JP5161174B2 - Route search device, speech recognition device, method and program thereof - Google Patents

Description

  The present invention relates to a technique for finding a route that matches an input sequence among a plurality of routes from a start point node to an end point node of a directed graph when a directed graph and an input sequence are given. The present invention also relates to a speech recognition technique using a technique for finding this route.

  When a model of series data is expressed by a directed graph and given an input sequence, in order to determine whether the input sequence is compatible with the model, cumulative weights are selected from various paths in the directed graph. There is a method of searching for a route that becomes the minimum or maximum and measuring the degree of fitness with the searched cumulative weight. The cumulative weight is a value obtained by sequentially accumulating the weight assigned to the node or arc on the route and the distance (or similarity) between the label assigned to the arc on the route and the element of the input sequence. There is also a method for estimating what kind of input sequence is the label sequence on the searched route. Such a method models the characteristics of series data belonging to a certain category and automatically detects series data having the characteristics of that category from a large number of series data. It is used for the purpose of discriminating whether it is done. However, the route in the present invention represents a transition process from a certain node to a certain node, and may pass through the same arc many times.

  For example, in speech recognition, rules such as phoneme and word arrangement are represented by a directed graph, the phoneme and word chain probabilities (logarithmic values) are given to the arc as weights, and the phoneme and word acoustic patterns and their distribution functions are represented. A label (e.g. pointing to the distribution function of the short-time spectral pattern of phonemes / a /) is assigned to each corresponding arc. The similarity between the feature vector representing the short-time spectrum pattern of the input speech and the label is a value when the feature vector is input to the distribution function corresponding to the label. In addition, a start point node and an end point node are also set according to the rules of phoneme and word arrangement. From the prepared directed graph for speech recognition, a path with the minimum or maximum cumulative weight from the start node to the end node is obtained for the time series of the short-time spectrum pattern of the input speech. Thus, a label string (phoneme string) on the route can be output as a speech recognition result.

  However, there are many paths in a large directed graph, and it is not easy to search for a path with a large or small cumulative weight. For example, when enumerating all possible paths and trying to calculate the cumulative weight, the number of paths listed increases exponentially with length. For example, consider a simple directed graph as shown in FIG. The circle in FIG. 7 represents a node, the arc with an arrow represents an arc with a direction, and the alphabet arranged near each arc represents the label of the arc. At this time, if an input sequence having a length of 10 is given, the total number of paths is the 10th power of the number of arcs (= 9765625). Even in such a simple directed graph, the number of different paths becomes enormous.

  Therefore, a beam search is performed in which the elements of the input sequence are read in order, the path is extended by one arc from the start node in accordance with the reading, and the process is performed while eliminating the less likely path based on the cumulative weight each time. Often used. Here, the less likely route is a route with a larger cumulative weight compared to other routes when searching for a route with a smaller cumulative weight, and a route with a higher cumulative weight compared to other routes when searching for a route with a larger cumulative weight. It is assumed that the route has a small cumulative weight.

  A beam search method in a general graph search problem is described in Non-Patent Document 1, for example. A beam search method when an input sequence is given to a directed graph is described in Non-Patent Document 2, for example.

  Hereinafter, a conventional beam search procedure will be described with reference to FIG. Hereinafter, it is assumed that the route to be searched is a route with a small cumulative weight. When searching for a path with a large accumulated weight, the sign of the weight is reversed, the sign of the distance between the arc label and the input element is reversed or their similarity is used, the minimum value calculation is replaced with the maximum value calculation, etc. This can be achieved with simple changes. In the explanation, one route in the middle of a search is called a hypothesis. A hypothesis that reaches the end node from the start node is called a complete hypothesis.

  As shown in FIG. 8, it is first determined whether or not there is an input sequence to be read (step S1). If there is an input sequence to be read, a set of hypotheses including only a zero-length hypothesis that first reaches the start node is Are prepared as a set of hypotheses (step S2). Next, the input sequence is read (step S3), and it is determined whether there is an element that can be read from the input sequence (step S4). If there is an element to be read in the input series, hypothesis update (step S5) and hypothesis pruning (step S6) are performed.

  In step S5, each hypothesis included in the current hypothesis set is added with each arc that can be transitioned from the last node reached by each hypothesis and the transition destination node of each arc, and a new hypothesis is obtained. To do. Further, the cumulative weight of the hypothesis, the weight of the transitionable arc, and the value obtained by adding the distance between the current input element and the label of the transitionable arc are set as the new hypothesis cumulative weight. In this manner, the hypothesis is updated by replacing the current hypothesis set with the generated new hypothesis set.

