RU61924U1 - STATISTICAL SPEECH MODEL - Google Patents

Abstract

The utility model relates to the field of speech technology, the model generates a sound stream and can be used for analysis, synthesis and speech recognition, as well as assessing the quality of vocoders and communication channels. The task that the utility model aims to solve is to create a statistical speech model that combines the elements of a speech synthesizer, statistical data, and large-volume speech data bodies. The technical result is achieved due to the fact that in a static speech model that includes an interface unit, a speaker selection unit containing a speaker selection generator, a sound selection unit generating samples of sounds that are converted into audio signals with predetermined properties and a database in the speech stream generating unit containing descriptions of typical speakers and other necessary information, changes were made, namely:

- a module for statistics of population parameters of various regions, the output of which is connected to the input of the speaker selection generator, is additionally introduced into the speaker selection block;

- additional blocks have been introduced: samples of typical speakers and storage prosodik selected sounds or chains of sounds;

- two modules are additionally introduced into the sound selection block: rules for following sounds and rules for naming allophones.

In addition, changes were made to the structure of some blocks. The utility model has several advantages, which include the following:

- a larger volume of the original speech corpus;

- inclusion in the database of additional statistical information for each AP;

- the availability of descriptions of the intonation contours for each AP;

- the ability to simultaneously work with structural elements of different sizes and formats, and a number of others.

In addition, the statistical speech model is language independent, as all algorithms and its interfaces will be preserved. Currently, the proposed statistical model is being tested by using a text-based speech synthesis system and an objective assessment system for vocoder quality.

Description

The utility model relates to the field of speech technology. The model generates a sound stream and can be used for analysis, synthesis and speech recognition, as well as assessing the quality of vocoders and communication channels.

The question of what is a “statistical model of speech” can be approached from different angles. From an applied point of view, it is a means of studying and modeling the processes of speech activity, a basis for building systems of synthesis, analysis, and speech recognition. However, we can consider the statistical model on a global scale as a cast of the current state of the language, preserving it for posterity.

Traditionally, the following four areas are referred to speech technologies:

- speech synthesis (primarily text-to-speech);

- speech recognition;

- voice identification;

- speech processing

Until recently, programs that synthesize and recognize human speech, for the most part, could be attributed to "demonstration" systems. They may or may not take the form of practical decisions, but in any case, their creation did not imply a profit. The main goal of the developers of such solutions was to demonstrate the level of scientific achievement.

To teach a computer to communicate in a natural language is an extremely interesting and important task. The main form of expression of thoughts, goals and desires for a person is speech, as the most natural and convenient form of information transfer. When entering commands from the keyboard, a person is forced to formulate thoughts in a strict grammatical form, to control the correctness of the typed text on the screen. Manual input is a long and tedious task - an ordinary person is able to enter 10-20 words per minute, while up to 200 words can be spoken in a minute. In addition, not all users can and want to use the keyboard.

To work with an unprepared user, speech input is most attractive - in a natural language, almost any tasks from various fields of human activity can be formulated. Speech input does not cancel other means of information exchange (I. Zharkov, P. Skrelin, M. Gusev “Voice of Time”, Computer Press No. 8, 2005).

Formally, the speech recognition process can be described in just a few sentences. In theory, everything is simple, but if you go to practice, as it turns out a lot

“Minor” problems, which to this day do not allow researchers to obtain a recognition percentage suitable for use in commercial systems.

With the identification (verification) of the speaker’s voice - the problem is the same as with recognition - a high percentage of errors - no one wants to protect their bank account with a voice password if, say, every thousandth person can access it.

To date, two basic principles of building speech synthesis systems are known - this is “by the rules” synthesis and compilation synthesis. Synthesis “by the rules” is based on the formation of the physical characteristics of speech sounds based on their mathematical descriptions and has a low naturalness. Comparative synthesis involves cutting segments from a natural speech sequence and their subsequent processing and gluing. Digital processing of the speech signal allows us to solve the problem of changing the frequency of the fundamental tone and the duration of the signal fragment.

