RU2119196C1 - Method and system for lexical interpretation of fused speech - Google Patents

Abstract

FIELD: automatic control and computer engineering; speech understanding systems, control systems for process equipment and robots, computer hardware and software, computer-aided speech translation, reference systems, etc. SUBSTANCE: omitting phonemic conversion level, lexical decoding network builds hypotheses about probable beginning, continuation, or end of words in speech and most probable sequences of standard words corresponding to delivered speech are compiled. Pronounced words may continuously follow each other in any order or may be separated by pauses or words which do not belong to assigned set of words. Proposed lexical decoding network is essentially integrated data base incorporating orthographic presentation of assigned set of words, expected acoustic presentation of assigned set of words in the form of sequence of standard meanings of speech signal parameter determining acoustic conditions and combining phonetic transcription, phonologic rules, and vocabulary for assigned set of words. System implementing this method has series-connected acoustic and lexical analyzers. Acoustic analyzer incorporates pre-processing unit, frequency spectrum analyzer, spectrum value storage buffer, weight coefficient and current acoustic condition calculators. Lexical analyzer has expected acoustic condition identifier, unit for comparison with standard, comparison estimate storage unit, control unit, optimal estimate selection and top marking unit, boundary top data base storage unit, check unit, probable top storage unit, local top data base storage unit, acoustic condition data base storage unit, word data base storage unit, lexical hypotheses shaper, lexical hypotheses storage unit, and output unit. EFFECT: improved accuracy and speed of interpreting fused Russian speech. 2 cl, 19 dwg

Description

 The invention relates to the field of automation and computer technology and can be used in speech understanding systems, process equipment control systems, robots, computer tools, automatic voice translation, in help systems, etc.

 A known method of lexical interpretation of continuous speech, implemented in the system of automatic understanding of English speech HEARSAY II [1].

 The essence of the method is that periodically utter a speech utterance, which is digitized at fixed time intervals with a given quantization frequency in this interval. Then, samples of this acoustic digitized signal are taken and, based on their totality, the current values of the parameters of the input speech signal are calculated, which transform the class of syllables, called the syllotype. After that, for each syllotype, when constructing the lexical hypothesis, all words that contain the stressed syllable belonging to this class of syllotypes are revealed. Polysyllabic words are rejected if they do not agree well with related syllotypes. The definition of syllables is based on the grouping of phonemes into phonetic classes. The pronunciation of each word that belongs to the pronunciation dictionary is converted into a sequence of syllotypes by distributing all phonemes by their classes. The sequence of parameter values of an unknown speech utterance defines hypotheses about the syllotypes used to construct hypotheses about words.

 A feature of the known method is that variations in the pronunciation of words are taken into account by applying wide classes of phonemes and including variants of pronunciation of words in the dictionary. Classes of phonemes suggest that each syllotype belongs to only one class of syllotypes.

 However, this method has drawbacks: it is impossible to separate syllables and phonemes strictly into classes, since there are phonemes that can be attributed to two neighboring classes. This leads to the fact that the differences between the classes are erased and the clarity of the difference between the syllabi is reduced, as a result of which the accuracy of the lexical interpretation of continuous speech is reduced.

 A known method of lexical interpretation of continuous speech is a feature, which consists in a direct transition from recognized sounds in the utterance to pronunciation of words, taking into account changes in these sounds during co-articulation. This method is implemented in the system of automatic speech understanding DRAGON [2].

 The essence of the method is that periodically utter a speech utterance, which is digitized at fixed time intervals with a given quantization frequency in this interval. Next, samples of this acoustic digitized signal are taken and, based on their totality, the current values of the parameters of the input speech signal are calculated, which are converted into a phoneme. After this, a sequence of phonemes is formed, and using a lexical decoding network, which is a model for pronouncing a word, hypotheses about possible words in a statement are built.

 To build a network of lexical decoding, a canonical pronunciation is taken and phonological rules are applied to it to represent the most complete probable pronunciation model of a word. When using the canonical pronunciation dictionary (the word subnet dictionary), each word subnet is replaced to the node. As a result, we get a network in which each node is an individual phoneme. Possible phonetic realizations of the word are formed by repeatedly applying phonological rules to the main pronunciation.

 Each rule provides an alternative pronunciation for a sequence of phonemes. For each phonological rule, the entire network is scanned to find any nodes that satisfy the context. All this leads to a decrease in the speed and accuracy of the lexical interpretation of continuous speech.

 Closest to the claimed method, taken as a prototype, is the method of lexical interpretation of continuous speech, implemented in the CASPERS system [3]. The essence of the method is that periodically utter a speech utterance, which is digitized at fixed time intervals with a given quantization frequency in this interval. Next, take samples of this acoustic digitized signal, the totality of which calculates the current values of the parameters of the input speech signal, which are converted into a phoneme. After this, a sequence of phonemes is formed, and, using the lexical decoding scheme, hypotheses about possible words in the utterance are built. In this case, the lexical decoding scheme is a tree containing all the expected phonetic realizations of the words of a given dictionary. Words that have the same first sounds are placed at the same starting point in the tree. Further, the end of each branch of the tree representing the pronunciation of the word is connected with all the initial forms of words, using a set of phonological rules. As a result, a network of phonetic solutions is created.

 The definition of the original expression is based on the search for the optimal sequence of phonemes in the network of phonetic solutions. Moreover, for accounting inside the dictionary phonological phenomena, as well as changes in word endings due to the influence of previous and subsequent words, the expected phonetic realizations of the word are represented by expanding the main pronunciation with several alternative pronunciations. Such a dictionary expansion is done automatically, using phonological rules.

 However, it is necessary to have some heuristic comparison strategy for selecting words corresponding to the phonetic notation of an unknown expression. To do this, it is necessary to introduce a penalty measure for erroneous identification, possible cases of adding or skipping sounds, since the automatic phonetic analyzer makes many errors of this type. Errors in phonetic transcription can ultimately lead to fatal mismatch with the correct word.

The disadvantages of the above methods and prototype are the low speed, lack of accuracy of the lexical interpretation of continuous speech, due to the following:
- phonetic transcription, which serves as input for constructing lexical hypotheses, contains substitution errors, extra sounds and missing sounds, reducing the similarity of the interpreted word to the correct hypothesis and increasing the similarity of the interpreted word to erroneous ones, especially when the dictionary is large;
- repeated application of phonological rules to the dictionary of pronunciation of words entails a slowdown in the process of lexical interpretation of continuous speech;
- the expected phonetic realization of a word depends on the context of the sentence in which it occurs. Word boundaries in continuous speech are completely absent in transcription, since the acoustic signs of their positions are poorly expressed;
- the position of the boundary between the durations of the phonetic groups depends on the speed of speech, the position of the syntactic boundaries, stressed syllables and the local phonetic environment.