  In step S6, the number of hypotheses is reduced by excluding hypotheses with a large accumulated weight from the updated new hypothesis set.

  If there is no element to be read in the input sequence, that is, when the input element has been read to the end, the complete hypothesis is selected and output (step S7). That is, one or more complete hypotheses are selected from the complete hypotheses in the current hypothesis set in ascending order of cumulative weight, and these are output as candidate routes from the start node to the end node (step S7). ). Then, it returns to step S1. In step S1, if there is no input sequence to be read next, the process ends.

  For the pruning of the hypothesis (step S6), generally, the cumulative weight of the hypothesis of interest is a value obtained by adding a positive constant (also called a beam width) to the smallest cumulative weight in the current set of hypotheses. Is a positive constant that has a rank in the cumulative weight in the current hypothesis set of the hypothesis of interest (in contrast to the beam width for the cumulative weight, the number beam width There is a way to eliminate the hypothesis if it is larger than

  Eliminating hypotheses that are less probable during the search, that is, hypotheses that are unlikely to become complete hypotheses with a small cumulative weight even if this hypothesis is extended in the future, is generally called hypothesis pruning.

  When searching for a complete hypothesis having a small cumulative weight by the beam search method, it is possible to reduce the amount of calculation by pruning a large number of hypotheses from a set of hypotheses in pruning the hypotheses. This is because updating the hypothesis that accounts for the majority of calculations generates new hypotheses for each hypothesis included in the current set of hypotheses as many arcs that can be transitioned from the node that the hypothesis has reached. This is because the smaller the number of hypotheses included, the smaller the calculation amount.

Xuedong Huang, Alex Acero, Hsiao-Wuen Hon, “Spoken Language Processing”, Prentice Hall, Upper Saddle River, NJ 07458, Chapter 12, Basic Search Algorithms, Section 12.1.3.2, pp.606-608 Takaaki Hori, Chiori Hori, Yasuhiro Minami, Atsushi Nakamura, `` Efficient WFST-based one-pass decoding with on-the-fly hypothesis rescoring in extremely large vocabulary continuous speech recognition '', IEEE Transactions on Audio, Speech, and Language Processing, Vol .15, No.14, pp.1352-1365, 2007, Section IV B (pp.1357-1359)

  Although it is possible to reduce the number of hypotheses by reducing the beam width or number beam width to make the hypothesis more easily pruned, it can become a complete hypothesis with a small cumulative weight when the hypothesis is extended in the future. Pruning the hypothesis causes a search error that makes it impossible to obtain the desired complete hypothesis. That is, when the beam width is narrowed, the amount of search processing is reduced, but the risk of search errors increases.

  An object of the present invention is to provide a route search device, a speech recognition device, and a method and a program thereof that are less likely to cause a search error than before.

  When pruning hypotheses, the features for each hypothesis are extracted. The output value when the extracted feature is input to the discriminant function that outputs an indicator as to whether the hypothesis is to be excluded or not by using the feature as an input is calculated and used as a discriminant value for each hypothesis. If the discriminant value is larger than the predetermined threshold value, and if it is large, remove the hypothesis corresponding to the discriminant value from the hypothesis set, or if the discriminant value is smaller than the predetermined threshold value and smaller Excludes the hypothesis corresponding to the discriminant value from the set of hypotheses.

  By using attributes other than the hypothetical cumulative weight, for example, which node has been reached, which arc has been passed, etc., more accurate pruning determination can be performed, and the risk of search error is smaller than before. can do.

The functional block diagram of the example of a route search apparatus. The functional block diagram of the example of a speech recognition apparatus. The flowchart of the example of a route search method. The flowchart of the example of the speech recognition method. Diagram showing an example of a directed graph for speech recognition The figure which shows the comparative evaluation result of the recognition processing time of a conventional method and this invention. Diagram illustrating a directed graph The flowchart of the example of the conventional route search method.

[Route search apparatus and method]
FIG. 1 shows a functional block diagram of an example of a route search apparatus. FIG. 3 shows a flowchart of an example of the route search method.
The route search apparatus and method of the present invention differs from the conventional one in that it mainly includes the processing of the hypothesis pruning unit 6 and the hypothesis pruning discrimination function update unit 8 as necessary. In the following description, parts different from the conventional ones will be mainly described, and duplicate explanations may be omitted for the same parts as the conventional ones.
The route search device 100 includes a storage unit 1, a control unit 2, an initial hypothesis generation unit 3, an input series reading unit 4, a hypothesis update unit 5, a hypothesis pruning unit 6, a complete hypothesis output unit 7, and a hypothesis pruning discrimination function update unit. 8 is included, for example.