A vocoder models speech as a system response to arousal over short time intervals. Examples of vocoder systems include linear prediction vocoders, homomorphic vocoders, channel vocoders, sine wave encoders (SSCs), multi-band excitation (IPM) vocoders, and advanced multi-band excitation vocoders (UsMOS). In these vocoders, speech is divided into short segments (usually 10-40 ms), each segment being characterized by a set of model parameters. These parameters usually represent several basic elements of each speech segment, for example, segment pitch, speech state, and spectral envelope.

A vocoder may use one of many known representations for each of these parameters. For example, a step may be represented by a step period, a fundamental frequency, or a delay in long-term prediction. Similarly, a speech state can be represented by one or several voiced / unvoiced decisions, a measure of speech probability, or the ratio of periodic energy to stochastic. The spectral envelope is often presented in the form of a filter response with a transfer characteristic with one pole, but can also be represented by a set of spectral amplitudes or other spectral measurements (RF Patent No. 2214048).

A multiband excitation (MPV) speech model has the ability to provide high-quality speech and operate at medium to low bit rates. This model uses a flexible speech structure that allows

receive more natural-sounding speech and makes it more resistant to the presence of acoustic background noise.

Most modern speech encoders are based on some model of the speech path used to form the encoded speech signal. The parameters and signals of the model are quantized, and information describing them is transmitted over the channel. The dominant encoder model in cellular telephony applications is Code Excited Linear Prediction (LPCV).

There are many known implementations of the code based on the LPKV model, for example, RF patent No. 2223555. Coded speech is generated by an excitation signal supplied through a pole synthesizing filter with an order usually equal to 10. The excitation signal is generated as the sum of two signals that are selected from the corresponding codebooks (one fixed and one adaptive) and then multiplied by the corresponding amplification factors. Codebook signals typically have a duration of 5 ms (subframe), while a synthesis filter is usually adjusted every 20 ms (frame). The parameters associated with the LPKV model are synthesis filter coefficients, codebook entries, and gain factors.

As a result of a search conducted on literature and patent sources, from the current level of technology there is no unambiguous prototype, because identified analogues belong to individual parts of the model, therefore, do not allow to qualitatively solve the problem.

The task that the utility model aims to solve is to create a statistical speech model that combines the elements of a speech synthesizer, statistical data and speech data corps of a large volume, with the following objectives: - to improve the quality of synthetic speech generated by synthesis systems using the model;

- creating sound streams for training speech recognition systems;

- Creation of sound streams for testing and evaluating the quality of vocoders and communication channels.

An analysis of the descriptions of analogues showed that, in a statistical model, to solve the problem, the following well-known blocks can be used: a database containing descriptions of speakers, blocks for selecting sounds, a block for generating a speech stream, an interface block for coordinated control of the components of the model and its interaction with the user.

The user of the model can be an automatic device or an operator that generates a request by using the interface.

The technical result is achieved due to the fact that in a statistical speech model, including an interface unit, a speaker selection unit containing a speaker sample generator, a sound selection unit that generates samples of sounds that are converted into audio signals with predetermined properties and a database in the speech stream generation unit containing descriptions of typical speakers and other necessary information, changes were made, namely:

- a module for statistics of population parameters is added to the speaker selection block, the output of which is connected to the input of the speaker sample generator;

- between the speaker selection block and the sound selection block, a sample block of typical speakers is additionally included;

- between the unit for selecting sounds and the unit for forming a speech stream, an additional prosodyki storage unit is also included, also connected to one of the inputs of the interface unit;

- two modules are additionally introduced into the sound selection block: the rules for following sounds and the rules for naming allophones, the outputs of which are connected to the inputs of the corresponding modules.

Accordingly, with these additional requirements, and also, taking into account the possibility of interaction of individual blocks with the interface block during operation, as well as with individual elements of the statistical model, the modular structure of the main blocks of the model was chosen, and the structural diagram of the connection of individual modules in the blocks may differ slightly, depending on a specific example of the implementation of a statistical speech model.

Most of the blocks of the model have a complex structure due to the large set of functions performed by them. They will be described in more detail in an example implementation of a statistical speech model.

The utility model is illustrated in the following Figures 1-4, and the implementation of Fig. 5, 6.