 In the lexical interpretation of continuous speech, a problem arises, the essence of which is that making decisions on phonetic damage partially depends on a higher level factor that cannot be determined until decisions are made on the phonetic level. The solution to this problem boils down to the need to make decisions at the phonetic and higher levels simultaneously.

 The description of the proposed method of lexical interpretation of continuous speech includes eight figures: FIG. 1 is a general representation of the vowel phoneme; FIG. 2 - spelling and phonetic representation of tokens, FIG. 3 - modeling graph; FIG. 4, 5 - graph of alternative representations; FIG. 6 - a network of alternative representations; FIG. 7 - network lexical decoding; FIG. 8 is an example of a data structure.

 The proposed method of lexical interpretation of continuous speech consists in periodically uttering a speech utterance, which is digitized at fixed time intervals with a given quantization frequency in this interval. Then take samples of this acoustic digitized signal, the totality of which calculates the current values of the parameters of the input speech signal that determine the current acoustic state.

 The method is characterized in that, bypassing the level of the phonemic transformation, simultaneously using the calculated values of the parameters of the input speech signal, using the lexical decoding network, hypotheses are built about the possible beginning, continuation, or end of words in the speech utterance, and the most probable sequences of reference words corresponding to the spoken are made up speech utterance. In this case, the spoken words can continuously follow each other in any order, or can be separated by pauses, or by words that do not belong to a given set of words. The proposed lexical decoding network is an integrated database containing spelling representations of a given set of words, expected acoustic representations of a given set of words in the form of sequences of reference values of speech signal parameters that determine acoustic states and combines phonetic transcription, phonological rules and vocabulary for a given set of words.

 The result of the invention is to increase the accuracy of the lexical interpretation of the unified speech of the Russian language and ensuring performance as close as possible to real time. The result is achieved using the lexical decoding network (SLD), the lexemes of which are presented as a sequence of acoustic states (AS), taking into account the dictionary phonetic phenomena, as well as phonetic phenomena that occur at word boundaries.

 It is proposed to call the acoustic state a set of values of the parameters of a speech signal (RS), which characterizes a time interval commensurate with the period of the fundamental tone. The essence of this approach is to present the RS with a finite number of pre-selected types of speakers. The number of different speakers should be chosen in such a way as to reflect the entire significant variety of impulse responses of the vocal tract during speech formation.

 The acoustic representation of continuous speech signals is based on the principle of sequential decomposition of phonemes into allophones, and allophones into their constituent speakers. Allophones are easily distinguishable acoustically, as a result of which the need to apply the rules at lower levels disappears. They contain information about the boundaries between syllables and words. It is proposed that such information be obtained by representing allophones in the form of three consecutive speakers: initial, middle and final. In this case, the type of middle speaker depends only on the type of allophone selected, and the type of initial or final transitional speaker depends, moreover, on the type of previous and subsequent phonemes.

 A different degree of detail of the decomposition of each phoneme into allophones, and of allophones into AS is possible. As an example, we present one of the possible decomposition options, which is sufficient to provide the necessary variety of realizations of each phoneme and allophone in the lexical interpretation of the coherent speech of the Russian language.

а также соответствующие им множества назализованных

Для русских согласных необходимо различать губное, зубное, альвеолярное, велярное и латеральное место образования. Таким образом, для описания переходных (начального или конечного) интервалов РС каждого аллофона гласной необходимо иметь до 5-ти различных типов АС. Общее представление каждой гласной фонемы в виде набора АС, необходимых для акустического описания слитной речи русского языка, представлено фиг. 1 на примере гласной /А/. Аналогичным образом предлагается определять три временных интервала РС (начальный, серединный и конечный) для описания согласных звуков.Of the Russian vowel phonemes, we choose a lot of allophones hard - {A, O, Y, E, I, S} and soft -

as well as the corresponding sets of nasalized

For Russian consonants, it is necessary to distinguish between the labial, dental, alveolar, velar and lateral places of formation. Thus, to describe the transition (initial or final) MS intervals of each vowel allophone, it is necessary to have up to 5 different types of speakers. A general representation of each vowel phoneme in the form of a set of speakers necessary for the acoustic description of the continuous speech of the Russian language is presented in FIG. 1 by the example of the vowel / A /. Similarly, it is proposed to define three time intervals of the RS (initial, middle and final) for the description of consonants.

 The representation of Russian phonemes in the form of AS allows a significant change in their number, which is due to the effect of co-articulation with the previous and subsequent phonemes.

множества фонем m - го и n - го типов, где q - индекс, определяющий тип АС, q = 1, 2, 3 (q = 1 - начальное АС; q = 2 - серединное АС; q = 3 - конечное АС); φ - индекс, определяющий фонему φ = 1, 2, ..., Ф ; m - индекс, определяющий множество предшествующих фонем, m = 1, 2, ..., M; n - индекс, определяющий множество последующих фонем, n = 1, 2, ..., N. Тогда в общем случае АС можно представить в виде многозначной функции:

Формула (1) приобретает конкретный вид для каждой фонемы. Проиллюстрируем это на примере фонемы /А/ для трех АС. Для начального АС:

где: Ω₁ = {П, Б, Ф, В, Л} - множество твердых губных и боковых согласных фонем, Ω₂ = {Т, Д, С, З, Р, Ц, Ч, Ж, К, Г, Х} - множество твердых зубных, альвеолярных и небных согласных фонем, Ω₃ = {П', Б', Ф', В'} - множество мягких губных согласных, Ω₄ = { Т', Д', С', З', Р', Ш', Ч'} - множество мягких зубных и альвеолярных согласных фонем, Ω₅ = {К', Г', Х'} - множество мягких небных согласных фонем, Ω₆ = {Л'} - единичное множество мягких боковых согласных фонем, Ω₇ = {М} - единичное множество твердых губных носовых согласных фонем, Ω₈ = { Н} - единичное множество твердых зубных носовых согласных фонем, Ω₉ = { М'} - единичное множество мягких губных носовых согласных фонем, Ω₁₀ = { Н'} - единичное множество мягких зубных носовых согласных фонем, Ω₁₁ = {А, О, У, Э, И, Ы, #} - множество гласных фонем и паузы.Let V $\genfrac{}{}{0ex}{}{\text{q}}{\text{φ}}$ - AS required to determine the qth interval of the PC φ_th phoneme; ω _φ-1 , ω _{φ + 1} - previous and subsequent phonemes;

sets of phonemes of the mth and nth types, where q is the index defining the type of AS, q = 1, 2, 3 (q = 1 is the initial AS; q = 2 is the middle AS; q = 3 is the final AS); φ is the index defining the phoneme φ = 1, 2, ..., Ф; m is the index defining the set of previous phonemes, m = 1, 2, ..., M; n is the index defining the set of subsequent phonemes, n = 1, 2, ..., N. Then, in the general case, AS can be represented as a multi-valued function:

