JPH01177600A - Voice recognition error correcting device - Google Patents

Abstract

PURPOSE:To practically improve the recognition performance of an acoustic recognizing part by using a reverse propagation network model to correct a time series of symbols including erroneous recognition obtained as the acoustic recognition result. CONSTITUTION:A symbol string as the acoustic recognition result is stored in an input buffer part 1, and an input window part 2 successively segments a fixed-length symbol string from the input buffer part 1 while shifting the start point by every one symbol, and a corresponding symbol in the input buffer part 1 is written with the output symbol each time when the inference result of the input is outputted. The output of a reverse propagation network model part 3 is stored in an output buffer part 4, and a first control part 6 shifts the start position of the input window part 2 by one symbol to perform the next correcting operation each time when one symbol is outputted. When detecting that correction is performed up to the layer symbol in the input buffer part 1, a second control part 7 writes back stored contents of the output buffer part 4 to the input buffer part 1 and performs the correcting operation again, and contents of the output buffer part 4 are started to write in a correction result output part 8 after this process is repeated a fixed number of times. Thus, the recognition performance of the acoustic recognizing part is improved.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech recognition error correction device. phoneme symbol strings written as a result of recognition, word symbol strings written as a result of word recognition, etc.)
This invention relates to an improvement of a speech recognition error correction device that corrects the error by taking into consideration the preceding and following contexts in a time series.

[Conventional technology]

As a method for correcting errors by taking into account the preceding and following contexts in the time series, the probability of appearance (conditional probability) of the central symbol when the symbol sequence before and after one is determined is calculated from the data to be recognized and tabulated. When a time series containing errors is given, there is a method of modifying the symbol sequence by using tabulated conditional probabilities to maximize the posterior probability. For example, when correction is performed by considering three symbols before and after one, the conditional probability P is expressed as follows.

(Formula 1) P(solslszsss4ssSssy
) Here, si represents the i-th symbol, and P represents the probability of mistaking the symbol So to 34. central symbol S
The correction result for 4 is (SsSzSsSaSs).

It can be decided by closing. That is, the correction result S. is △ (Equation 2) S, = argmax [P (so
l 51S2S3O 3a 5586'? )) is given by.

[Problem that the invention seeks to solve]

However, in the above method, the size of the table of conditional probabilities increases exponentially when the contexts taken into account are expanded, which is not practical.In other words, the length of the contexts taken into account is

Definition formula of conditional probability where symbol type is M (formula l
), the size of the table is ~O(M) (where ~0() indicates the order of size). Moreover, the amount of calculation required for optimization calculations to maximize the posterior probability cannot be ignored. Furthermore, when many errors are included in the preceding and following contexts, stable error correction becomes difficult.

The purpose of the present invention is to avoid the above-mentioned conditional probability table having an enormous storage capacity, which is difficult to realize, and to perform error correction using the recognition results of the acoustic recognition unit. By performing supervised learning, it is possible to perform error correction that adapts to the tendency of recognition errors in the acoustic recognition unit.Also, since no optimization calculation is required during error correction, it is possible to significantly reduce the amount of calculation. It is an object of the present invention to provide a recognition error correction device that makes it possible to perform stable error correction using a context with fewer errors by sequentially correcting an input symbol string using correction results.

If the recognition error correction device according to the present invention is used as a post-processing section of the acoustic recognition section, substantially the same effect as that of improving the recognition performance of the acoustic recognition section can be obtained.

[Means for solving problems]

The speech recognition error correction device according to the present invention is used in speech recognition to correct recognition errors included in a time series of symbols obtained as a result of recognition.

an input buffer section for storing the time series; an input window section for cutting out the fixed-length symbol string by sequentially shifting the starting point by one symbol from the beginning of the time series of symbols stored in the input buffer section; V5- A backpropagation network that performs supervised learning using a symbol string containing errors in advance so as to input the fixed-length symbol string obtained as the output of the input window and output the correct answer for the central symbol. - a model section, a rewriting section that rewrites the corresponding symbol in the input buffer section to a modified symbol at the time when the backpropagation network model section outputs a symbol; a first control unit that causes the backpropagation network e model unit to perform a correction operation on the next symbol by shifting the starting point of an input window unit for cutting out a symbol string by one symbol; and a symbol output by the backpropagation network model unit. an output buffer section for storing the sequence, and when it is detected that the symbol at the end of the symbol string in the input buffer section has been modified, the contents of the output buffer section are written back to the input buffer section, and the modification operation is performed again. The present invention comprises: a second control section that repeats the correction operation; and a correction result output section that outputs the contents of the output buffer section as a correction result when the correction operation is repeated a certain number of times.