The control unit 2 determines whether there is an input sequence to be read (step S1).
If there is an input sequence to be read, the initial hypothesis generation unit 3 stores the set of hypotheses including only the zero-length hypothesis that first reaches the start node as the current hypothesis set in the storage unit 1 (step S2). . The cumulative weight of the hypothesis of length 0 that reaches the start point node first is assumed to be 0, for example.
The input sequence reading unit 4 reads the input sequence (step S3) and determines whether there is an element that can be read from the input sequence (step S4).
When there is an element to be read in the input series, the hypothesis updating unit 5 updates the hypothesis (step S5), and the hypothesis pruning unit 6 performs the hypothesis pruning (step S6).

  The hypothesis updating unit 5 includes, for each hypothesis included in the current hypothesis set read from the storage unit 1, each arc that can be transitioned from the last node reached by each hypothesis, and the transition destination node of each arc. To make a new hypothesis. Further, the cumulative weight of the hypothesis, the weight of the transitionable arc, and the value obtained by adding the distance between the current input element and the label of the transitionable arc are set as the new hypothesis cumulative weight. In this manner, the hypothesis is updated by replacing the current hypothesis set with the generated new hypothesis set. A newly generated set of hypotheses is stored in the storage unit 1.

  The hypothesis pruning unit 6 reduces the number of hypotheses by excluding hypotheses having a large cumulative weight from the updated new hypothesis set.

  When there is no element to be read in the input sequence, that is, when the input element has been read to the end, the complete hypothesis output unit 7 selects and outputs a complete hypothesis (step S7). That is, one or more complete hypotheses are selected from the complete hypotheses in the current hypothesis set in ascending order of cumulative weight, and these are output as candidate routes from the start node to the end node. Then, it returns to step S1. In step S1, if there is no input sequence to be read next, the process ends.

  Hereinafter, the hypothetical pruning unit 6 will be described in detail. As illustrated in FIG. 2, the hypothesis pruning unit 6 includes a hypothesis feature extracting unit 61 that extracts a feature for each hypothesis, and a discriminant function that outputs an index as to whether or not the hypothesis is excluded with the feature as an input. A hypothesis pruning discriminant value calculation unit 62 that calculates an output value when the extracted feature is input and uses it as a discriminant value for each hypothesis, and determines whether the discriminant value is larger or smaller than a predetermined threshold value In the case where it is larger or smaller, a hypothesis pruning determination unit 63 that excludes the hypothesis corresponding to the discriminant value from the set of hypotheses is included.

  The hypothesis feature extraction unit 61 extracts features for individual hypotheses included in the current set of hypotheses (step S61). If the hypothesis is h, the current set of hypotheses is H, the cumulative weight of hypothesis h is w [h], and the rank in the set of cumulative weights of hypothesis h is rank (h), A feature vector such as can be extracted.

The first-dimensional feature w [h] − (min _g∈H w [g] + B) of the feature vector φ (h, H) is a positive constant B (beam) to the minimum cumulative weight of the hypotheses in the set H. Is a value obtained by adding a value obtained by adding (equivalent to the width) to the cumulative weight w [h] of the hypothesis h. If this feature is positive, the hypothesis h is pruned according to the conventional beam width B. Gives the discriminant value of not cutting.

  The next second-dimensional feature rank (h, H) -R is a value obtained by subtracting a positive constant R (corresponding to the number beam width) from the rank in the cumulative weight set H of the hypothesis h. If the feature is positive, a determination value is given that the hypothesis h is pruned according to the number beam width R, and if the feature is negative, the pruning value is not pruned.

These two features were also used in the conventional beam search, but in the present invention, more features such as the node n [h] and arc a [h] where the hypothesis h finally arrives are used. In the example of the feature vector φ (h, H) described above, f _{node [i]} (n [h]) returns 1 if the node n [h] is the i-th node of the directed graph, and returns 0 otherwise. Similarly, far _[j] (a [h]) is a function that returns 1 if a [h] is the j-th arc of the directed graph, and returns 0 otherwise. Here, it is assumed that a unique number is assigned to each arc or node of the directed graph, and the total number of nodes of the directed graph is N, and the total number of arcs is A. Therefore, 1 ≦ i ≦ N and 1 ≦ j ≦ A. As other features, the length of hypothesis h (the number of arcs on the path), the label of arc a [h], and the like can be introduced as similar functions.