Formula (1) takes on a specific form for each phoneme. We illustrate this with the phoneme / A / for three speakers as an example. For the initial speaker:

where: Ω ₁ = {П, Б, Ф, В, Л} is the set of solid labial and lateral consonant phonemes, Ω ₂ = {Т, Д, С, З, Р, Ц, Ч, Ж, К, Г, Х } - the set of hard dental, alveolar and palatine consonant phonemes, Ω ₃ = {P ', B', Ф ', В'} - the set of soft labial consonants, Ω ₄ = {T ', D', C ', Z', P ', W', H '} - the set of soft dental and alveolar consonant phonemes, Ω ₅ = {K', G ', X'} - the set of soft palatine consonant phonemes, Ω ₆ = {L '} - the unit set of soft lateral consonant phonemes, Ω ₇ = {M} is the unit set of hard labial nasal consonants, Ω ₈ = {H} is the unit set of solid dental nasal consonants phonemes, Ω ₉ = {M '} is the unit set of soft labial nasal consonant phonemes, Ω ₁₀ = {H'} is the unit set of soft labial nasal consonant phonemes, Ω ₁₁ = {A, O, Y, E, I, S , #} - a lot of vowel phonemes and pauses.

где:
Ω₁₂ = {А, О, У, Э, И, Ы, Л, Р, В, З, Ж, Б, Д, П, Т, Г, Ф, К, С, Ш, Х, Ц} - множество твердых неносовых согласных и гласных фонем, Ω₁₃ = {Л', Р', В', З', Ж', Б', Д', Г', П', Т', К', Ф', С', Ш', Х', Ч'} - множество мягких согласных фонем, Ω₁₄ = {М', Н'} - множество мягких носовых согласных фонем, Ω₁₅ = {М, Н} - множество твердых носовых согласных.For the middle speaker phoneme / A / formula (1) has the form:

Where:
Ω ₁₂ = {A, O, Y, E, I, S, L, P, B, H, G, B, D, R, T, G, F, K, C, W, X, C} - the set solid non-nasal consonants and vowel phonemes, Ω ₁₃ = {L ', P', B ', Z', F ', B', D ', G', P ', T', K ', F', C ', W ', X', H '} - the set of soft consonant phonemes, Ω ₁₄ = {M', H '} - the set of soft nasal consonant phonemes, Ω ₁₅ = {M, H} - the set of hard nasal consonants.

где
Ω₁ - Ω₁₀ те же, что и в формуле (2), а множества Ω₁₆ - Ω₂₀ являются единичными и содержат соответственно гласные - {А, О, У, Э, И, Ы}.For the final AS phoneme / A / formula (1) has the form:

Where
Ω ₁ - Ω ₁₀ are the same as in formula (2), and the sets Ω ₁₆ - Ω ₂₀ are single and contain respectively vowels - {A, O, Y, E, I, S}.

 Similarly to formulas (2), (3), (4), for each phoneme, corresponding expressions can be written taking into account the rules of their allophonic variability.

 Based on the foregoing, SLD is formed. The formation of SLD occurs by performing a sequence of operations: creating a database of words; presentation of the speech utterance as a sequence of words, determination of the acoustic state as a set of parameter values for the time interval of the MS; creation of a database of standards of acoustic states for the phonetic and phonological description of Russian words; representation of a word as a sequence of acoustic states. The essence of these operations is as follows.

 1) Create a database of words necessary for verbal communication, containing the number of the word - l, for which are defined: spelling representation, pronunciation options with the corresponding numbers - j.

2) A speech utterance is represented by a sequence of words allowing continuous spoken words to be spoken one after another in any order, either with separation by pauses, or with separation of words that do not belong to a given set (database) of words:
W = c $\genfrac{}{}{0ex}{}{\text{j}}{\text{l, 1}}$ , ..., C $\genfrac{}{}{0ex}{}{\text{j}}{\text{l, i}}$ , ..., C $\genfrac{}{}{0ex}{}{\text{j}}{\text{l, I}}$ (5)
Where:
W is a speech utterance; C is the word;
l is the word number in the database of words l = 0, 1, 2, ..., L;
j is the pronunciation number of the lth word, j = 0, 1, 2, ..., J;
i is the serial number of the word in the statement, i = 1, 2, 3, ..., I;
3) Determine the acoustic state as a set of values of the parameters of the time interval of the PC:
V = (x ₁ , x ₂ , x ₃ , ... x _R ) + Q, (6)
where, for example: x ₁ = F ₀ is the frequency of the fundamental tone; x ₂ = A ₀ is the amplitude of the fundamental tone; x ₂ = F ₁ , x ₄ = F ₂ , where F ₁ , F ₂ are the frequencies of the formants; x ₅ = A ₁ , x ₆ = A ₂ are the amplitudes of the first and second formants, respectively; x ₇ = B ₁ , x ₈ = B ₂ - transmission bandwidth of the first and second formant, respectively; x ₉ = Z is the number of transitions through zero; x ₁₀ - rate of pronunciation, etc .; Q is the noise.

 4) Create a database of standards of acoustic states containing the number of speakers, the name of the speaker with a set of parameter values for the time interval of the speech signal.

где
0≤h≤H, 0≤l≤L, 1≤k(j)≤K, 0≤j≤J, 1≤i≤I, (8)
C - слово;
V - акустическое состояние;
i - порядковый номер слова в высказывании, i = 1, 2, 3, ...,I;
h - номер АС в базе данных эталонов АС, h = 0, 1, 2, ..., H;
l - номер слова в базе данных слов, l = 0, 1, 2, ..., L;
j - номер произношения l-го слова, j = 0, 1, 2, ..., J;
b - тип начального АС, выбираемый в соответствии с формулами (1), (2) и в соответствии с произношением j для l-го слова;
e - тип конечного АС, выбираемый в соответствии с формулами (1), (4) и в соответствии с произношением j для l-го слова;
q - индекс, определяющий тип АС, выбираемый в соответствии с формулами (1) - (4) и в соответствии с произношением j для l-го слова; q = 1, 2, 3;
k - число акустических состояний в слове, изменяющееся в зависимости от j для C $\genfrac{}{}{0ex}{}{\text{j}}{\text{l}}$ , k = 1, 2, 3, ..., K;
Если i=1, то речевое высказывание состоит из одного слова. Тогда:
W = C $\genfrac{}{}{0ex}{}{\text{j}}{\text{l,1}}$ ; (9)

