CN1755795B - Method of constructing digital voice database of Chinese characters, Chinese digital series synthesis system and method - Google Patents

Abstract

The disclosed method to construct Chinese digital pronunciation database comprises: generating an original pronunciation database contained a plurality of digital unit string opposite to Chinese digital pronunciation; selecting digital units from organic database for first, medium and tail of said Chinese string as binary group digital units; similarly, selecting single group digital units; cutting down binary digital units with weak influence degree of adjacent units; constructing objective pronunciation database with cut binary units and single units. This invention keeps high-natural synthesis language, and decreases database scale fit to embedded device with small memory.

Description

Method for constructing digital sound library and system and method for synthesizing Chinese digital strings

technical field

The present invention relates to a method for constructing a Chinese character digital sound library and a Chinese digital string synthesis system and method, in particular to a method for constructing a Chinese character digital sound library with high naturalness suitable for a small memory environment, and a method suitable for A system and method for synthesizing high-naturalness Chinese character and numeral strings in a small memory environment.

Background technique

With the development of software and hardware technologies, the computing power and storage capacity of current information products (such as PDAs and smart phones) have been greatly improved, which in turn makes the demand for improving user interfaces more intense. Text-to-Speech (TTS) technology is used to convert text input into human-audible synthetic speech output with certain intelligibility and naturalness. The computer system applying this technology can provide dynamically generated voice responses while greatly reducing the number of pre-recorded (Prerecorded) voice files, thus enabling information products to have a better human-computer interaction interface.

Digit string synthesis is an important application of Chinese text-to-speech conversion technology, such as providing phone number reading function for high-end mobile phones or telephone directory desks, reporting securities trading prices and industrial instrument readings, etc. Since the number strings in a sentence usually contain more information, it is easier to attract the attention of the listener. Objectively, the number string synthesis is required to have better sound quality, that is, higher naturalness and intelligibility.

Traditionally, depending on the application platform and occasion, the techniques for building a synthetic system are different. The unit splicing method based on a large-scale sound bank can provide high sound quality and speed, so it is suitable for server or desktop computing environments; while the synthesis method based on the channel model is more suitable for embedded applications due to lower resource requirements , but can only provide lower quality output speech at the same time. Therefore, there is a need for a high-naturalness Chinese digit string synthesis technology that can be applied to a small memory environment.

Contents of the invention

Therefore, the problem to be solved by the present invention is that the existing method for synthesizing Chinese numeral strings cannot be applied to mobile communication terminals. This is because the memory of the mobile device is relatively small, and the large-scale sound library applicable to the desktop computing environment cannot be directly transplanted. to the mobile communication terminal.

That is, the present invention cuts down the established original sound bank, while ensuring the quality of the synthesized speech, the size of the sound bank can be adapted to a small memory application environment.

In one aspect of the present invention, a kind of method of constructing the Chinese numeral sound bank is proposed, comprising the steps of: producing an original sound bank comprising a plurality of numeral unit strings, said numeral unit string being the representation of the pronunciation of the corresponding Chinese numeral string; From the original sound bank, the digital units in which are selected are respectively located at the beginning of the string, in the string and at the end of the string; from the original sound bank, the digital units are respectively located in the Chinese digital string. The one-tuple number unit of the string head, the string and the end of the string; such a two-tuple number unit is cut, and the degree of influence between adjacent number units in the two-tuple number unit is weak; by the two-tuple number unit after cutting The tuple number units and the selected one-tuple number units constitute the target sound bank.

Utilizing the method for constructing a Chinese character digital sound library according to the present invention, a Chinese digital sound library that meets the small memory requirement of an embedded device can be established while maintaining a relatively high degree of naturalness.

In addition, in the method for constructing the digital sound bank of Chinese characters according to the present invention, it also includes the step of establishing an index structure for describing its position for each digital unit, and the index structure includes the offset of the digital unit in the target sound bank , the duration of the number unit and the context structure of the number unit. Wherein, the context structure of the digital unit includes the code of the previous digital unit, the code of the next digital unit and the code of the current digital unit. The number unit includes the pronunciation representations of the following Chinese characters: zero, one, two, three, four, five, six, seven, eight, nine, ten, hundred, thousand, ten thousand, hundred million, dot, minute, zhi, and.

In addition, in the method for constructing Chinese character digital sound bank according to the present invention, the context unit of the digital unit in the two-tuple digital unit selected from the original sound bank in the original sound bank is the boundary of the digital string or the pair A number unit in a tuple number unit affects another pronunciation unit weakly.