In Fig. 1 - the general structure of the statistical model of speech;

Fig. 2 - structure of the speaker selection block;

Fig. 3 - structure of the block for selecting sounds;

In Fig. 4 - the structure of the block forming the speech flow;

Fig. 5 is an example of the use of a statistical model in the synthesis of speech by text;

Fig. 6 is an example of using a statistical model in a vocoder quality assessment system.

The general structure of the statistical model shown in Figure 1.

Interface unit 1 provides interaction with the outside world (or the User). Block No. 1 also synchronizes the operation of the remaining blocks of the statistical model.

The speaker selection block (block 2) generates a sample of typical speakers (TD) or a sequence of TD indices. Depending on the team, either a representative sample of APs or a sample consisting of one AP can be generated. A representative sample is representative in the sense that the distribution of speech parameters in the sample will correspond to the distribution of speech parameters of the population described by the model.

The sequence of TD identifiers generated at the output of this block is stored for further use in the TD sample block (block 3).

The block for selecting sounds (block 4) forms a prosodic (descriptions of sounds). Depending on the command, the prosodic is formed either for a representative sample of sounds, or for a given sequence of sounds, or for one given sound.

The prosodica is stored in the prosodiki buffer (block 5) until further use. The block for the formation of the speech stream (block 6) converts the descriptions of sounds into samples of the sound signal, and the result is output to the interface unit 1.

The block of descriptions of typical announcers (block 7) stores the descriptions of the APs and returns upon request: the necessary parts of the descriptions, information about their number, a list of speakers. It also contains other necessary information for the functioning of the model, i.e. block 7 is the database of the statistical model.

The statistical speech model is not a thing in itself; it is designed to work as part of various systems in which it is required to simulate the speech flow that are users of the model.

So, the User can execute the following types of requests:

1. request a list of typical speakers (TD) presented in the model;

2. synthesize individual sounds with the voice of any AP;

3. synthesize a chain of sounds with the voice of any AP;

4. generate a sound stream characterizing one AP;

5. generate a sound stream characterizing the population described by the model;

6. cancel the generation of sound stream.

Speaker selection block 2 (Fig. 2) contains two modules: generator 2.1. speaker samples and module 2.2 - statistics of population parameters of regions.

When generating a representative sample, the speaker selection block receives statistical data on the population described by the model (module 2.2): region, age and gender composition, educational level, dialect / dialect, occupation, etc. Part of the AP descriptions (block 7) used by the module consists of ranges of statistical parameters that allow you to determine the percentage of the population corresponding to each AP and include the AP sample in an amount proportional to the composition of the population.

Having determined how many and what typical speakers need to be included in the sample, the speaker selection block randomly gives their indices to block 3 for storage.

After all the necessary indexes are saved, the block reports to the interface block 1 about the completion of the command.

When you select one specific speaker, the speaker selection unit simply transfers its index for storage to block 3 and reports to block 1 about the completion of work.

Block 4 - the block for selecting sounds, shown in Fig. 3, consists of the following modules connected in series: chain formation - 4.1; attribution of intonation contours - 4.2; naming allophones - 4.3; determining the duration and energy of the sound chain - 4.4; imposition of intonation contours - 4.5.

In addition, to form chains of sounds, the signal from module 4.6, the rules for following sounds, is received at the corresponding input of module 4.1, and the signals from module 4.7, the rules for naming allophones, are input to module 4.3.

Sound selection block 4 operates in one of two modes: distribution generation mode and synthesizer mode, when sound and parameters are issued by interface unit 1. Let us consider these modes in more detail.

In the distribution generation mode for each AP (the indices of which are taken from block 3), samples of sounds with parameters are formed (module 4.1 works), as in the speech of the AP. To do this, information on the frequency of sounds is taken from the TD description block (block 7) by index, and, taking into account the rules for following (module 4.6), chains of sounds from pause to pause are prepared. The lengths of the chains are also determined by the parameters of the AP.

If for some reason there is no information in the TD parameters about the frequency of sounds or the length of chains, then statistical data obtained on a large volume of texts are used. Average statistics is part of block 7 of the TD descriptions.