  Next, the hypothesis pruning discriminant value calculation unit 62 determines the value of a discriminant function that determines whether or not the hypothesis is pruned based on the feature vector φ (h, H) extracted by the hypothesis feature extraction unit 61. Calculate (step S62). The discriminant function has a parameter vector Λ, and the discriminant criteria can be changed variously by changing Λ. For example, when a linear discriminant function is used, a coefficient vector (hereinafter referred to as a feature weight coefficient vector) for each dimension of the feature vector φ (h, H) is a parameter vector Λ, and the linear discriminant function is Λ and φ (h, H ) Of the inner product Λ · φ (h, H). For example, Λ is a vector as follows.

Here, _[ lambda _{] weight} , _[ lambda _{] rank} , _[ lambda _{] node [1]} , _[ lambda _{] node [2]} ,..., _[ Lambda _{] node [N]} , _[ lambda _{] arc [1]} , _[ lambda _{] arc [2]} , ..., _[ lambda _{] arc [A]} , The feature weighting coefficient for each dimension of the feature vector φ (h, H) in the equation (1). The parameter of this discriminant function can be obtained in advance using learning data.
In the examples of equations (1) and (2), these inner products can be calculated as follows.

Subsequently, the hypothesis pruning determination unit 63 determines that the hypothesis is pruned when the value of such a discriminant function is greater than a certain threshold (for example, 0), and is not pruned when the value is smaller. In the example of the expression (3), in addition to the determination based on the conventional pruning determination criteria such as the beam width and the number beam width, the hypothesis corresponds to the node n [h] or the arc a [h] that the hypothesis finally reached. The values of λ _{n [h]} and λ _{a [h]} affect the determination. As described above, the pruning determination of the hypothesis is performed based on different determination criteria depending on the node or arc that the hypothesis has reached, so that more accurate pruning determination can be performed.
The linear discriminant functions shown in the examples of the equations (2) and (3) are linear equations. However, higher-order polynomials may be used as long as the function gives some discriminant value, or some nonlinear function is used. May be.

Next, the hypothetical pruning discriminant function updating unit 8 that updates the discriminant function parameters will be described. Let d (φ (h, H); Λ) be a discriminant function of the feature vector φ (h, H) of the hypothesis h. In the case of a linear discriminant function, d (φ (h, H); Λ) = Λ · φ (h, H).
Also, a search error risk function risk (d (φ (h, H); Λ)) representing the possibility of a search error occurring is defined according to the value of the discriminant function. When d (φ (h, H); Λ) is small, this function is less likely to cause the hypothesis h to be pruned, so when d (φ (h, H); Λ) is large. Since the hypothesis h is pruned and the risk of a search error increases, it is assumed that the value is greater than zero.

  As this search error risk function, for example, a sigmoid function of Expression (4), a hinge function of Expression (5), or the like can be used.

Here, x = d (φ (h, H); Λ), c represents a constant. In any case, the value approaches 0 when x is small, and takes a value larger than 0 when x is large.
To find parameters with a lower search error risk, the sum of the search error risks for the hypothesis features on the path of the complete hypothesis with a smaller cumulative weight

Update the parameter so that Here, K is a set of complete hypotheses obtained by reading the input sequence O to the end, G (k | K, O) represents the importance of the complete hypothesis k in K, and the smaller the cumulative weight is, the smaller the cumulative weight is. A large value of 0 or more is assumed. For example, when the arc weight of the directed graph or the distance between the arc label and each element of the input series is calculated as a negative logarithmic probability, G (k | K, O) is

It is good also as posterior probability of k like. On the other hand, H _t is a set of current hypotheses updated by the t-th element of O, and the condition “h is a part of k” is the path from the start node to the node where the path of hypothesis h is k and the node where h has reached It means that they match. That is, only the hypotheses that have reached the complete hypothesis in the middle of the search are summed.

For example, Λ for reducing L (Λ, O) is, for example, a partial differentiation of L (Λ, O) by each element of Λ, and multiplying the value by a learning coefficient (a positive constant) to each element of Λ L (Λ, O) can be reduced by subtracting from. This is the principle of the stochastic descent method.
When partial differentiation of equation (6) with element λ _x of parameter vector Λ (x indicates an arbitrary feature),

It becomes. λ _x uses the partial differentiation of L (Λ, O) by λ _x

Update like this. Here, η is a positive constant representing a learning coefficient.
For example, when the linear discriminant function of Expression (3) is used and the sigmoid function of Expression (4) is used as the search error risk function, L (Λ, O) is converted to feature weights λ _weight , λ _rank , and node n When the partial differentiation is performed with respect to the feature weight λ _n and the feature weight λ _a of the arc a,

It can be calculated as follows.
Further, the sum of pruning error risks for the learning data O using learning data O = {O ₁ , O ₂ ,..., O _Y } consisting of Y input sequences prepared in advance.