где
V $\genfrac{}{}{0ex}{}{\text{b}}{\text{h,l,1}}$ - акустическое состояние, связанное с описанием перехода от паузы к началу первой фонемы C $\genfrac{}{}{0ex}{}{\text{j}}{\text{l,1}}$ - го слова;

последовательность акустических состояний C $\genfrac{}{}{0ex}{}{\text{j}}{\text{l,1}}$ - го слова;
V $\genfrac{}{}{0ex}{}{\text{e}}{\text{h,l,K}}$ - акустическое состояние, связанное с описанием перехода от конца последней фонемы C $\genfrac{}{}{0ex}{}{\text{j}}{\text{l,1}}$ -го слова к паузе;
Если 1< i≤l, то допустимые V $\genfrac{}{}{0ex}{}{\text{b}}{\text{h,l,1}}$ - начальные и V $\genfrac{}{}{0ex}{}{\text{e}}{\text{h,l,K}}$ - конечные АС определяют с учетом всех возможных (грамматически правильных и неправильных) последовательностей слов из базы данных слов. При этом для определения допустимых V $\genfrac{}{}{0ex}{}{\text{b}}{\text{h,l,1}}$ - начальных АС учитывают фонологические явления, которые могут образоваться из-за влияния всех возможных "предыдущих" слов на C $\genfrac{}{}{0ex}{}{\text{j}}{\text{l,i}}$ - ое слово. Для определения допустимых V $\genfrac{}{}{0ex}{}{\text{e}}{\text{h,l,K}}$ - конечных АС учитывают фонологические явления, которые могут образоваться из-за влияния всех возможных "последующих" на C $\genfrac{}{}{0ex}{}{\text{j}}{\text{l,i}}$ - ое слово.5) Represent words as a sequence of acoustic states:

Where
0≤h≤H, 0≤l≤L, 1≤k (j) ≤K, 0≤j≤J, 1≤i≤I, (8)
C is the word;
V is the acoustic state;
i is the serial number of the word in the statement, i = 1, 2, 3, ..., I;
h is the number of speakers in the database of speaker standards, h = 0, 1, 2, ..., H;
l is the word number in the database of words, l = 0, 1, 2, ..., L;
j is the pronunciation number of the lth word, j = 0, 1, 2, ..., J;
b - the type of initial speaker, selected in accordance with formulas (1), (2) and in accordance with the pronunciation of j for the l-th word;
e is the type of final speaker selected in accordance with formulas (1), (4) and in accordance with the pronunciation of j for the lth word;
q is the index defining the type of speaker selected in accordance with formulas (1) - (4) and in accordance with the pronunciation of j for the l-th word; q is 1, 2, 3;
k is the number of acoustic states in the word, which varies depending on j for C $\genfrac{}{}{0ex}{}{\text{j}}{\text{l}}$ , k = 1, 2, 3, ..., K;
If i = 1, then the speech utterance consists of one word. Then:
W = c $\genfrac{}{}{0ex}{}{\text{j}}{\text{l, 1}}$ ; (nine)

Where
V $\genfrac{}{}{0ex}{}{\text{b}}{\text{h, l, 1}}$ - acoustic state associated with the description of the transition from pause to the beginning of the first phoneme C $\genfrac{}{}{0ex}{}{\text{j}}{\text{l, 1}}$ - th word;

sequence of acoustic states C $\genfrac{}{}{0ex}{}{\text{j}}{\text{l, 1}}$ - th word;
V $\genfrac{}{}{0ex}{}{\text{e}}{\text{h, l, K}}$ is the acoustic state associated with the description of the transition from the end of the last phoneme C $\genfrac{}{}{0ex}{}{\text{j}}{\text{l, 1}}$ words to pause;
If 1 <i≤l, then admissible V $\genfrac{}{}{0ex}{}{\text{b}}{\text{h, l, 1}}$ - initial and V $\genfrac{}{}{0ex}{}{\text{e}}{\text{h, l, K}}$ - final speakers are determined taking into account all possible (grammatically correct and incorrect) sequences of words from the database of words. Moreover, to determine the permissible V $\genfrac{}{}{0ex}{}{\text{b}}{\text{h, l, 1}}$ - initial speakers take into account phonological phenomena that can be formed due to the influence of all possible “previous” words on C $\genfrac{}{}{0ex}{}{\text{j}}{\text{l, i}}$ - th word. To determine the permissible V $\genfrac{}{}{0ex}{}{\text{e}}{\text{h, l, K}}$ - final speakers take into account phonological phenomena that can be formed due to the influence of all possible “subsequent” ones on C $\genfrac{}{}{0ex}{}{\text{j}}{\text{l, i}}$ - th word.

Thus, each word contains three sections in a speech utterance: initial, middle and final. Moreover, for a fixed value i = 1, the speech utterance consists of one word and contains the start and end portions of the speech utterance associated with a pause, and for 1 <i≤I the word C $\genfrac{}{}{0ex}{}{\text{j}}{\text{l, i}}$ contains the initial and final sections of the speech utterance associated with either the “previous” and “subsequent” words, or with pauses, respectively.

 6) Produce descriptions of transitions from acoustic states using a set of phonetic and phonological rules of the Russian language and p.1-p.5.

 7) Create a network of lexical decoding taking into account clause 5, clause 6, with the subsequent formation of a database of local vertices and a database of boundary vertices.

 The stages of building a lexical decoding network are represented by six figures: FIG. 2 - spelling and phonemic representation of tokens, FIG. 3 - a modeling graph whose vertices are allophones, and arcs are pointers to the following possible allophones; FIG. 4, FIG. 5 is a graph of alternative representations, the vertices of which are AS, and the arcs are pointers to the following possible AS; FIG. 6 - a network of alternative representations, the vertices of which are AS, and the arcs are pointers to the following possible AS; FIG. 7 - a network of lexical decoding, the vertices of which are speakers, and the arcs are pointers to the following possible acoustic states.

 The stages of constructing a lexical decoding network are presented using an example of expressions applicable to control the capture motion of the manipulator left and right, as well as indications of the output of the capture program. For example, “Take capture to the left,” Output capture to the right, “Show capture program,” “Show capture output program,” etc.