In addition, in the method for constructing Chinese character digital sound bank according to the present invention, the context unit of the digital unit in the one-tuple digital unit selected from the original sound bank in the original sound bank is the boundary of the digital string or is a pair of one-tuple A number unit in a weakly affects another number unit.

In addition, in the method for constructing a Chinese character digital sound library according to the present invention, the digital units in the cut-down two-tuple digital units include at least: zero, one, five, six, and tone-modified yao2, wu2, wan4, yi4 and you4.

In another aspect of the present invention, a method for synthesizing Chinese numeral strings is proposed, comprising the steps of: performing prosodic grouping according to the length and type of the input numeral strings to obtain numeral string groups; extracting the context features of numbers in the numeral string groups , the context feature includes the encoding of the previous number, the encoding of the next number and the encoding of the current number; according to the context feature, select the corresponding number unit from the sound library, and then obtain the number unit of each number in the number string group ; The waveforms of the digital units of the various numbers are spliced together and output; wherein, the sound library is constructed using the method for constructing a Chinese digital sound library as described above.

Utilizing the method for synthesizing Chinese numeral strings according to the present invention, on the one hand, the reading process can be made more pleasing and natural, and on the other hand, the noise generated due to the mismatch of energy and period during splicing can be eliminated.

In addition, in the method for synthesizing Chinese numeral strings according to the present invention, when a numeral is at the beginning of a numeral string or a numeral string group, the encoding of the previous numeral in its context feature is a boundary mark. When a number is at the end of a number string or a number string group, the encoding of the next number in its context feature is a boundary mark.

In addition, in the method for constructing a Chinese character digital sound library according to the present invention, selecting a suitable digital unit from the original sound library includes scoring all candidates of the current digital digital unit, and selecting the highest-scoring candidate digital unit. Scoring is performed by how well the contextual features of the current digit match the contextual features of the candidate digital unit.

In addition, in the method for constructing Chinese character digital sound bank according to the present invention, if two digital units selected from the sound bank are adjacent to two adjacent numbers, then the waveforms of the two digital units are directly spliced, otherwise , performing windowing smoothing processing when splicing the waveforms of the two digital units. The length of the window used in the windowed smoothing process is 20 milliseconds, and the length of the overlapping region between the windows is 0-10 milliseconds.

In yet another aspect of the present invention, a system corresponding to the above-mentioned method for synthesizing Chinese numeral strings is proposed, including: a grouping device, which is used to perform prosodic grouping of numeral strings according to the length of the numeral string and the type of the numeral string, to obtain numerals string group; extracting means, for extracting the context feature of the number in the number string group, said context feature comprises the coding of the previous number, the coding of the next number and the coding of the current number; the selection means, for according to said The context feature selects the corresponding digital unit from the sound bank, and then obtains the digital unit of each number in the digital string group; the splicing device is used to splice and output the waveforms of the digital units of each number; wherein, the sound bank It is constructed using the method for constructing the Chinese digital sound bank as described above.

Utilizing the system for synthesizing number strings of the present invention, on the one hand, it can make the process of reading aloud more pleasing and natural, and on the other hand, it can eliminate the noise generated due to the mismatch of energy and period when splicing.

Description of drawings

The table shown in Fig. 1 has provided the numeral unit in the Chinese numeral string synthesis;

Fig. 2 shows the oscillogram when the pronunciation of adjacent numerals is not cohesive;

Fig. 3 shows the oscillogram when the pronunciation of adjacent numeral is glued;

Fig. 4 shows the consonant classification table of numeral unit;

The table shown in Figure 5 shows the degree of influence between the pronunciation of Chinese characters;

Fig. 6 is the schematic diagram of sound library index structure;

Fig. 7 is the format of the subitem 'feature' in the sound bank index structure described in Fig. 6;

Fig. 8 is the flowchart of the method for constructing Chinese character digital sound storehouse of the present invention;

Fig. 9 is the flow chart of Chinese character numeral string synthesis method of the present invention;

Fig. 10 shows the process of performing prosodic grouping of input digit strings according to one embodiment;

Fig. 11 shows the prosodic grouping of the input Chinese numeral string in the form of a table;

Fig. 12 shows the process of selecting a suitable digital unit from the sound library according to one embodiment of the present invention;

Fig. 13 shows a schematic diagram of splicing the waveforms of selected digital units in the Chinese character array string synthesis method of the present invention;

Fig. 14 is a structural block diagram of the Chinese character digital synthesis system of the present invention;

Fig. 15 shows the evaluation results of the output of the Chinese character numeral synthesis system of the present invention.