Each chain is assigned an intonation circuit (module 4.2). The parameters of the intonation circuits and information about their use in speech are taken from the descriptions of the TD. If the parameters of the intonation circuits or the distribution of the durations of sounds or energy are absent in the description of this AP, the average values are taken, which are part of block 7 of the AP descriptions.

Taking into account the contexts (sounds left and right) and the rules for naming allophones (module 4.7), the names of the sounds are converted to the names of allophones in module 4.3.

For all sounds of the chain, on the basis of information on the parameters of the AP, durations and energies are assigned (module 4.4). After that, the intonation circuit (IR) is superimposed. As a result of applying intonation circuits to chains of sounds (block 4.5), the main tone is determined for each sound.

As a result, at the output of module 4.5, all necessary prosodic parameters are known: duration, pitch, energy, and TD identifier. The generated prosodic is stored in block 5. Upon completion of the prosodic generation, block 4 reports on the execution of the command in interface block 1.

The distribution of prosodic parameters of sounds in chains coincides with the distribution in real speech.

In synthesizer mode, either separate sounds or chains of sounds from pause to pause are received at the input of block 4 of sound selection. If you select a single sound, all prosodic parameters (duration, pitch frequency (BST), energy) are indicated by block 1. Block 4 only names the allophone using the context “pause” - “sound” - “pause”, transfers the description to storage block 5, and reports on the execution of the command in block 1.

In the case of generating a sequence of sounds, block 4 forms a prosody in accordance with the parameters of the intonation circuit, determines the names of allophones and transfers them to the input of block 5. After which it reports the execution of the command to block 1.

Different procedures are used to determine the duration of sounds when working in synthesizer mode and in sample generation mode. If during the generation of the sample, the durations of sounds are determined by the statistical distribution, then in the synthesizer mode strict rules are used, and the central values of the distributions. The rules for determining the duration are determined by the specific implementation (module 4.6).

Block 6 - the block of the formation of the speech stream shown in Fig. 4 consists of series-connected modules: 6.1 - the formation of the duration of sounds; 6.2 - change in frequency response; 6.3 - the formation of the amplitude envelope; 6.4 - processing joints of sounds; 6.5 - cast to the format specified by the interface unit 1.

Block 6 of the formation of the speech stream, having received the prosodic (from block 5) and the format of the sound stream (from block 1), extracts sound samples with markup from the description of the AP (from block 7). Each sound is reduced to a duration determined by prosodiki parameters

in module 6.1. For sounds of different types, different algorithms (strategies) for changing the duration are used, providing minimal distortion of the sound quality. Concrete algorithms are an attribute of implementation.

After the durations of allophones are formed, they are reduced to the specified frequencies of the fundamental tone in module 6.2, and the frequency response does not remain constant over the entire allophone, but changes in accordance with the movement specified in prosodic parameters. To minimize distortion of sounds, modifications of the frequency response of sounds of different types can be carried out using various algorithms. Concrete algorithms are an attribute of implementation.

Further, taking into account the energy parameters specified in the prosodic, the amplitude envelope of the chain sounds is formed (module 6.3), and the morph of the sound junctions is performed to minimize noise at the joints (module 6.4).

The audio signal is converted to the format specified by the interface unit 1 and transmitted to it for later use. Audio format conversion is performed in module 6.5.

The generated speech stream from block 6 enters the interface unit 1. Block 1 transmits the speech stream to the User as it arrives. Upon receipt of a report from block 6, the User is informed of the completion of the formation of the speech stream.

The following is a description of the implementation of the statistical speech model. made on a personal computer in accordance with the description of the modules and interaction schemes of the blocks given above. As blocks of the model, software tools and computer equipment were used. Software can be developed to control the operation of individual blocks of the model.

For a better understanding, the following are definitions of the terms contained in the description of the statistical speech model.

Allophone - (from the Greek. Allos - different. Different and phone - sound, option, phoneme variety, due to this phonetic environment.

Phoneme - (from the Greek. Phonema - sound), the basic unit of the sound system of the language, the limiting element highlighted by the linear division of speech.