[Effect]

The basic principle of the present invention is that in speech recognition, a backpropagation network is used that performs supervised learning in advance to analyze the time series of symbols, including erroneous recognition, obtained as acoustic recognition results.
This is an attempt to correct the problem using a model. The principle of the present invention will be explained in detail below.

The symbol string obtained as the output of the acoustic recognition section when input speech is recognized contains several errors reflecting the error tendency of the acoustic recognition section due to recognition errors of the acoustic recognition section, which are unavoidable under the present circumstances. The present invention corrects the time series of symbols containing errors by taking into consideration the context before and after the symbol, thereby essentially improving the recognition performance of the acoustic recognition unit.

For correction, a backpropagation network model, which was devised as a model for associative memory and pattern recognition, is used. For details of this model, please refer to [European Magazine Complex Systems%198IIEg1, pp.145-168J ('P
arallel Networks thatLear
n to Pronounce English T
ext', T, J.

Sejnowski h C, R, Rosenberg
, Complex Systems.

Vol. 1 (1987) 145-168).

Generally, a model is hierarchically composed of three types of layers as shown in Figure 2: an input crab layer, a hidden unit layer, and a hidden unit layer.
It is called the output layer.

Each layer has processing units called units.

Each unit receives an input from a unit in a layer adjacent to the input layer, and outputs an output to a unit in a layer adjacent to the output layer. The input/output response relationship of each unit is given as follows.

(Formula 4) y"=f(x dimension)) (Formula 5) f(x)=(1+e-")-'Here, X is the input to the unit, y is the output of the unit, and θ is the threshold value of the unit. , the superscript represents the hierarchy from the input layer (n==l,...,N).

The subscript is the number representing the unit in the layer, and the load representing the connection to unit j of the n-th layer, f(x) is (Equation 5
) is a nonlinear saturation type response function common to each unit. After all, each unit determines its output by a certain amount of threshold logic using as input the difference between the weighted sum of outputs of adjacent upper layer units and a predetermined threshold.

When data is given to the input layer of this model.

The information (data) II'i is sequentially processed in adjacent lower layers and propagated to the output layer. The output of this output layer unit becomes the model's inference result for the given input data.

In the present invention, when a fixed-length symbol string cut out from a symbol string containing ib is presented to the input layer, the error correction result (inference result) for the center symbol of the fixed-length symbol string input to the output layer is output. Construct a model that will

Next, a learning method (backpropagation learning) that determines the connections between units and units so that the model performs the desired inference operation.A fixed length of symbols extracted from the error-containing symbol string that is the output of the actual acoustic recognition unit for the incorrect input voice. This is pseudo data in which errors are stochastically added to a symbol string with no errors, assuming a symbol string of , or an M direction of fla between symbols. These data are presented to the input layer, the correct answer for the central symbol is presented to the output layer, and backpropagation learning is repeatedly performed. In the backpropagation method, a desired inference result (output data) for input data is given as a teacher signal, and connections between units are repeatedly corrected in a direction that reduces the difference (error) between the model's inference result and the teacher signal. In reality, the connection between units that minimizes the error function determined from the output yi of the model in the output layer (Nth layer) defined by the following equation and the desired output (answer) yI for the given input is created. Respond to what you discover.

(Formula 6) n= (1/2) (y mer y, )2
Since this function depends on all inter-unit connections through y(N), minimization can be performed using E as the evaluation function. The resulting backpropagation learning algorithm is detailed in the above-mentioned literature.