Detailed ways

Embodiments of the present invention will be described in detail below with reference to the accompanying drawings, and unnecessary details and functions for the present invention will be omitted during the description to prevent confusion in the understanding of the present invention.

[Definition of digital unit and description of its characteristics]

Chinese includes 'one', 'two', 'three', 'four', 'five', 'six', 'seven', 'eight', 'nine' (or the capitalization of these numbers, such as one, two, etc. etc.) a total of ten basic numbers, and units and marks such as "ten", "hundred", "thousand", "ten thousand", "billion", "point", "minute", "zhi", "again". The number string formed by these nineteen units includes a character string that is read aloud by value, such as an amount, or that is read aloud digit by digit, such as a telephone number. In order to make the synthetic output as natural as possible, it is necessary to consider the three tone transposition phenomenon in the actual Chinese pronunciation, that is, two adjacent upper syllables, the previous syllable becomes Yangping when pronounced, such as "Jiujiu" jiu2 jiu3" (here the Arabic numerals behind the pinyin letters indicate the pitch of the pinyin, 1 means Yinping, which is one tone; 2 means Yangping, which is two tones; 3 means upper tone, which is three tones; 4 means going tones, which is four tones ) instead of "jiu3 jiu3".

Considering the factor of tone transposition, the above unit should also include: "5" is performing a three-tone transposition "wu2", "nine" is performing a three-tone transposition "jiu2", "dot" is performing a three-tone transposition "dian2" , "hundred" is performing three-tone transposition to produce "bai2". In addition, in order to distinguish it from "seven" more easily, "one" is pronounced "yao1" in some occasions. In addition, when a number is at the beginning or end of a number string, the above or below is actually a string group boundary, so it is also classified into one category, represented by "#", although it is not pronounced, it is still seen Make a special number unit.

To sum up, a Chinese numeral string synthesis system involves a total of 25 numeral units, including 24 pronounced numeral units and one silent numeral unit, as shown in FIG. 1 .

Except for single digits, numeric units appear as strings of digits. In the case of strings of numbers, there will be interactions between a number unit and its context's number units. Therefore, it can be described by the context of a digital unit in the digital string, that is, the previous digital unit, the next digital unit and the digital unit itself constitute a vector as the feature of the digital unit.

It should be noted that the 'number' and the 'number unit' in the present invention correspond to each other, since there are no different numbers with the same pronunciation, so, in some cases, the 'number' and the 'number unit' have the same The meaning does not need to be specially distinguished. Likewise, the meanings of 'number string' and 'number unit string' mentioned in the present invention are the same in some cases.

【Pronunciation Impact Factor Table】

As mentioned above, when the number string is read continuously, the numbers will affect each other's pronunciation, so that the pronunciation of the number is different from that of the number alone, even very different. From the waveform point of view, they may be clearly separated from each other, but adhesions may occur.

Fig. 2 and Fig. 3 respectively give the examples of the adjacent digits being clearly separated from each other and sticking together when they are continuously read out. As shown in Figure 2, when we read '02 (zero two)' continuously, the pronunciation of the numeral '0 (zero)' and the pronunciation of the numeral '2 (two)' have no influence on each other, and they can be easily distinguished from each other. However, when '25 (two five)' is read continuously, the pronunciation of the number '2 (two)' and the pronunciation of the number '5 (five)' stick together, and the decomposition line between the two cannot be clearly distinguished from the waveform. ,As shown in Figure 3.

Therefore, in order to measure the degree of pronunciation influence between adjacent digital units, a table of pronunciation influence factors can be established. This table will be used for soundbank pruning when constructing the soundbank, and for evaluating candidate digital units in the soundbank when synthesizing numbers.

In order to establish the table of pronunciation influencing factors, firstly, the pinyin of all number units is classified according to the pronunciation principle. Fig. 4 shows the consonant classification table of the number unit.

Combined with the classification made in Figure 4, the waveform analysis of the digital units in the original sound bank is carried out. The results of the analysis of 13 basic number units are shown in Figure 5. When the latter number is a side sound and a semi-vowel, the pronunciation of the previous number is more affected, that is, the pronunciation has a stronger adhesion relationship, while other The mutual influence between adjacent numbers is relatively small.

In Figure 5, the adhesion relationship between two adjacent numbers when they are pronounced continuously, Y indicates a stronger influence, X indicates a weaker influence, among the rest of the basic number units, the stronger influence includes: "wan4", "yi4" , "you4". Similarly, all 24 numerical units with actual pronunciation can be analyzed to obtain the degree of pronunciation influence among them.