 At the first stage (Fig. 2) determine the necessary dictionary for verbal communication. The spelling and phonemic representation of each token is determined. At the second stage (Fig. 3), for each token with possible endings, a modeling graph of the expected allophonic representations is constructed, the vertices of which are allophones, and the arcs indicate the next possible allophones. In FIG. 3, rectangles indicate the vertices with the names of allophones, and the numbers indicate the numbers of tokens corresponding to the numbers of tokens from FIG. 2. After this, the sequence of allophones is replaced by a sequence of acoustic states (Fig. 4, Fig. 5) for all lexical units of the dictionary used with possible endings and build them in the form of a decision tree. In this case, words having the same first sounds are placed in the same initial vertex of the tree. For example, the words “show” and “program” have the first common sound - “p”. Further, all possible endings of each word are connected to the root of the tree and, using phonological rules, a network of alternative representations is constructed for all possible (grammatically correct and incorrect) sequences of words from the dictionary. A fragment of a network of alternative representations is depicted in FIG. 6.

 As a result of applying phonological rules, local and boundary vertices are formed.

The local vertex should be considered an object associated with AS type V $\genfrac{}{}{0ex}{}{\text{q}}{\text{h, l, k (j)}}$ . Moreover, each local vertex, as an example, is assigned: the number of the vertex; speaker number in the speaker database; binary border code; binary pause code; an array (list) of words belonging to the AS of a given vertex with a sign of the end of each of the list of words; an array (list) of subsequent possible vertices, a pointer to a database of boundary vertices.

The boundary vertex should be considered the local vertex associated with transitional speakers of type V $\genfrac{}{}{0ex}{}{\text{b}}{\text{h, l, 1}}$ and v $\genfrac{}{}{0ex}{}{\text{e}}{\text{h, l, K}}$ formed at the beginning and end of each word, respectively, in which the binary boundary code is not equal to zero. In this case, the boundary vertex is not assigned an array (list) of words associated with the AS of this vertex.

 At the final stage of building an SLD, the root of the network of alternative representations is connected to all boundary vertices.

 Thus, they get SLD, which is a dictionary with a built-in phonetic transcriptor, phonology rules and vocabulary for a given set of words. A fragment of a lexical decoding network is shown in FIG. 7. In FIG. 4 - FIG. 7 rectangles indicate the vertices with the names of the speakers, and the numbers indicate breaks in the connections.

 In accordance with FIG. 7 the initial (root) vertex is a pause. Each vertex in the SLD column is an object associated with one section of a quantized statement (phrase). Each vertex in the second column contains ASs associated with the following possible states, etc. Each vertex admits a transition into itself and bypassing itself (in Fig. 7 this is not shown so as not to clutter up the circuit). This leads to the fact that two or more vertices can be associated with the same AS. Thus, in the process of isolating V, additional ASs may arise, while the absence of ASs leads to significant problems. Therefore, potentially absent speakers should be considered as additional in the process of creating an SLD.

 Such a network explicitly takes into account the co-articulation effects that occur both inside words and at their borders, and allows, bypassing the level of phonetic transformation, to form possible variants of the lexical interpretation of the input statement of continuous speech.

 To determine the possible variants of the lexical interpretation of the original expression, it is necessary to find the optimal sequence of vertices (path) in the SLD. SLD uses a vocabulary representation that combines the common parts of different words. Therefore, the procedure for viewing the entire dictionary is easily implemented from a computational point of view and does not require a separate consideration of each word. At the same time, acoustic-phonetic knowledge is manifested in a convenient and accessible form that simplifies the process of optimizing the choice of the best path.

 Based on the foregoing, a database of local and a database of boundary vertices is created. In this case, each vertex is assigned a weight coefficient η based on the AS. Next, vertices are classified by increasing the weight coefficient η (with the corresponding renumbering). As a result, the number of the boundary vertex in the database of boundary vertices (BDGV) determines the number of vertices in the database of local vertices (BDLV). An example of the data structure used in the WLD is shown in FIG. eight.

 The proposed method of lexical interpretation of continuous speech, based on the use of SLD, implements a sequential reduction of the initial set of speakers and words according to the criteria of acoustic similarity.

Its essence is as follows. A speech utterance is uttered, which is digitized at fixed time intervals with a given quantization frequency in this interval. Next, take samples of this acoustic digitized signal, the totality of which calculate the weight coefficient η. This coefficient determines the probable search area for vertices in the BDLV. At the same time, the current AS - V _T is calculated from the obtained set of samples. Find a vertex in BDLV with reference AC - V _e, V _T like. If V _{T is} not similar to the reference AS of the expected vertices in the BDLF, then the correction of the search area of the expected vertices in the BDLF is performed. If it is not possible to find the reference speakers of the expected vertices in the BDLVs like V _T , then the search is performed in the BDGVs. If vertices with reference speakers similar to V _T are found, then according to estimates of the proximity measure of the current speaker and expected standards, hypotheses are formed about words that are acoustically similar in their initial speakers to the current one. After that, from the set of hypotheses about words that have been formed at this point, samples are selected that are acoustically similar in their next reference speakers to the next current speaker. In this case, sequences of words are formed taking into account the alternation of boundaries that make up their words in accordance with (7), (10). If it is not possible to find reference speakers like V _T , neither in BDLV, nor in BDGV, then add the corresponding labels in the generated word sequences, which indicate not found speakers and their corresponding words. This process continues until an interphrase pause is detected. The sequences of words obtained at this point constitute a set of possible lexical hypotheses or variants of the lexical interpretation of the input statement. This set of lexical hypotheses can be subjected to further analysis according to grammatical, syntactic, semantic and pragmatic criteria.

 Description of the system of lexical interpretation of continuous speech (SLISR) of the Russian language that implements the proposed method includes eleven figures; FIG. 9 is a structural diagram of a system, FIG. 10 is a block diagram of an acoustic analyzer unit; FIG. 11 is a block diagram of a lexical analyzer block; FIG. 12 - FIG. 19 is a flowchart of the SLISR operation algorithm.

 The lexical interpretation system of continuous speech using the SLD and the data structure shown in FIG. 8 is shown in FIG. 9. It consists of an acoustic analyzer represented by block 1 and a lexical analyzer represented by block 2. The system allows one to form variants of possible sequences of words corresponding to the uttered statement based on information about the sequence of detected acoustic states.

 Block 1 is designed to determine acoustic states in sound signals and contains two inputs and two outputs.

 Block 2 is designed to determine words from a given dictionary acoustically similar to spoken and contains two inputs and two outputs. The input 1 of block 1 is connected to the microphone, and the input 2 is connected to the output 2 of block 2. The outputs 1 and 2 of block 1 are connected to the inputs 1 and 2 of block 2, respectively. From the output 1 of block 2 get the desired result.