To reduce the element λ _{x of the} parameter vector

It can also be updated as follows.

Further, the step of performing a search process using the parameter vector Λ and the learning data O to obtain the sum of the differential values of Expression (8) and the step of updating Λ of Expression (9) may be repeated. By repeating this, L (Λ, O) can be further reduced.
However, if L (Λ, O) is repeatedly reduced, the pruning determination becomes sweet as a whole, so the number of hypotheses that are not pruned increases and the amount of calculation gradually increases. Therefore, a penalty term that avoids monotonic decrease of L (Λ, O) due to repetition may be introduced.
For example, we introduce a penalty term into equation (6)

The following formula can be used. The penalty term (second term) on the right side is the square norm of Λ, which is the value obtained by summing all the elements after squaring each element of Λ. This penalty term is minimized when all the elements of Λ are 0. Therefore, when the decrease in the first term due to the update of Λ matches the increase in the second term due to the non-zero element of Λ Becomes 0, and L ^ (Λ, O) takes the minimum value. That is, it is possible to reduce the total search error risk while suppressing an increase in calculation amount. However, γ is a weighting factor of the penalty term, and is a positive constant that balances the first term and the second term.
Further, a penalty term reflecting the value of the discriminant function may be used as in the following equation.

Equation (13) is a penalty that works so that the sum of discriminant function values approaches zero. That is, it can be expected that the calculation amount is kept constant without being biased positively or negatively as a whole.
For the total pruning error risk with such a penalty term, as in equation (9),

The parameter can be updated as follows. The derivative of equation (11) is

The derivative of equation (13) is as follows when the discriminant function is a linear discriminant function:

It becomes. However, φ _x (h, H _t ) represents the value of the element x of the feature vector.
Even when learning data O = {O ₁ , O ₂ ,..., O _Y } composed of Y input sequences is used, as in the equation (11),

To update the parameter.

[Voice recognition apparatus and method]
The route search apparatus and method can be applied to a speech recognition apparatus and method.
FIG. 2 is a functional block diagram of an example of a speech recognition apparatus, and FIG. 4 is a flowchart of an example of a speech recognition method.
The speech recognition device includes, for example, an acoustic model storage unit 11, a directed graph storage unit 12, a speech signal input unit 13, a speech feature vector extraction unit 14, a route search device 100, a recognition result output unit 15, and learning data 16.

  The control unit 2 of the route search apparatus 100 reads an acoustic model for calculating the distance between the element of the input sequence and the arc label of the directed graph from the acoustic model storage unit 11 (step A1). Here, the acoustic model refers to, for example, a standard speech feature vector or a distribution of a speech fixed unit (for example, phoneme) and a certain input feature vector, and the feature vector is used as the speech fixed unit. It returns what is plausible by numerical values such as distance and probability.

  As an acoustic model representing a set of speech feature vectors of various speech fixed units (for example, phonemes), for example, a hidden Markov model method for modeling a set of speech feature vector time series of speech fixed units based on probability / statistical theory (Hidden Markov Model, hereinafter referred to as HMM) is the mainstream. Details of the HMM method are disclosed in, for example, “Recognition of Speech by Stochastic Model” by Seiichi Nakagawa, edited by the Institute of Electronics, Information and Communication Engineers.

  Next, the control unit 2 of the route search apparatus 100 reads a directed graph for speech recognition from the directed graph storage unit 12 (step A2).

  A method for constructing a directed graph for speech recognition is, for example, describing a word pronunciation dictionary or a language model with a weighted finite state transducer (WFST) and combining them into a single weighted finite state. The method of constructing the transducer is described in, for example, academic papers, M. Mohri, F. Pereira, M. Riley “Weighted finite-state transducers in speech recognition”, Computer Speech and Language, Vol. 16, No. 1, pp. 69--88 (2002). A weighted finite state transducer is a type of directed graph.

  In general, a label indicating one state of the HMM is given to each arc of the directed graph for speech recognition. The state of the HMM is a speech fixed unit corresponding to a rough position (for example, the first half, the middle, and the second half) of a certain phoneme in each phoneme, and each state is a probability density distribution (for example, a multidimensional mixed normal) of a speech feature vector. Distribution). In this embodiment, the sign of the logarithm of the probability density of the speech feature vector calculated using the probability density distribution of the state corresponding to the distance between the speech feature vector added to the cumulative weight of the hypothesis and the label of the arc. Calculated as the inverted value.