In the embodiment of the pronunciation influencing factor, only the 'strong' and 'weak' forms are used to describe the degree of correlation between two adjacent basic number units. As an optional embodiment, a more detailed description method may be used to describe the degree of correlation between adjacent digital units, for example, three levels or more, and a weighted form.

[Construction of the original sound bank]

In order to finally obtain a high-natural pronunciation of synthetic digital strings, in addition to requiring the original sound library to have a sufficiently high recording quality, it is also required that the sound library can cover as many pronunciation phenomena as possible, that is, collocations that may exist in real pronunciation. The original sound bank designed based on this principle can ensure that when synthesizing a number string, each number can find a candidate unit that matches as much as possible in the sound bank. When doing cuts on this sound bank to reduce the sound bank, it can also ensure the quality of the target sound bank to the greatest extent.

The digital script of the original sound bank should be long enough to cover how the numbers are connected in different collocations. In the number string, the position of each number can be divided into: head, middle and tail. Similarly, in the digital unit string, the position of each digital unit can be divided into: head, middle and tail. Since a three-digit number string can provide the pronunciation of three numbers in different positions and in different contexts, the original sound bank changes from "000" to "999" in turn, and includes the Arabic numeral "1" pronounced as "unit". time combination. Of course, those skilled in the art can use longer digital strings to construct the original sound bank.

After the script design of the sound library is completed, a high-quality original sound library for the synthesis of Chinese numeral strings can be obtained by broadcasters who meet certain pronunciation quality to read and record under certain recording conditions, and then mark the syllable boundaries.

[Construction of the target sound bank]

Although the original sound bank covers a large number of language phenomena, it also contains a lot of redundancy. Therefore, in practice, it needs to be cut to form the target sound bank. The target sound bank is relatively small, but it is guaranteed to cover enough multilingual phenomena.

When two consecutively pronounced Chinese numbers are glued together, the pronunciation of the previous number will have a greater impact on the pronunciation of the latter number, and vice versa. In this case, the acoustic characteristics of the digital utterances will be significantly different from those of the independent utterances. Therefore, in order to obtain a high degree of naturalness, the present invention proposes to construct the target sound library by selecting and retaining binary links of digital pronunciations as basic units.

All binary links have better coverage, but require a larger storage space. In embedded and other application environments, the limitation of storage space makes it necessary to further cut and compress the binary chain.

According to the Pronunciation Affecting Factor Table, for two consecutive digital units with cohesive relationship, retaining their pronunciation samples can improve the naturalness of real synthesis. Such number units include: 0, 1, 5, 6, yao2, wu2, wan4, yi4 and you4, and all the number pairs connected with these number units are selected into the target sound bank. In addition, due to the special tone transposition of the upper tone, when constructing the target sound bank, it is specially reserved, that is, the combination of all digital units that may have three tone transpositions such as wu2-wu3, wu2-wu2, wu2-jiu3, wu2- jiu2, jiu2-wu3, etc. are selected into the target sound bank.

Studies have shown that for two consecutive numbers that are not isolated, their acoustic properties are not consistent in their different contexts. For example, the glued number "35" in "356" differs from "35" in "353" because the subsequent number "6" or "3" still has an effect on the pronunciation of the previous number. Therefore, choosing a context with a weaker impact will enable the selection of a more general numerical sample, according to the impact factor table.

The research also shows that two consecutive numbers with strong pronunciation cohesion have different pronunciations in different positions in the number string group. For example, the acoustic performance of the "35" in "353" and the "35" in "335" are not consistent. Retaining these two different types of pronunciation samples can improve the naturalness of the synthesis.

When the cohesive relationship between a number unit and subsequent number units is not strong, its pronunciation will be relatively independent. When the position of these number units in the number string is different, the pronunciation is also different. For example, the pronunciation of "9" in the middle of the number string is shorter than that of "9" at the end.

The more important reason is that the auditory pause of a digital unit is more natural than the pause produced by inserting silence, and an important factor of the auditory pause is the position where the digital unit appears in the digital string. Therefore, these digital units should retain their pronunciation samples at the head, middle and tail respectively. The selection method is:

1) If the number is at the head of the number string, the number should have a weaker bonding relationship with the pronunciation of the next number in the string;

2) If the number is located in the middle of the number string, then the number should have a weak bonding relationship with the pronunciation of the numbers before and after it;

3) If the number is at the end of the number string, the number should have a weaker bonding relationship with the pronunciation of the previous number.