Block 1, the block diagram of which is shown in FIG. 10, contains: block 3 - preprocessing, block 4 - frequency spectrum analyzer, block 5 - buffer memory values of the spectrum, block 6 - calculator weight coefficient η, block 7 - calculator of the current acoustic state V _T.

 Block 2, the block diagram of which is shown in FIG. 11, contains: block 8 - determinant of expected acoustic states, block 9 - comparison with the standard, block 10 - memory block 1, block 11 - control unit, block 12 - block for selecting the optimal assessment and marking of vertices, block 13 - database storage block boundary vertices, block 14 — check block, block 15 — memory block 2, block 16 — block of storage of the database of local vertices, block 17 — block of storage of the database of acoustic states, block 18 — block of storage of the database of words, block 19 — lexical shaper hypotheses, block 20 - memory block 3, block 21 - output block.

 Block 3 is designed to digitize and filter acoustic signals.

 Block 8 is intended for receiving data from block 6, block 11 and block 16, organizing data requests in blocks 13 and 16, as well as issuing data related to the determination of the following possible vertices with their numbers and AC parameters.

 Block 9 is designed to calculate an estimate of the degree of coincidence between the acoustic characteristics of the expected AC standards and the current portion of the speech signal.

 Block 10 is intended for temporary recording, storage, reading and transmission of estimates of the degree of coincidence between the acoustic characteristics of the expected AC standards and the current section of the speech signal, as well as the vertices to which they belong.

 Block 11 is designed to generate data requests about vertices using blocks 8, 12, 14, 15, 19, as well as control blocks 10, 15, 20.

 Block 12 is designed to select the best estimate of the degree of coincidence, with the corresponding vertex numbers available to block 10, as well as marking the vertices.

 Block 14 is designed to check the peaks for the content of non-zero values of binary codes "pause" and "border".

 Block 15 is designed to temporarily record, store, read and transmit possible vertices with acoustic states similar to the current portion of the speech signal.

 Block 19 is intended to form variants of sequences of words acoustically similar to a spoken utterance.

Block 20 is intended for temporary recording, storage, reading and transmission of variants of sequences of words (lexical hypotheses) acoustically similar to a pronounced utterance. \
Block 21 is designed to display the results of the lexical interpretation of continuous speech.

 The work of the lexical interpretation system of continuous speech is as follows (see Fig. 9, 10). The input statement from the microphone is fed to the input of block 3 of the acoustic analyzer 1. Block 3 converts the input signals into digital form and performs their filtering. Next, the signals from the output of block 3 are fed to the input of block 4 to highlight the frequency spectrum. The signals from the output of block 4 are fed to the input of block 5. From the output of block 5, the signals are fed to input 1 of block 6 and input 1 of block 7.

Block 6 calculates the weight coefficient η used to search for the input vertex of the first column of the SLD (see Fig. 11, Fig. 7). The calculated value of the weight coefficient from the output 2 of block 6 goes to the input 2 of block 8. From the output of 2 block 8, the value of the weight coefficient η goes to the input 3 of block 16. From the output of 3 block 16, the number of the nearest vertex goes to the input 3 of block 8. Next, the block 8 determines the vertex numbers n _min and n _max , respectively indicating the upper and lower boundaries of the region in which it is necessary to search for the initial acoustic state. After this, block 8 generates data requests for vertices whose numbers belong to the search area for the initial acoustic state and sends them from its output 2 to input 3 of block 16. Based on the received vertex numbers, block 16 determines the corresponding numbers of reference acoustic states V _,, their names and values speaker parameters. Block 16 from its output 3 supplies this data to the input 3 of block 8. At the output 4 of block 8, a permission signal is generated that is input to input 2 of block 6. In turn, block 6 at output 1 forms a permission signal that goes to input 2 of block 7.

Block 7 calculates the current values of the parameters of the AC state V _T and from its output feeds them to the input 1 of block 9. At the same time, block 8, having determined the values of the parameters of the expected standards AC - V _E , together with the corresponding numbers of vertices, from its output 1 in sequence, starting with V _e with the vertex number n _min feeds them to input 2 of block 9.

 Block 9 calculates an estimate of the α degree of coincidence of the current and reference acoustic state. The value of this estimate, together with the corresponding vertex number, from the output of block 9 goes to input 1 of block 10.

Block 11 checks the contents of block 10 to reach the upper limit of the search region n _max . If n _{max is} not reached, then there is a further comparison of the expected speakers with the current one. If n _{max is} reached, then the output 6 of block 11 transfers the data contained in block 10, which are received at the input of block 12.

Block 12 checks the data coming from block 10 through block 11 for an estimate of α that exceeds the threshold value ε. If no such estimate is found, then block 12 analyzes the increase (decrease) of α with increasing n _min . After that, at the output 2 of block 12, signals are generated that change the boundaries of the search area, which are fed to the input 5 of block 11. Block 11, changing the boundaries of the search at its output 1, generates a control signal that goes to input 2 of block 10 and zeroes the contents of the block 10. At the same time, at the output 1 of block 11, a signal is generated that enters the input 1 of block 8, which allows the determination of the next possible vertex. In the case when _VE acoustically similar V _T block 8 in block 16 is not found, then block 8 searches them in block 13.

If the estimate α exceeding the threshold value ε is not found and the limits on the permissible search region in block 13 and block 16 are exceeded, then block 12 at output 1 generates data defining V _T as the previous AS of the vertex with the name previous. In this case, the node with the name previous is assigned the label of the unknown speaker. At the output 2 of block 12, a signal is generated informing of an unknown speaker, which is fed to input 5 of block 11. Block 11 at output 5 generates a control signal that is fed to input 2 of block 15. Based on this signal, block 15 at input 1 receives information from the output 1 of block 12 through block 14.

 Block 11 from output 5 sends a signal to input 2 of block 15. Block 11 at input 3 receives data from output 2 of block 15. Next, block 11 checks this data for the contents of the vertex named previous.

If the vertex with the name previous does not contain information about V _акуст acoustically similar to V _T , then block 11 at output 2, having prepared a signal for processing the next (new) part of the PC, feeds it to input 1 of block 8. In this case, the output 4 of block 8 forms a permission signal, which is input to input 2 of block 6. Block 6 from output 2, after calculating the weight coefficient η, feeds it to input 2 of block 8. From the output 2 of block 8, the value of the weight coefficient η goes to input 3 of block 16. From output 3 of block 16, input 3 of block 8 receives the number of the nearest vertices . Next, block 8 determines the vertex numbers n _min and n _max , designating respectively the upper and lower boundaries of the region in which it is necessary to search for the initial acoustic state. After this, block 8 generates data requests for vertices whose numbers belong to the search area for the initial acoustic state and sends them from its output 2 to input 3 of block 16. Using the received vertex numbers, block 16 determines the corresponding numbers of reference acoustic states - V _Э , their names and AC parameter values. Block 16 from its output 3 supplies this data to the input 3 of block 8. At the output 4 of block 8, a permission signal is generated that is input to input 2 of block 6. In turn, block 6 at output 1 forms a permission signal that goes to input 2 of block 7.