  A directed graph for speech recognition is configured as shown in FIG. 5, for example. The start node is indicated by a black circle, the end node is indicated by a double circle, and the other nodes are indicated by circles. A label is given to the arc connecting the nodes, and the phoneme name and the position in the phoneme are indicated. For example, a2 indicates the probability density distribution of the second state of the HMM of phoneme a.

  When two labels are written with “:” as in “a1: red”, the distance between the label and the input vector is calculated based on a1 before “:”, and after “:”. “Red” is a label that becomes a recognition result when a complete hypothesis passing through this arc is selected as a search result. In the example of Fig. 5, there are a route that passes up and a route that passes down. When the complete hypothesis passing above is selected, "red" is the recognition result, and when the complete hypothesis passing below is selected. , “It is blue” is the recognition result.

  The voice signal input unit 13 inputs voice, and the voice feature vector extraction unit 14 extracts a time series of feature vectors of the short-time spectrum pattern of the voice signal sent from the voice signal input unit 13 (step A3). Feature vectors used for speech recognition include a mel cepstrum (referred to as mel-frequency cepstral coefficients (MFCC)), delta MFCC, LPC cepstrum, logarithmic power obtained by analyzing a speech signal every short time (for example, 10 milliseconds). and so on.

  The route search apparatus 100 uses a time series of feature vectors obtained from the speech feature vector extraction unit 14 as an input series. The route search device 100 reads a time series of feature vectors sent from the speech feature vector extraction unit 14 using the acoustic model read from the acoustic model storage unit 11 and the directed graph for speech recognition read from the directed graph storage unit 12. The complete hypothesis with the minimum cumulative weight is obtained and sent to the recognition result output unit 15.

Since the functional configuration and processing (step S1 to step S8) of the route search device 100 are the same as those described in the section [Route search device and method] above, duplicate description is omitted here. Note that the label is a label corresponding to an acoustic pattern such as a phoneme or a state of a hidden Markov model, and the distance between the label and the element of the input sequence is the Euclidean distance or the hidden from the average feature vector of the phoneme corresponding to the label. The log likelihood is calculated from the probability density function assigned to the state of the Markov model.
The recognition result output unit 15 outputs a label corresponding to the complete hypothesis as a speech recognition result (step A3).

[Experimental result]
A speech recognition apparatus was constructed in the form shown in FIG. The acoustic model has HMMs for 43 phonemes, and each phoneme has three states. Each state has a phoneme context (what is the previous phoneme and what is the phoneme that follows)? Accordingly, one of the 3064 probability density distributions is assigned.

  The directed graph used a weighted finite state transducer constructed by the method shown in Document 1 using a pronunciation dictionary composed of 20,000 words and a trigram language model that gives a three-chain probability of words. The number of nodes in the directed graph is 72469, and the number of arcs is 102073.

  The feature vector time series of speech is the MFCC 12 dimension obtained by analyzing the speech signal every 10 milliseconds, the delta MFCC 12 dimension which is the primary regression coefficient when looking at the two frames before and after the time series direction of each dimension of the MFCC, each dimension The delta delta MFCC 12-dimensional, which is a primary regression coefficient when viewing two frames before and after in the time series direction, and a 39-dimensional vector combined with logarithmic power are extracted as input sequences.

  Further, the hypothetical pruning discriminant function updating unit 8 (FIG. 1) updated the parameters of the hypothetical pruning discriminant function 100 times using the 33820 sentences read by the male speaker in advance as learning data. At this time, as a feature vector of hypothesis h,

Was used.
As described above, the first-dimensional feature w [h] − (min _g∈H w [g] + B) is a positive constant B (corresponding to the beam width) to the minimum value of the hypothetical cumulative weight in the set H. Is a value obtained by subtracting h from the cumulative weight w [h] of h, and if this feature is positive, h is pruned according to the conventional beam width B, and if it is negative, it is not determined to be pruned. give. _{Far [j]} (a [h]) in the second and subsequent dimensions is a function that returns 1 if a [h] is the j-th arc of the directed graph, and returns 0 otherwise.
As feature weights as parameters of hypothetical pruning discriminant function

Was used. The hypothetical pruning discriminant function is a linear discriminant function Λ · φ (h, H), and the sigmoid function formula (4) is used as the risk function. The update of the parameters was performed using the derivative of equation (15) and the update equation of equation (16) based on the sum of the search error risks of equation (13).

  FIG. 6 shows the processing time required for the speech recognition processing when the subject inputs speech that reads out 100 sentences in a newspaper article using this speech recognition apparatus. However, the processing time is a value (actual time ratio) obtained by dividing the processing time required for speech recognition by the actual speech time. The word error rate is a ratio of errors per word when the correct sentence is compared with the recognition result, and is usually calculated as follows.