As an example, "3" as a contextual number can ensure that the pronunciation of the current number is less affected by the context, that is, the "3" located in different positions is intercepted by "×33", "3×3" and "33×". ". Since "wu2" and "jiu2" are not pronounced independently, these two number units are not considered when intercepting the single. The above and below of this type of numerical unit are considered as head and tail position marks or middle position marks.

After the sound bank unit is selected, it is extracted from the original sound bank and saved in a certain order to generate an independent target sound bank, and an index structure is used to describe the position information of each digital unit. The index structure is described in Figure 6.

The item 'feature' in the table of FIG. 6 describes the context of the digital unit, and this value can be described by a value of 16 bits width as shown in FIG. 7 . The assignment of the three subkeys in 'Characteristics' obeys the following rules:

1) When the number unit is at the beginning of the string, the previous number in the 'feature' is marked as #;

2) When the number unit is at the end of the string, the last number in the 'feature' is marked as #;

3) The current digital unit is not located at the beginning of the string, but is positioned at the first unit of the selected one-tuple or two-tuple, then the previous digital unit of the 'feature' is marked as &;

4) the current digital unit is not located at the end of the string, but is positioned at the second unit of the selected one-tuple or two-tuple, then the latter digital unit of its 'feature' is marked as &;

5) Otherwise, the previous number unit and the next number unit in the 'feature' are taken as the actual context unit of the current number unit.

Fig. 8 shows a flowchart of a method for constructing a Chinese digital phonetic library according to an embodiment of the present invention. The method for constructing a Chinese character digital sound library of the present invention will be described in detail below with reference to FIG. 8 .

In step S81, an original sound library including a plurality of digital cell strings is generated. For example, the number string read aloud changes from "000" to "999" in turn, and includes the combination of the Arabic numeral "1" when it is pronounced as "unit". At the same time, it is recorded to determine the boundary mark, thereby forming a copy for the Chinese character number string. Synthesized raw soundbank.

Next, in step S82, select the two-tuple digital unit that is positioned at string beginning, in the string and the string end from the original sound storehouse in the unit of two-tuple, just select two from the three-element digital unit string of the original sound storehouse. Tuple number units, and when selecting, consider the situation that the numbers in the two-tuples are located at the beginning, middle and end of the three-tuple number unit string respectively.

Then, select a tuple of digital units from the original sound bank in step S83, that is, select corresponding numerical units according to the different positions of a single number in the triplet number string. After the above two selection steps, the intermediate sound bank including the two-tuple digital unit string and the one-tuple digital unit is obtained. Although the sound bank at this time is smaller than the original sound bank, it still exists in the two-tuple digital unit string Redundancy requires further pruning.

In step S84, because the degree of influence between the two array units in the binary group is strong or weak, for the binary number strings with weak mutual influence, it can be synthesized with a single number without reducing the quality of the synthetic result. Naturalness. Therefore, the binary number unit strings that have weak influence on each other are cut out. Finally, in step S85, the final target sound bank is composed of the trimmed binary digital unit string and the one-tuple digital unit selected from the original sound bank.

In addition, it is necessary to index the positions of the digital units in the target sound bank to form an index file, so as to find suitable candidate digital units when synthesizing digital strings.

The sound library constructed through the above process occupies a small storage space and maintains a high degree of naturalness, which can meet the requirements of embedded systems.

[Synthesis of Chinese number strings]

The process of synthesizing Chinese numeral strings according to an embodiment of the present invention includes: digit string input, prosody grouping, feature extraction, unit selection, waveform splicing and output. Its process is shown in Figure 9.

1. Digital string input

Usually, the stored number string is read from a specific area of the memory according to the instruction, and the length and type of the number string can be known in advance, for example, telephone number, ID number, driver's license number, securities data and industrial instrument readings, etc. .

2. Rhythmic grouping

Under normal circumstances, in order for the listener to have enough time to understand the content read aloud, and to make the reading process more pleasing and natural, longer digit strings should be split into shorter digit string groups. In this way, the number strings read aloud naturally have a certain rhythmic structure. The prosodic structure first depends on the syntax and semantics of the number string, such as the information grouping in the ID number, and secondly depends on the reading habit and ability, such as the grouping of 8-digit phone numbers.

Usually, the prosodic grouping of digit strings has certain rules. As an example of prosodic grouping, Fig. 10 shows the process of prosodic grouping the input phone numbers.