Block 7 calculates the current values of the parameters of the AC state V _T and from its output feeds them to the input 1 of block 9. At the same time, block 8, having determined the values of the parameters of the expected standards AC - V _E , together with the corresponding numbers of vertices, from its output 1 in sequence, starting with V _e with the vertex number n _min feeds them to input 2 of block 9.

If the vertex with the name previous contains information about V _акуст acoustically similar to V _T , then block 11 at output 2, having prepared a signal for processing the next (new) section of the PC, feeds it to input 1 of block 8. In this case, at the output 4 of block 8, a permission signal is generated that is input to block 2 input 6. In turn, block 6 at output 1 forms a permission signal that goes to block 2 input 7.

Simultaneously, the block 7 calculates the current value of acoustic states V _T, and unit 8 determines the expected peaks V _E, the following vertex for the previous parameters in block 16. The values of V _e, together with the corresponding numbers of vertices with Released 1 unit 8 sequentially input to block 9 2 , and the value of V _T from the output of block 6 goes to input 1 of block 9.

If block 12, after checking the contents of block 10, finds an estimate of α that exceeds the threshold value of ε, then block 12 takes the value of α as optimal - α _opt . In this case, block 12 redefines the vertex number corresponding to α _opt as optimal - n _opt , marks the value n _opt with the current name and transfers it from output 1 to input 1 of block 14. Block 14 at output 2 forms a request for binary vertex codes and sends it to input 2 of block 16. From output 1 of block 16, input 2 of block 14 receives the value of the vertex number by which the request was made, as well as the values of the binary codes “pause” and “border”. Block 14 checks the values of the binary codes at the incoming vertex. After that, at the output 3 of block 14, a signal is generated that goes to the input 4 of block 11, by which the block 11 from the output 5 sends a signal to the input 2 of the block 15, allowing the block 15 at the input 1 to receive data from the output 1 of the block 14. Block 19 input 2 reads data from output 1 of block 15.

 If the current peak is not a pause, then block 19 from output 2 supplies the value of the vertex number to input 1 of block 16, by which block 16 from output 2 sends the vertex number to the input 3 of block 19, as well as a list of words with the corresponding signs of their end with this number. Block 19 on the basis of the received words forms an array (list) of the corresponding sequences, each of which is assigned the name PreviousPhrase and the number corresponding to this name. Next, block 19 from output 1 writes the generated array of word sequences, with the corresponding names and numbers to the input 1 of block 20. At the output 3 of block 19, a signal is generated that goes to input 2 of block 11, by which block 11 assigns the value of the vertex with the name current to the previous variable and from output 5 it writes to block 15. After that, block 11 at output 2, having prepared a signal for processing the next (new) section of the PC, feeds it to input 1 of block 8.

 If the current peak is a pause, then block 19 from output 3 sends a signal to input 2 of block 11, according to which block 11 at output 3 generates an enable signal and feeds it to input 1 of block 21. Based on this signal, block 21 reads data from the output of the block 20 and displays the results of variants of possible sequences of words that are acoustically similar to a spoken utterance.

 A more detailed algorithm for the operation of SLISR is presented in the flowchart of FIG. 12 - FIG. 19. Conventions in the presented algorithm are given on pages 29 - 31.

 The proposed SLISR, which uses SLD, in essence allows you to track several acoustically similar trajectories (sequences of vertices) from which you can choose the most optimal one. For this, it is necessary to modernize the lexical analyzer by introducing a path selection block into it.

 The system searches by sorting all the valid vertices (or only in the selected area) containing speakers that can follow the initial one. The search for the optimal AS sequence is carried out within a certain part of the SLD. Due to the fact that at each step of processing the input data several possible speaker options are selected, there is no need to go back.

 The advantages of the proposed system are that it allows lexical interpretation of continuous speech with higher speed and higher probability.

 Block 3 is a standard analog-to-digital converter for inputting acoustic signals into a computer and a set of filters, which can be implemented both hardware and software. Blocks 5, 10, 13, 15, 16, 17, 18, 20 - are memory blocks and can be made, such as, for example, in the form of storage devices, boards, nodes, etc. and depending on the volume of words used, they can be implemented on the basis of large, medium, and small integrated circuits, with their corresponding peripherals, or on the basis of drives on magneto-optical, electronic disks, etc. with their respective periphery. Blocks 4, 6 - 9, 11, 12, 14, 19 can be implemented both hardware and software. The software implementation of these blocks is presented in the form of a flowchart of the algorithm of operation in FIG. 12 - 19. Block 21 can be implemented as a device with a visual display of information (for example, a display), with its corresponding periphery, or in the form of an interface that provides logical or physical interaction between SLISR and a system: speech recognition or understanding, control of technological equipment or a robot , computer facilities, automatic voice translation, etc.

- символ окончания слова;
⊕ - символ конкатенации (склейки);
Unknown - метка не найденного АС;
η - весовой коэффициент;
n - номер ближайшей вершины;
V_т - текущее АС;
V_э - эталонное АС;
λ - коэффициент, определяющий смещение границ области поиска;
n_min - номер вершины, обозначающий нижнюю границу области поиска;

- номер вершины, обозначающий критическое значение нижней границы области поиска;
n_max - номер вершины, обозначающий верхнюю границу области поиска;

- номер вершины, обозначающий критическое значение верхней границы области поиска;
α - оценка степени совпадения текущего и эталонного акустического состояния;
ε - пороговое значение оценки степени совпадения текущего и эталонного акустического состояния;
α_opt - оптимальное значение оценки степени совпадения текущего и эталонного акустического состояния;
n_opt - номер вершины, соответствующий α_opt
Библиографические данные
1. Lesser V.R., Fennel R.D., Erman L.D., Reddy D.R., Organization of the HEARSAY II Speech Understanding System, IEEE Trans. ASSP, 23, 1, pp. 11-24, 1975.Legend
Previous - previous vertex;
Current - current peak;
Size - the size of the array (list) of the following possible vertices;
Edge - the size of the array (list) of boundary vertices;
SizeWord - the size of the array (list) of expected words;
SizePhrase - the size of the array (list) of expected sequences of words;
NewSizePhrase - size of the new array (list) of expected sequences of words;
Words [] - an array (list) of expected words;
PreviousPhrase [] - an array (list) of previous sequences of words;
NewPhrase [] - a new array (list) of expected sequences of words;
zPhrase - counter of zero sequences of words;
CurrentPhrase - current word sequences;
CurPhraseWord - the last word from the CurrentPhrase sequence;
Ptr - indicator of the current vertex of the array of the next possible vertices coming from the previous vertex Previous;
WordPtr - pointer to the current (expected) word;
PhrasePtr - pointer to the current (expected) sequence of words;
NewPhrasePrt - pointer to a new (expected) sequence of words;

