses the person-under-voice and reads the phonetic database analysis = The sound of the text data is based on the speech "2 breeding materials, and it is twisted. It carries the a prime data of I and the library, and synthesizes the text data I. It includes the following steps: τ is the sound of the phoneme data Li A ^ and the mute phoneme code; and the ear elementary stone horse "the phoneme code is based on the phoneme code of the phoneme data via-& Liren 2 :: -voice sound" and according to the voice data; J sound ... Based on Rong mute phoneme code generation—quiet • Please call for the coding of speech phonemes and linguist method described in item i of the patent scope, where the sampling rate for sampling the language is 8 thousand per second-due to D, as requested The combination of speech phonemes described in item 1 of the patent scope Page 526466 —— VI. Method of patent application f Method 'The compression coding of vowels is based on-basic weekly reference ^ Vibration field number and a spectrum parameter to encode; the second vowelless ==, The flat code is encoded by the far base period parameter and the one-bit address parameter; the time length encoding of the phoneme is encoded by the base period parameter. > Requirement 4. The base period parameter and the amplitude parameter of the phoneme in the speech phoneme described in item 3 of the scope of the patent application and the phoneme in the speech combination are calculated in units of: frame, and the parameter value is calculated step by step. . 5 · = Declares that the coding of speech phonemes and speech combination '.' As described in item 3 of the patent scope, and the encoding of mid-frequency reading parameters are coded using a linear predictive coding (6 Coding, LPC) method. Synthesize the speech phoneme encoding and speech Γ L described in item 1 or 3, where the address parameter is the storage address of the sounded phoneme of the sampled speech. 7 · ίΓ = encoding and speech of the phoneme of speech as described in item 1 or 3 of the circle. The time parameter in 1 /, is to record the sampled δ of the speech; the application time is long and wide. The method, 1 in the ^, item 1 of the phoneme coding and speech synthesis 9 :::: prime value is defined as i 'the synthesizer ί MG ί 1 or 3 of the speech phoneme encoding and speech parameters and The synthesis of frequency speech is based on the parameters of the base cycle, and the amplitude system includes :, π M to synthesize the voiced speech, where the speech synthesizer M, page 22, 526466 6. Patent application scope-a pulse sequence generator for The week parameter output is an excitation signal; a channel filter is used as a filter parameter of the channel filter according to the spectral parameter to receive the excitation signal and output it as a voice signal; and The multiplier is used to multiply the voice signal by the amplitude parameter to output a restored voice. 10. The method for encoding and synthesizing speech phonemes as described in item 1 or 3 of the scope of patent application, wherein the generation of the silent speech is based on reading a silent speech phoneme in the speech database according to the address parameter, and The silent speech phoneme generates the silent speech. Π. The method for encoding and synthesizing speech phonemes as described in item 1 or 3 of the scope of the patent application, wherein the generation of the mute speech is based on the time parameter to output a silent sound with an amplitude of 0 that corresponds to the time length of the time parameter Page 23