First, a telephone number having a predetermined length L is read in. Then judge whether its length L is greater than 6, and if it is greater than 6, group it according to 3 or 4. For example, for an 8-digit phone number, group by 4-4.

If the length L of the phone number is not greater than 6, it is judged whether L is equal to 6, and if it is equal to 6, group or not group according to 3-3.

If the length L of the phone number is less than or equal to 6, it is judged whether it is equal to 5, and if it is equal to 5, grouping is performed according to 2-3 or not.

Finally, if the length L of the phone number is less than 5, such as a phone number within a company, no further grouping is done.

In addition, as shown in FIG. 11 , a more detailed situation of prosodic grouping of telephone numbers is also specifically listed. Through prosodic grouping, the features of the digits are more able to describe the characteristics of the digits themselves, thus helping to select the correct candidates from the sound bank.

3. Feature extraction

After prosodic grouping the input digit strings, the features of each digit, that is, its context, can be extracted. When dealing with numbers, it is necessary to consider the upper tone transposition, that is, when wu3, jiu3, bai3, and dian3 are pronounced continuously, the previous tone is converted into wu2, jiu2, bai2, and dian2. The principle of feature extraction is basically the same as the process described above for the feature of the digital unit in the "Construction of the target sound bank", but only when the number is at the beginning of the string or group, the last digital unit of the feature is set to #, When the number unit is at the end of the string or group, the next number unit of the setting feature is #. After the digit feature extraction is completed, the contextual feature description of each digit is obtained, and this feature will be used to select a suitable candidate unit from the sound bank.

4. Unit selection

The unit selection process is to select the most suitable candidate digital unit from the target sound bank. The unit selection process first finds all candidates of the current digital unit to be matched through the sound bank index file, then scores and judges each candidate in turn, and finds the candidate with the highest score as the final target candidate.

The scoring method considers the matching degree of the context, and the possible and feasible scoring functions are as follows:

S=Sp+Sn

Where Sp is the evaluation of whether the previous number unit in the context matches or not, and Sn is the evaluation of whether the next number unit in the context matches or not. The scoring method of Sp and Sn is to check whether they match completely at first, and if they do not match completely, then check whether the context of the candidate number of the sound bank is the tuple boundary mark &. As shown in Figure 12, the specific scoring method is as follows:

If the previous digit unit matches, then Sp=3;

Otherwise, if the previous digital unit of the candidate digital unit in the target sound bank is &, then Sp=2;

Otherwise, Sp = 1;

If the latter digit unit matches, then Sn=3;

Otherwise, if the last digital unit of the candidate digital unit in the target sound bank is &, then Sn=2;

Otherwise, Sn=1;

Then, sum Sp and Sn to get the final score of the candidate digital unit.

By repeating the above steps, the final equalization of all candidate digital units is obtained, and the digital unit with the highest score is selected, and its waveform is further obtained according to the index structure of the digital unit.

5. Wave splicing

The waveform splicing process is to sequentially splice and connect the waveforms of each selected candidate digital unit to form a completed speech waveform. In order to eliminate the noise and "click" sound caused by the mismatch of energy and period when two digital units are directly spliced, it is necessary to add a window to the splicing point for smoothing, and for two adjacent digital units in the original sound bank Then splice directly. A possible and feasible window function is shown in Figure 14, where the window length is about 20 milliseconds, the length of the overlapping region is between 0 and 10 milliseconds, and the attenuation method is a positive slope from the starting point of windowing to the end point of windowing. or reverse decay.

For the adjacent numbers that have a glue relationship, they have all been selected when building the target sound bank, so when the input number string contains such a number chain, the unit selection process can always effectively select the most suitable number unit from the target sound bank. The library is selected and can be spliced directly, which ensures the highest degree of naturalness.

6. Waveform output

During the waveform output process, output the spliced waveform directly from the speaker, or output it after being amplified.

In addition, Fig. 14 also shows a block diagram of the Chinese numeral string synthesis system according to the present invention.

As shown in FIG. 14 , the Chinese numeral string synthesis system 10 of the present invention includes: an input part 141 , a grouping part 142 , an extraction part 143 , a selection part 144 , a splicing part 145 , an output part 146 and a storage part 147 . These components described above are described below.

The input section 141 inputs a number string having a predetermined length, and the grouping section 142 rhythmically groups the number string according to its type and length. For example, for a 15-digit ID number, it is usually divided into 3-3-6-3, and for an 18-digit ID number, it is usually divided into 3-3-4-4-4. For phone numbers, the prosodic grouping is performed according to the situations listed in Fig. 11 .