  The number of replacement errors is the number of times the correct sentence word is recognized as another word in the recognition result, the number of insertion errors is the number of times a word that does not exist in the correct sentence is inserted into the recognition result, and the number of deletion errors is the correct sentence. This represents the number of times that the included word was not included in the recognition result. A smaller word error rate means higher recognition accuracy.

In the conventional beam search method, the pruning determination is performed only by the aforementioned w [h] − (min _gεH w [g] + B).

  From the results shown in FIG. 6, the real-time ratio (processing time) is equal between the conventional beam search method and this experimental example, and the word error rate is smaller in this experimental example. From this, it was shown that the search error can be reduced in the speech recognition according to this experimental example. The real-time ratio is the same because the sum of the search error risks in equation (13) is used. Due to the effect of the penalty term, the discriminant function value as a whole is suppressed from being biased to positive or negative, and the amount of calculation is reduced. This is because it is kept constant.

[Modifications, etc.]
In FIG. 1 and FIG. 2, data is directly exchanged between the respective units, but data may be transferred via a storage unit (not shown). That is, data generated or received by each unit may be stored in the storage unit, and each unit may read the data from the storage unit.

  Each of the route search device and the speech recognition device can be realized by a computer. In this case, the processing contents of the functions that each device should have are described by a program. Then, by executing this program on a computer, each processing function in these devices is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. In this embodiment, these apparatuses are configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

  The present invention is not limited to the above-described embodiment, and can be modified as appropriate without departing from the spirit of the present invention. For example, the various processes described above are not only executed in time series according to the description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes.

DESCRIPTION OF SYMBOLS 1 Memory | storage part 2 Control part 3 Initial hypothesis generation part 4 Input series reading part 5 Hypothesis update part 6 Hypothesis pruning part 61 Hypothesis feature extraction part 62 Hypothesis pruning discrimination value calculation part 63 Hypothesis pruning judgment part 7 Complete hypothesis output part 8 Hypothesis pruning discrimination function update unit 11 Acoustic model storage unit 12 Directed graph storage unit 13 Speech signal input unit 14 Speech feature vector extraction unit 15 Recognition result output unit 16 Learning data 100 Route search device

Claims (7)