Syntagma - (from the Greek. Syntagma, literally - together built, connected), in the broad sense - any sequence of linguistic elements connected by a defined-determinative relation. In a narrower sense, C. is a phrase singled out as part of a sentence (C. is predicative, attributive, object, etc.), and

 sentence is a syntagm chain sequence. If even simpler, then syntagma. is a sequence of sounds from pause to pause.

Sonant - among voiced consonants, a group of sonorous consonants, or sonants (for example, l, r, m, n, d), distinguished from noisy S. (both voiced and deaf) by the presence of a clear formant structure stands out. The latter brings sonants together with vowels, but they are distinguished by a lower total energy. To nasal ones, in particular, nasal.

Frictive, explosive, affricate - according to the nature of the noise-forming barriers, consonants are divided into closest, fricative and trembling. The first are formed due to the closure of two active pronunciation organs of speech, for example, the lower and upper lips (p, b, m,) or an active organ with a passive, for example, the tongue with the sky (t, d, k). A bow can end with a sharp opening, an explosion or a gradual opening, a transition to a gap. In the first case, explosive consonants arise (n, b, t, d), in the second - the so-called affricates, for example, Russian (c and h), which are kind of complex sounds, because have a commissive and crevice (fricative) elements.

Format - in this case, the sampling frequency of the audio signal and the number of bits allocated to the sample are implied.

Let us dwell in more detail on the operation algorithm of the speaker selection block (block No. 2).

The statistics on the composition and characteristics of the population are based on data obtained by the Goskomstat of Russia as a result of the 2002 All-Russian Population Census. To simplify this specific implementation, in module 2.2 it was decided to use only information on the age-sex composition of the population.

Accordingly, two criteria for gender and age were used to link TDs with population statistics. Six, very conditional, TDs were introduced.

Table number 1 TD No. Floor Age TD No. Floor Age one m younger than working age four well younger than working age

2 m at working age 5 well at working age 3 m older than working age 6 well older than working age

The sampling procedure of the AP implemented in block 2 works as follows:

1. Based on statistics on the age-sex composition of the population, the percentage of the population corresponding to each AP is determined;

2. The percentages are converted to integers (by multiplying by 10);

3. The values are minimized (the least common divisor (GCD) of all percent values is searched, after which they are all divided by it);

4. The sum of the percent values (N _TD ) is calculated and a uniform random number sensor is started. The sum of the percentages is equal to the length of the sample, so you need to set reasonable restrictions on the accuracy of their reduction to integers;

5. The intervals of values corresponding to the TD are constructed ([0, N _TD1 [, [N _TD , N _TD1 + N _TD2 [...);

6. N _APs of random number sensor values are generated. When the sensor value falls into the interval, the AP is included in the sample, which is transmitted to the sound selection unit 3 as the index of the selected speaker.

The speaker selection unit 2, depending on the command of the interface unit (unit 1), either selects one specific speaker or generates a sample in accordance with the algorithm described above (module 2.1). Statistical data (module 2.2) are designed as a data structure that is part of the database (block 7). The generated sample is saved in block 3 as a file.

Block 4 sound selection is a complex block consisting of separate modules that implement the algorithm of the block. At the command of the interface unit 1 in module 4.1, the formation of sound chains is carried out. For this, module 4.1 simultaneously with the command to form chains from block 3 receives indices of typical speakers and a description of their parameters from block 7, in particular, this is information about the frequency of sounds, which is individual for each speaker. If, for some reason, these statistics are not available, it is possible to replace them with statistics obtained on the basis of word processing. Naturally, the general statistics do not allow to fully simulate the parameters of the AP, but it becomes possible to work with voices, the data for which are not fully prepared.

The formation of sound chains is carried out taking into account the rules for following sounds (module 4.6), which ensures that, depending on the frequency of the sounds, chains of sounds are generated from pause to pause. The formed chains are fed to module 4.2, in which an intonation circuit (IR) is superimposed on each sound chain.

IR corresponds to the identifier of typical speakers whose indices were contained in block 3. In synthesizer mode, the identifier (number) of the intonation circuit is a parameter of the command. In the distribution generation mode (module 4.2), he himself selects the numbers of the intonation circuits, based on the information contained in the description of the AP (taken from block 7), and in the absence of it, from the general table. In the general case, the number of intonation circuits for each speaker will be different, and always even (one half for syntagmas with a hoot and the other without). Depending on the command and parameters, the IR assigns an IR identifier and a pause duration to each chain, after which the result is transmitted to the input of the module 4.3 for the name of allophones.