The extraction part 143 extracts the context features of the numbers to be synthesized, and the context features include: the code of the previous number, the code of the next number and the code of the number itself. Then, the selection part 144 selects the most suitable digital unit from the storage part 147 storing the target sound bank according to the context characteristics of the numbers to be synthesized, and the specific matching process is the same as that described above in conjunction with accompanying drawing 13 .

Next, the splicing section 145 splices the selected waveforms. If the number strings to be synthesized are adjacent to each other in the target sound bank, they are directly spliced together. Otherwise, window processing is used to eliminate the "click" sound caused by the mismatch of energy and period when two digital units are directly spliced. Finally, the output part 146 outputs the concatenated waveform directly from the speaker, or outputs it after being amplified.

In the above embodiment, the various components are separated, but it is clear to those skilled in the art that one or more of the above-mentioned parts can be integrated together, for example, the function of the input part 141 is integrated in the prosodic grouping , or directly integrate the function of the output part 146 into the waveform splicing part 145, and so on.

【Implementation Effect】

Figure 15 describes the relationship between the average number of pronunciation unit samples in the target sound bank and the subjective evaluation score of the system synthesized speech. The goal of subjective evaluation is naturalness, which is a method in which participants subjectively listen to and defend the system, give a score between 0 and 5, and finally obtain a final score by weighting all the scores.

It can be seen from Figure 15 that when the number of pronunciation unit samples reaches between 15 and 30, the subjective evaluation is already above 4 points, indicating that the sound quality is quite good.

According to Fig. 15, the summary of the implementation effect is as follows: According to the method provided by the present invention, a high degree of naturalness can also be obtained on a small sound bank. While maintaining high naturalness of synthesized speech, the size of the sound bank is small enough to be ported to embedded devices with small memory.

【Variety】

The present invention has been described in detail above in the form of embodiments, but the embodiments here are only for the purpose of clear illustration, and are not intended to limit the scope of the present invention. For example, 25 numeric units are defined in 'Definition of Numerical Units and Description of Their Characteristics', but this is only an application. In industrial applications, the numeric strings to be converted into text may have their own measurement units, For example 'degree', 'meter', 'Pa', etc. However, in everyday devices with text-to-speech functions, it is necessary to read units of measurement such as 'jin' and 'liang'. Such as the calling system of places such as banks and tax bureaus needs to synthesize voices such as '××× number'. That is to say, in a specific application, the number of digital units may be more than the 25 described above.

Accordingly, the present invention may have other specific forms without departing from its spirit and essential characteristics. Therefore, the present invention considers all aspects by way of example rather than limitation, so it is intended to include all changes within the scope of the invention and the equivalent meaning and scope of the claims determined by the claims rather than the above description or examples. .

Claims (31)