When performing correction using a trained model, 2
The starting point is shifted one symbol at a time from the symbol string that is the output of the acoustic recognition unit for input speech, and fixed-length symbol strings are sequentially cut out and input to the backpropagation network model. When the model outputs the correction result for the center symbol of the input (support) leg length symbol string, the corresponding symbol in the input symbol time series is rewritten with that symbol.

This means that the first half of the fixed-length symbol string presented to the input layer of the model by VC always consists of highly probable symbols that have been previously corrected, so the model cannot correct errors. It will be done more stably.

In this way, even the symbol string corrected by the model may still have errors that could not be corrected, so in order to correct the remaining errors, the entire symbol string that has been corrected by the model is given to the model again as input to perform error correction. By repeating this process, symbol strings with progressively fewer errors can be obtained.

〔Example〕

FIG. 1 is a schematic diagram showing an embodiment of a device implementing the present invention. The buffer section 1 stores the symbol string that is the acoustic recognition result, and the input window section 2 shifts the starting point one symbol at a time from the input buffer section 1 and sequentially cuts out fixed-length symbol strings to create a backpropagation network model. Every time unit 3 outputs an inference result for the input, it rewrites the corresponding symbol in the input buffer unit with the output symbol. The output buffer section 4 stores the output of the backpropagation network model section 3, and the first control section 6 causes the backpropagation network to be corrected. When the second control unit 7 detects that the symbols at the end of the input buffer unit 1 have been corrected, the second control unit 7 writes the memory contents of the output balance unit 4 back to the input buffer unit 1, causes the correction operation to be performed again, and keeps this process constant. After repeating the process several times, the contents of the output buffer section 4 are written to the modified result output section 8.

〔Effect of the invention〕

As described above, according to the present invention, it is possible to correct errors in the symbol string output from the acoustic recognition section in a bottom-up manner by utilizing the context. Furthermore, by writing the correction results back to the input symbol string symbol by symbol, it is possible to correct errors using a likely context, and by repeatedly re-inputting the entire output symbol of the model, errors can be corrected. By performing the correction, it is possible to obtain a correction result with fewer errors.

The effect of the present invention corresponds to improving the recognition performance of the /l'i sound recognition section, and makes it possible to achieve high accuracy as a whole of the speech recognition device.

Furthermore, the storage capacity required for execution is determined by the length of the context to be taken into account, where M is the type of symbol and H is the number of hidden units.The order of the storage capacity is as follows.

~O(L-M-)1)
13- Make possible.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing one embodiment of the present invention. FIG. 2 is a diagram showing the general configuration of a backpropagation network model. 1 is an input buffer section, 2 is an input window section, 3 is a backpropagation network model section, and 4 is an output buffer section. 5 is a rewriting unit, 6 is a first control unit, 7 is a second control unit, 8
Fi is a modified result output unit. Agent: Patent Attorney Susumu Uchihara 7＼, Kayue SoF Nasen

Claims (1)

[Claims] In speech recognition, when correcting recognition errors included in a time series of symbols obtained as a result of recognition, an input buffer section that stores the time series, and a time series of symbols stored in the input buffer section are used. An input window section that sequentially shifts the starting point one symbol at a time from the beginning of the series to cut out a fixed-length symbol string, and inputs the fixed-length symbol string obtained as the output of the input window section and calculates the correct answer for the center symbol. A backpropagation network model section that performs supervised learning using a symbol string containing errors in advance so as to output the corresponding symbol in the input buffer section at the time the backpropagation network model section outputs a symbol. a rewriting unit that rewrites the symbol sequence into a modified symbol, and then shifts the starting point of the input window unit that cuts out the fixed-length symbol string from the input buffer unit by one symbol and sends the next symbol to the backpropagation network model unit. a first control unit that performs a correction operation; an output buffer unit that stores the symbol string output by the backpropagation network model unit; and a first control unit that detects that a symbol at the end of the symbol string of the input buffer unit has been corrected. writing the contents of the output buffer section back to the input buffer section at a point in time;
A voice recognition system characterized by comprising: a second control section that repeats the correction operation again; and a correction result output section that outputs the contents of the output buffer section as a correction result when the correction operation is repeated a certain number of times. Error correction device.