Allophones are named on the basis of naming rules (module 4.7), which replaces the names of sounds with the names (or names) of combinatorial allophones (or simply allophones), after which, depending on the command, the chains are transferred to module 4.4 or to blocks 4.5 and / or interface block 1.

Module 4.7 can be implemented as a subprogram that sets the rules for naming allophones, i.e. the order in which the name is formed, but based on the sound itself and the sounds located to the left and right of it (context). Each sound has a name that becomes the core of the name of a combinatorial allophone. Neighboring sounds give context names (naming rules module 4.7), which are added to the kernel on the left and right, respectively.

For vowel allophones, vowels with different reductions are considered different allophones and give different names for the nuclei. For consonants, different sounds are allophones of hard and soft sounds, giving carved names of kernels (and contexts). In general terms, the name of a combinatorial allophone is written like this:

<name of the left context> <name of the kernel> <name of the right context>.

Zeros for an allophone there is no information about the sound located to the left and / or to the right, then they are replaced by pauses, and the names of the contexts are determined based on the assumption of proximity to the pauses.

Module 4.4 works according to a well-known algorithm. In it, he ascribes to all sounds in the formed chain the durations and energy coefficients, after which the results of processing the chains of sounds are transferred to module 4.5.

Depending on the operating mode, different strategies for determining the duration of sounds are used. In the generation mode, samples of sound durations are determined randomly using random number sensors, and statistics on the distribution of the duration of each sound (which is part of the TD description and is taken from block 7).

In synthesizer mode, the durations of sounds are determined by a complex algorithm that takes into account both the values of statistics and the length of the chains and the order of the sounds in them.

Then in module 4.5, according to the well-known algorithm, the values of the fundamental tone for all vowels and voiced consonants contained in the chains are determined and attributed to them the parameters of the intonation circuits, which are part of the description of the AP and are taken from block 7. After that. as the last chain was worked out, in block 1 a message is issued about the completion of the command.

The obtained parameters (descriptions) of the sound chains with the assigned values of the fundamental tone (OT) and superimposed IR, depending on the parameters of the command, are transmitted to block 5 and / or to interface block 1.

Block 5 is a memory block that stores descriptions (prosodik) of sounds obtained as a result of their processing in block 4, so its operation does not cause difficulties in understanding.

Block 6 of the formation of the speech stream is also a complex one, consisting of separate modules that implement the algorithm of its operation. The control signal from the interface unit 1 enters the module 6.1, in which the command is analyzed, and reads the prosody (description of the chains of sounds) from block 5, and also, if necessary, receives the description of the AP from block 7, after which the formation of the speech stream taking into account the format specified interface unit 1.

Based on prosodiki (for one sound), it selects the marking of the sound signal from the TD description, generates the necessary sound duration from the labels, and depending on the type of sound (explosive, fricatives, sonants, vowels, affricates), various algorithms are used to modify the duration.

After all the marks necessary for inclusion in the resulting signal are determined, a specific value of the length of the period is assigned to each label (if the sound is vowel or voiced consonant). The lengths of the periods are determined by prosodic data and the basic value of the fundamental tone (OT) indicated in the description of the AP. After that, the list of tags is passed to module 6.2.

Module 6.2, having received a list of labels corresponding to the sound, the sound type and the identifier of the AP from module 6.1, extracts audio data from the description of the AP for each label, after which it brings the lengths of the periods to the given values. For this, the well-known PSOLA algorithm (TD-PSOLA) was used with some changes and additions. This algorithm is preferred. it provides a sufficiently high quality of the transformed voice and does not require significant computing resources.

Differences from the standard algorithm are that when raising and lowering the frequency response, different window functions are used. When reducing the length of the period (increasing the frequency), the following cosine direct and reverse windows are used:

and with an increase in the length of the period (decrease in the frequency frequency), the following linear forward and reverse windows are used:

NData - min (NOldData, NNewData);

Wca - weighting factors of the direct window;

Wcb - weighting coefficients of the reverse window;

i is the index of the coefficient;

NOldData - length of the initial period;

NNewData - the length of the formed period.