1. A method for constructing a Chinese digital sound bank, comprising steps: Produce an original sound bank that comprises a plurality of numeral unit strings, and said numeral unit string is the representation of the pronunciation of corresponding Chinese numeral string; From the original sound library, the number units in which are selected are respectively located at the beginning of the string, in the string and at the end of the string; Select the one-tuple number units wherein the number units are respectively located at the beginning of the string, in the string and at the end of the string from the original sound bank; Pruning such a two-tuple digital unit, the degree of influence between adjacent digital units in the two-tuple digital unit is weak; The target sound bank is formed by the trimmed two-tuple digital units and the selected one-tuple digital units. 2. The method according to claim 1, further comprising the step of setting up an index structure for describing its position for each digital unit, said index structure comprising the offset of the digital unit in the target sound bank, the number of the digital unit Context structure for duration and number units. 3. The method according to claim 2, wherein the context structure of the digital unit includes a code of a previous digital unit, a code of a next digital unit and a code of a current digital unit. 4. The method according to claim 2, wherein said number unit comprises a representation of the pronunciation of the following Chinese characters: zero, one, two, three, four, five, six, seven, eight, nine, ten, hundred, thousand, Ten thousand, one hundred million, one point, one point, one, and another. 5. The method of claim 4, wherein the numeric unit further comprises a phonetic representation of the unit of measure of the string of numbers. 6. The method according to any one of claims 2-5, wherein the context unit of the number units in the binary number units selected from the original sound bank in the original sound bank is the boundary of the number string. 7. according to the described method of one of claim 2-5, wherein, the context unit in the context unit in the binary group digital unit of picking out from the original sound bank is to the binary group digital unit A number unit affects another pronunciation unit weakly. 8. The method according to any one of claims 2-5, wherein the context unit of the number unit in the tuple number unit selected from the original sound bank in the original sound bank is the boundary of the number string. 9. according to the described method of one of claim 2-5, wherein, the context unit of the digital unit in the one-tuple digital unit that is selected from the original sound bank is to the digital unit in the one-tuple influence Another digital unit that is weak. 10. The method according to any one of claims 2-5, wherein the digital units in the pruned two-tuple digital units at least include: zero, one, five, six and transposition yao2, wu2, wan4, yi4 and you4. 11. The method of claim 6, wherein each of the plurality of numeric unit strings in the original sound bank is a triplet numeric unit string. 12. The method of claim 7, wherein each of the plurality of numeric unit strings in the original sound bank is a triplet numeric unit string. 13. The method of claim 8, wherein each of the plurality of numeric unit strings in the original sound bank is a triplet numeric unit string. 14. The method of claim 9, wherein each of the plurality of numeric unit strings in the original sound bank is a triplet numeric unit string. 15. The method of claim 10, wherein each of the plurality of numeric unit strings in the original sound bank is a triplet numeric unit string. 16. A method for synthesizing a Chinese numeral string, comprising the steps of: Prosodic grouping is performed according to the length and type of the input number string to obtain a number string group; Extracting the contextual features of the numbers in the digital string group, the contextual features include the encoding of the previous number, the encoding of the next number and the encoding of the current number; Select the corresponding digital unit from the sound library according to the context feature, and then obtain the digital unit of each number in the digital string group; Stitching together the waveforms of the digital units of each number to output; Wherein, the sound bank is constructed using the method for constructing a Chinese digital sound bank according to claim 1. 17. The method according to claim 16, when a number is at the head of a number string or a group of number strings, the encoding of the previous number in its context feature is a boundary mark. 18. The method according to claim 16, when a number is at the end of a number string or a number string group, the encoding of the next number in its context feature is a boundary marker. 19. The method of claim 16, selecting a suitable digital unit from the original sound bank comprising selecting the highest scoring candidate digital unit by scoring all candidates for the current digital unit. 20. The method according to claim 19, scoring is performed according to the matching degree of the context features of the current number and the context features of the candidate number units. 21. The method according to claim 16, if the two digital units selected from the sound bank are adjacent for two adjacent numbers, then the waveforms of the two digital units are directly spliced, otherwise, when splicing the Windowed smoothing is performed on the waveforms of the two digital units. 22. The method according to claim 21, the length of the window used in the windowed smoothing process is 20 milliseconds, and the length of the overlapping region between the windows is 0-10 milliseconds. 23. The method of claim 16, wherein the type of string of numbers includes: telephone number, identification number, driver's license number, security data, and industrial meter readings. 24. A Chinese numeral string synthesis system, comprising: The grouping device is used for rhythmically grouping the number strings according to the length of the number string and the type of the number string to obtain the number string group; Extracting means for extracting the context features of the numbers in the number string, the context features including the encoding of the previous number, the encoding of the next number and the encoding of the current number; The selection device is used to select the corresponding digital unit from the sound bank according to the context feature, and then obtain the digital unit of each number in the digital string group; splicing device, for splicing and outputting the waveforms of the digital units of the respective numbers together; Wherein, the sound bank is constructed using the method for constructing a Chinese digital sound bank according to claim 1. 25. The system according to claim 24, when a number is at the beginning of a number string or a group of number strings, the encoding of the previous number in its context feature is a boundary marker. 26. The system according to claim 24, when a number is at the end of a number string or a number string group, the encoding of the next number in its context feature is a boundary marker. 27. The system according to claim 24, the selecting means selects the highest-scoring candidate digital unit by scoring all candidates for the digital unit of the current digit. 28. The system according to claim 27, wherein the selecting means performs scoring according to the matching degree of the contextual features of the current digit and the contextual features of the candidate digit units. 29. The system according to claim 24, if the two digital units selected from the sound bank of two adjacent numbers are adjacent, then the waveforms of the two digital units are directly spliced, otherwise, when splicing the Windowed smoothing is performed on the waveforms of the two digital units. 30. The system of claim 29, the length of the windows used in the windowed smoothing process is 20 milliseconds, and the length of the overlap between windows is 0-10 milliseconds. 31. The system of claim 24, wherein the types of strings of numbers include: telephone numbers, identification numbers, driver's license numbers, security data, and industrial meter readings.

Images