In a plurality of paths from the start point node to the end point node of a directed graph composed of a plurality of nodes including a start point node and an end point node, and arcs that connect the nodes and have weights and labels, a finite length In a route search device that finds a route that matches an input sequence,
An initial hypothesis generation unit having a hypothesis that is a path from the start point node to a certain node, and a hypothesis that reaches the start point node first;
An input sequence reading section for sequentially reading the elements of the input sequence;
Each hypothesis included in the set of hypotheses includes each arc that can transition from the last node reached by each hypothesis and the node to which each arc transitions, and a plurality of new hypotheses from each hypothesis. A value obtained by adding the cumulative weight of each hypothesis, the weight of each arc, and the distance between the label of each arc and the element of the read input sequence as the cumulative weight of the new hypothesis A hypothesis updating unit that replaces the new hypothesis with the new hypothesis set;
A hypothesis pruning unit that removes hypotheses with a large or small cumulative weight from the set of updated hypotheses;
Assuming that the hypothesis that has reached the end node is a complete hypothesis, after the elements to be read in the input series are eliminated, one or more complete hypotheses with a small or large cumulative weight among the complete hypotheses included in the set of hypotheses are output. A complete hypothesis output unit, and
The hypothesis pruning unit inputs the extracted feature to a hypothesis feature extracting unit that extracts a feature for each hypothesis and a discriminant function that outputs an indicator as to whether the hypothesis is excluded or not with the feature as an input. A hypothesis pruning discriminant value calculation unit that calculates an output value of each hypothesis and determines whether or not the discriminant value is greater than a predetermined threshold, and corresponds to the discriminant value A hypothesis pruning determination unit that removes a hypothesis from a set of hypotheses, or excludes a hypothesis corresponding to the discriminant value from the set of hypotheses if it is determined by determining whether the discriminant value is smaller than a predetermined threshold; seen including,
The features of the above hypothesis are
(a) The last node number that the hypothesis has reached,
(b) The last arc number that the hypothesis has reached,
(c) the length of the hypothesis and
(d) the label of the last arc that the hypothesis has reached,
A feature vector containing at least one of
The discriminant function is a function for obtaining an inner product of a feature vector and a parameter vector having the same dimension as the feature vector.
A route search apparatus characterized by that. The route search device according to claim 1,
The discriminant value for the hypothesis on the path of the complete hypothesis becomes smaller as the complete hypothesis with the smaller cumulative weight, or the discriminant value for the hypothesis on the path of the complete hypothesis becomes larger as the complete hypothesis with the larger cumulative weight, A hypothesis pruning discriminant function updating unit for updating the discriminant function;
A route search apparatus characterized by that. It includes a route searching apparatus according to claim 1 or 2,
The elements of the input sequence are feature vectors given by short-time spectral analysis of a speech signal,
The above label is a label corresponding to an acoustic pattern such as a phoneme or a state of a hidden Markov model,
The distance between the label and the element of the input series is a log likelihood calculated from the Euclidean distance from the average feature vector of the phoneme corresponding to the label or the probability density function assigned to the state of the hidden Markov model. ,
Search the path that best matches the input speech from the directed graph expressing the phoneme connections allowed as a language, and use the resulting label sequence on the path as the result of speech recognition.
A speech recognition apparatus characterized by that. In a plurality of paths from the start point node to the end point node of a directed graph composed of a plurality of nodes including a start point node and an end point node, and arcs that connect the nodes and have weights and labels, a finite length In a route search method for finding a route that matches an input sequence,
An initial hypothesis generation unit that assumes one path from the start node to a certain node as a hypothesis, and an initial hypothesis generation step in which a hypothesis that first reaches the start node is a set of hypotheses;
An input sequence reading step, in which the input sequence reading unit sequentially reads the elements of the input sequence;
The hypothesis updating unit adds each arc that can be transitioned from the last node reached by each hypothesis to each hypothesis included in the set of hypotheses and the transition destination node of each arc, and then adds a plurality of hypotheses from each hypothesis. A new hypothesis is generated, and the value obtained by adding the cumulative weight of each hypothesis, the weight of each arc, and the distance between the label of each arc and the element of the read input sequence is added to the new hypothesis. As a cumulative weight, a hypothesis updating step for replacing the new hypothesis set with the new hypothesis set,
A hypothesis pruning unit that removes hypotheses with a large or small cumulative weight from the updated set of hypotheses from the set;
The complete hypothesis output unit uses the hypothesis that has reached the end node as a complete hypothesis, and after the elements to be read in the input series are eliminated, one or more cumulative weights among the complete hypotheses included in the set of hypotheses are small or large A complete hypothesis output step for outputting a complete hypothesis of
In the hypothesis pruning step, when the extracted feature is input to a hypothesis feature extraction step that extracts a feature for each hypothesis and a discriminant function that outputs an index as to whether or not the hypothesis is excluded with the feature input A hypothesis pruning discriminant value calculating step for calculating the output value of each hypothesis and determining whether or not the discriminant value is greater than a predetermined threshold value and corresponding to the discriminant value A hypothesis pruning determination step unit that excludes a hypothesis from a set of hypotheses, or excludes a hypothesis corresponding to the determination value from the set of hypotheses if it is determined by determining whether the determination value is smaller than a predetermined threshold, and only including,
The features of the above hypothesis are
(a) The last node number that the hypothesis has reached,
(b) The last arc number that the hypothesis has reached,
(c) the length of the hypothesis and
(d) the label of the last arc that the hypothesis has reached,
A feature vector containing at least one of
The discriminant function is a function for obtaining an inner product of a feature vector and a parameter vector having the same dimension as the feature vector.
A route search method characterized by that. The route search method according to claim 4 ,
Update the discriminant function so that the discriminant value for the hypothesis on the path of the complete hypothesis decreases as the complete hypothesis with the smaller cumulative weight, or the discriminant value for the hypothesis on the path of the complete hypothesis as the complete hypothesis with the greater cumulative weight Further includes a hypothesis pruning discriminant function update step for updating the discriminant function so that
A route search method characterized by that. Including the route search method according to claim 4 or 5 ,
The elements of the above input sequence are feature vectors given by short-time spectral analysis of speech signals.
The above label is a label corresponding to an acoustic pattern such as a phoneme or a state of a hidden Markov model,
The distance between the label and the element of the input series is a log likelihood calculated from the Euclidean distance from the average feature vector of the phoneme corresponding to the label or the probability density function assigned to the state of the hidden Markov model. ,
Search the path that best matches the input speech from the directed graph expressing the phoneme connections allowed as a language, and use the resulting label sequence on the path as the result of speech recognition.
A speech recognition method characterized by the above. A program for causing a computer to function as the route search device according to claim 1 or 2 or the voice recognition device according to claim 3 .

Images