Another difference from the standard algorithm is that the samples of the generated period are normalized to the sum of the corresponding coefficients of the forward and reverse windows. In addition, the increase in the length of the period of the fundamental tone is performed not in one step, but in several.

After the durations of all periods are formed, labels, sound data and a description of the sound (prosodik) are transmitted to module 6.3. envelope formation.

To determine it, it uses only the data obtained from module 6.2 and the data generated at the previous stages of the model blocks. First, all samples of the signal are multiplied by the energy coefficient specified in prosodic module 4.4. Then additional energy coefficients for each period are additionally determined and applied in order to obtain a smooth transition of the level from the previous sound. After all the energy coefficients are applied, the maximum and minimum values of the amplitude of the last period, used to determine additional energy coefficients, are calculated and stored in the processing of the next sound.

The results (labels, sound data and sound description) are transmitted to module 6.4, in which the sounds are docked.

Module 6.4 uses only the data received from module 6.3 and the data stored in the previous processing step. If it is impossible to produce a morph of the previous and processed sound (for example, due to the fact that the latter has a short duration), then the stored sound data is given to module 6.5. If morph is possible. the stored sound data is mixed with the beginning of the processed sound, after which the sound data, excluding the part stored for docking with the next sound, is transmitted in module 6.5.

In module 6.5. the sound signal received from module 6.4 is brought to a predetermined format, the parameters of which are received in a command from the interface unit 1. The sound signal is brought to the sampling frequency and the number of bits assigned to each sample specified in the format, after which a report about the end of the formation of the speech signal ..

The database (block 7) is a complex database containing descriptions of APs and general tables. TD descriptions and general tables are created manually.

The country house database includes the following general tables: intonation circuits, energies, durations and frequencies of sounds;

Typical speaker descriptions include:

required components:

- base value of the frequency of the fundamental tone (OT);

- a sign of the sound base format (allophonic / suballophonic), and individual parameters of the audio signal processing algorithms;

- sound base, consisting of sound fragments corresponding to various sounds of Russian speech with the necessary markup for periods of OT, etc .;

- parameters for linking APs with population statistics.

optional components are tables:

- intonation types, energies, durations and frequencies of sounds;

To work with the model, all general tables must be filled in and a description of at least one AP created.

Sound base formation is a long and very laborious manual process. It is known that in Russian there are 6 vowels and 36 consonants. However, there will be many more allophones. So the nature of vowel reduction depends on its quality and rhythmic position. When forming a sound base It is necessary to take into account the position of the vowel with respect to stress. The following vowel reductions can be distinguished:

- stressed vowels “a0”, “o0”, “u0”, “e0”, “i0”, “у0”;

- the first pre-shock “a1”, and the pre-shock “a1”, “u1”, “e1”, “i1”, “y1”;

- second pre-shock "a2":

- shocked "a4". “O4”, “u4”, “e4”, “i4”. "Y4";

However, not only the rhythmic position affects the quality of the vowel. Phonetic realizations of sounds (both vowels and consonants) also depend on immediate contexts. Moreover, for stressed and unstressed vowels, different sets of contexts can be distinguished. Some sounds, which are different contexts for drums, are the same contexts for drums.

In addition to the fact that the fragment corresponding to the sound of speech must be found, highlighted and named, it is necessary to mark it accordingly. The markup differs for different sounds.

We give two examples of the application of the proposed statistical speech model, although there may be much more. So, on the basis of the described implementation, a system for synthesizing Russian speech in the text was developed (Fig. 5). The statistical model also found application as a test signal generator in a system for assessing the quality of vocoders and communication channels (Fig. 6).

Speech synthesis

The statistical model was used to build a system of high-quality text-based speech synthesis. High quality means high naturalness and intelligibility of synthesized speech.

The general scheme of the developed system can be represented by Figure 5. It contains: an operator, an interface module, a linguistic processor, and a statistical speech model.