- the word ending symbol;
⊕ - symbol of concatenation (gluing);
Unknown - label of the speaker not found;
η is the weight coefficient;
n is the number of the nearest vertex;
V _t - current speaker;
V _e - reference speaker;
λ is a coefficient determining the displacement of the boundaries of the search region;
n _min is the vertex number denoting the lower boundary of the search region;

- the vertex number denoting the critical value of the lower boundary of the search area;
n _max is the vertex number denoting the upper boundary of the search region;

- the vertex number denoting the critical value of the upper boundary of the search area;
α - assessment of the degree of coincidence of the current and reference acoustic state;
ε is the threshold value for assessing the degree of coincidence of the current and reference acoustic state;
α _opt is the optimal value for assessing the degree of coincidence of the current and reference acoustic state;
n _opt is the vertex number corresponding to α _opt
Bibliographic data
1. Lesser VR, Fennel RD, Erman LD, Reddy DR, Organization of the HEARSAY II Speech Understanding System, IEEE Trans. ASSP, 23, 1, pp. 11-24, 1975.

 2. Baker J. K., The DRAGON System - An overview, IEEE Trans. ASSP, 23, No. February 1, 1975, pp. 24 - 29.

 3. Klowstad J. W., Mondshein L. F., The CASPERS Linguistic Analysys System, IEEE Trans. ASSP, 23, No. February 1, 1975, pp. 118 - 123.

Claims (2)

 1. The method of lexical interpretation of continuous speech, which consists in periodically uttering a speech saying that is digitized at fixed time intervals with a given quantization frequency in this interval, then samples of this acoustic digitized signal are taken, from the totality of which the current values of the parameters of the input speech signal are calculated , compare the obtained values of the parameters of the input speech signal with the reference, pre-formed by the network of lexical decoding of a given a set of words, and based on the results of comparison, hypotheses are built about possible words in a speech utterance, characterized in that based on the lexical decoding network, hypotheses are made about the possible beginning, continuation, or end of layers in a speech utterance based on the calculated values of the input speech signal, and at the same time make up the most probable sequences of reference words corresponding to the spoken utterance, while the spoken words can continuously follow each other the friend in any order is either separated by pauses or words that do not belong to a given set of words, and the lexical decoding network is an integrated database containing spelling representations of a given set of words, expected acoustic representations of a given set of words in the form of sequences of reference values of speech signal parameters, defining acoustic states and combining phonetic transcriptions, phonological rules and vocabulary for a given set of words.  2. The system that implements the method according to claim 1, containing a series-connected acoustic analyzer containing a pre-processing unit, a frequency spectrum analyzer, a buffer for storing spectrum values, a lexical analyzer containing a comparison unit with a standard, a word database storage unit, characterized in that calculators of the weight coefficient and the current acoustic state are entered into the acoustic analyzer, and a determinant of expected acoustic states, a memory block of comparison estimates are entered into the lexical analyzer I, the control unit, the unit for selecting the optimal assessment and labeling of vertices, the storage unit for the database of boundary vertices, the verification unit, the memory unit for possible vertices, the storage unit for the database of local vertices, the storage unit for the database of standards of acoustic states, the lexical hypothesis generator, the lexical memory unit hypotheses, an output unit, while the first inputs of the weight factor and current acoustic state calculators are connected to the output of the spectrum storage buffer, the first output of the weight factor calculator connected to the control input of the current acoustic state calculator, the second output of the weight factor calculator is connected to the second input of the expected acoustic state determinant, the first output of which is connected to the second input of the comparison unit with the reference, the output of the current acoustic state calculator is connected to the first input of the comparison unit with the standard, the second the output of the determinant of expected acoustic states is connected to the third input of the storage unit of the database of local vertices, the third output of the determinant of the given acoustic states is connected to the input of the boundary vertex database storage unit, the fourth output of the expected acoustic states determinant is connected to the control input of the weight factor calculator, the output of the comparison unit with the reference is connected to the input of the comparison estimates memory block, the output of which is connected to the first input of the control unit, the first the output of the control unit is connected to the control input of the comparison estimates memory block, the second output of the control unit is connected to the first input of the expected condition, the third output of the control unit is connected to the first input of the output unit, the fourth output of the control unit is connected to the control input of the lexical hypothesis memory unit, the fifth output of the control unit is connected to the control input of the memory unit of possible vertices, the sixth output of the control unit is connected to the input of the optimal selection unit assessment and marking of vertices, the first output of which is connected to the first input of the check unit, the second output of the block for selecting the optimal assessment and marking of vertices is connected to the fifth input of the block and the control, the first output of the verification unit is connected to the input of the memory block of possible vertices, the second output of the verification unit is connected to the second input of the storage unit of the local vertex database, the third output of the verification unit is connected to the fourth input of the control unit, the output of the boundary vertex database storage unit is connected to the fourth input of the local vertex database storage unit, the first output of the possible vertex memory block is connected to the second input of the lexical hypothesis generator, the second output of the possible vertex memory block is connected nen with the third input of the control unit, the first output of the lexical hypothesis memory unit is connected to the first input of the lexical hypothesis generator, the first output of the lexical hypothesis generator is connected to the input of the lexical hypothesis memory, the second output of which is connected to the second input of the output unit, the second output of the lexical hypothesis generator is connected with the first input of the local vertex database storage unit, the third output of the lexical hypothesis generator is connected to the second input of the control unit, the first output of the xp unit An analysis of the local vertex database is connected to the second input of the verification unit, the second output of the local vertex database storage unit is connected to the third input of the lexical hypothesis generator, the third output of the local vertex database storage unit is connected to the third input of the expected acoustic state determinant, and the fourth output of the base storage unit local vertex data is connected to the input of the word database storage unit, the output of which is connected to the fifth input of the local vertex database storage unit, the fifth output is and database storage connected to the local peaks entry database storage state of the reference acoustic unit, an output of which is connected to a sixth input data base storage unit of local peaks.

Images