The synthesis system operates as follows. The interface module requests a list of typical speakers (TD) from the statistical speech model. Then the operator selects one AP from the list, sets the audio signal format and enters the synthesized text. Then, the interface module transfers the synthesized text to the linguistic processor (LP). The LP divides the text into syntagmas, ascribes intonation types to them, stresses and transcribes them and transfers them to the statistical model.

The statistical model, based on the algorithms laid down in it, forms the prosodic and the sound stream. Sound stream and prosodik are transmitted to the interface module. At the request of the operator, the sound stream can either be played back by the sound card of the computer, or saved to a file for future use.

Prosodic data can be saved and opened for viewing and editing in a regular text editor.

Evaluation of vocoder quality

The statistical model is used as a source (generator) of a test signal for quality assessment, vocoders and communication systems.

The scheme for using the statistical model in the vocoder quality assessment system is shown in Figure 6. The system consists of: a vocoder, a quality assessment module, a statistical speech model, and files for storing audio streams (the original and processed vocoder).

The quality assessment module issues a command to the statistical speech model to generate a sound stream, with parameters specific to the population model being described. The sound stream generated by the model is saved to a file. The file is fed to the vocoder and to the quality assessment module. An audio signal that has passed the encoding and decoding procedures implemented by the vocoder is also supplied to the quality assessment module. The quality assessment module compares the signals and provides an estimate.

The advantages of the proposed utility model include the following:

- a larger volume of the original speech corpus;

- the use of a large number of classification features in the segmentation and description of the speech corpus;

- inclusion in the database of additional statistical information for each AP;

- the availability of descriptions of the intonation contours for each AP;

- the ability to simultaneously work with structural elements of different sizes and formats:

- accounting for statistical information in the formation of the sound stream;

- different approaches to changing the duration of sounds of different types, minimizing the distortion of the perceptual properties of sounds;

- different approaches to changing the frequency response of sounds of different types, minimizing the distortion of their perceptual properties;

- the ability to select sounds from the database (or even chains of sounds) that require the least modification;

- a variety of contextual implementations of sounds will allow us to synthesize a speech stream with high naturalness;

- The achievement of high naturalness will also be facilitated by the use of intonation contours, specially selected for each AP.

In addition, the statistical model of speech is language independent. The language with which the model will work is determined only by the data that the model is filled with, and all algorithms and interfaces will be preserved.

Currently, the proposed statistical model is being tested by using a text-based speech synthesis system and an objective assessment system for vocoder quality.

Claims (4)

1. A statistical model of speech, including an interface unit connected by corresponding inputs and outputs to a selection unit forming a sample of speakers, a unit for selecting sounds that select sounds and determine their parameters, with a unit for generating a speech stream that performs actions on elements of speech signals and with a database containing descriptions of typical speakers, which is also connected to the inputs of these blocks, characterized in that the parameter statistics module additionally included in the speaker selection block is For example, between the speaker selection block and the sound selection block, a sample speaker selection block is included, and between the sound selection block and the speech stream generation block, the prosodiki storage block is additionally included, and the rules for naming allophones and following sounds are added to the block for selecting sounds. 2. The statistical speech model according to claim 1, characterized in that the speaker selection block consists of modules: a speaker sample generator and population parameter statistics, the output of the population parameter statistics module being connected to the input of the speaker sample generator, the output of which is connected to the input of the sample block announcers. 3. The statistical speech model of claim 1, characterized in that the unit for selecting sounds consists of series-connected modules: forming chains, attributing an intonation contour, naming allophones, determining the duration, applying intonation contours, as well as modules of rules for following sounds and naming rules for allophones, moreover, the outputs of the last two modules are connected to additional inputs of the chaining module and the allophone naming module, and the additional output of the allophone naming module is connected to the output module for applying intonation circuits, the outputs of which are connected to the prosodiki storage unit and the interface unit. 4. The statistical speech model according to claim 1, characterized in that the speech stream generation unit comprises series-connected modules: duration formation, pitch changes, envelope formation, joint processing and reduction to a given format, wherein the duration formation module is connected to the output prosodiki storage unit, and the format conversion module with the output of an interface unit that sets the reduction format, as well as with the corresponding inputs of the specified block that take the sign of completion command and generated speech flow.