JP6078402B2 - Speech recognition performance estimation apparatus, method and program thereof - Google Patents

Description

  The present invention relates to a speech recognition performance estimation device, method and program for estimating recognition accuracy of speech recognition.

  FIG. 1 illustrates the relationship between the recognition rate of speech recognition and the amount of learning data. The vertical axis represents the recognition rate [%], and the horizontal axis represents the learning data amount [time]. FIG. 1 shows the progress of learning of an acoustic model used for speech recognition. In acoustic model learning, if the amount of data in the learning set is increased, initially the recognition accuracy improves rapidly (acoustic model learning proceeds), but the recognition accuracy gain obtained by increasing the data gradually increases. The recognition accuracy is saturated when it reaches a certain recognition accuracy. Here, the recognition accuracy at the time of saturation is referred to as “recognition accuracy upper limit”.

  This learning data consists of a set of learning speech data and a correct text for learning that a person listens to the learning speech data and transcribes it into characters. The learning data is generally created using the voice of the introduction destination where the voice recognition system is introduced. Creating a correct text for learning requires a large cost. Therefore, there is a strong demand to minimize the amount of learning data.

  In order to determine whether the amount of learning data is a necessary and sufficient amount, it is necessary to know the “recognition accuracy upper limit” in advance. Conventionally, as a method for estimating the “recognition accuracy upper limit”, for example, a method is known in which the recognition accuracy measured by the closed evaluation described in Non-Patent Document 1 is used as the estimated value of the “recognition accuracy upper limit”.

  Closed evaluation refers to creating an acoustic model (closed acoustic model) learned using an evaluation set consisting of a set of evaluation speech and correct text for evaluation, and measuring the recognition accuracy of the evaluation set using the closed acoustic model. This is a performance evaluation method. Generally, the known data used for model learning can be recognized with high accuracy by the learned acoustic model. Since the recognition accuracy of known data measured by closed evaluation is a recognition accuracy under a kind of ideal condition, it is used as an estimated value of the upper limit of recognition accuracy.

Sako, Satoshi Yamagata, Tetsuya Takiguchi, Yasuo Ariki, `` System Requirements Detection by Integration with Speech Recognition '' Information Processing Society of Japan, SLP, Spoken Language Information Processing 2007 (129), pp.143-148, 2007.

  In general, the “recognition accuracy upper limit” obtained by the closed evaluation is higher than the actual “recognition accuracy upper limit”. The “recognition accuracy upper limit” measured in the closed evaluation is a result of recognizing known data. However, actual acoustic model learning is performed using data different from the evaluation set, and unknown data different from that used for learning is input during recognition accuracy evaluation. Therefore, recognition accuracy during evaluation is measured by closed evaluation. The value is always lower than the “recognition accuracy upper limit”. Judging whether the amount of learning data is excessive or insufficient based on the higher “recognition accuracy upper limit” than actual, the learning accuracy has progressed to some extent, and the recognition accuracy upper limit has been reduced, despite the fact that the recognition allowance is less. There is a problem in that it is determined that the learning set needs to be increased further from the gap, and unnecessary transcription work is continued, resulting in unnecessary costs.

  The present invention has been made in view of this problem, and an object of the present invention is to provide a speech recognition performance estimation apparatus capable of accurately estimating the “recognition accuracy upper limit”, and a method and program thereof.

The speech recognition performance estimation device of the present invention includes a closed acoustic model multiple generation unit, a closed acoustic model group recording unit, a learning set speech recognition accuracy calculation unit, an acoustic model selection unit, and a speech recognition accuracy calculation unit. To do. The closed acoustic model multiple generation unit receives the initial acoustic model and the evaluation set, extracts the acoustic feature amount from the evaluation sound data constituting the evaluation set, and models between the acoustic feature amount and the initial acoustic model N (N ≧ 2) closed acoustic models are generated and output to the closed acoustic model group recording unit. The learning set speech recognition accuracy calculation unit outputs N speech recognition accuracy obtained by performing speech recognition on the learning set using the N closed acoustic models recorded in the closed acoustic model group recording unit with the learning set as an input. To do. The acoustic model selection unit receives the N speech recognition accuracies as input, refers to the closed acoustic model group recording unit, selects a closed acoustic model corresponding to the maximum value of the N speech recognition accuracies, and is selected closed Output as an acoustic model. The speech recognition accuracy calculation unit outputs the speech recognition accuracy obtained by recognizing the evaluation set using the selected closed acoustic model as an estimated upper limit of the recognition accuracy.

  According to the speech recognition performance estimation device of the present invention, N acoustic models that are models between the speech data acoustic feature for evaluation and the initial acoustic model are generated, and a learning set is generated from the N acoustic models. The acoustic model showing the maximum recognition accuracy is selected, and the recognition accuracy obtained by voice recognition of the evaluation set using the acoustic model is output as an estimated value of “recognition accuracy upper limit”. Therefore, the approximate “recognition accuracy upper limit” can be known in advance. As a result, since the allowance for the recognition rate can be grasped in advance, it is possible to prevent the occurrence of extra transcription work and to suppress the development cost when the speech recognition system is introduced.

The figure which shows the example of the relationship between the recognition rate of speech recognition, and learning data amount. The figure which shows the function structural example of the speech recognition performance estimation apparatus 100 of this invention. The figure which shows the operation | movement flow of the speech recognition performance estimation apparatus 100. The figure which shows the function structural example of the closed acoustic model multiple production | generation part 110. FIG. The figure which shows the function structural example of the speech recognition performance estimation apparatus 200,200 'of this invention. The figure which shows the function structural example of the speech recognition precision calculation part 230. FIG. The figure which shows the function structural example of the speech recognition performance estimation apparatus 300 of this invention. The figure which shows the function structural example of the speech recognition performance estimation apparatus 400 of this invention. The figure which shows the function structural example of the speech recognition performance estimation apparatus 500 of this invention. The figure which shows the operation | movement flow of the speech recognition performance estimation apparatus 500.

  Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.

  FIG. 2 shows a functional configuration example of the speech recognition performance estimation apparatus 100 of the present invention. The operation flow is shown in FIG. The speech recognition performance estimation apparatus 100 includes a closed acoustic model multiple generation unit 110, a closed acoustic model group recording unit 120, a learning set speech recognition accuracy calculation unit 130, an acoustic model selection unit 140, and a selected closed acoustic model 150. A voice recognition accuracy calculation unit 160 and a control unit 170. The speech recognition performance estimation apparatus 100 is realized by reading a predetermined program into a computer composed of, for example, a ROM, a RAM, a CPU, and the like, and executing the program by the CPU.

  The evaluation speech data and the like input to the speech recognition performance estimation apparatus 100 is a digital signal that has been digitized at a sampling frequency of 16 kHz, for example. Also, the speech recognition processing performed by the speech recognition performance estimation apparatus 100 is performed in units of frames with 256 discrete points collected, for example, as one frame.

  The closed acoustic model multiple generation unit 110 receives an initial acoustic model and an evaluation set, extracts an acoustic feature amount from evaluation speech data constituting the evaluation set, and models between the acoustic feature amount and the initial acoustic model. N (N ≧ 2) closed acoustic models are generated and output to the closed acoustic model group recording unit 120 (step S110). Here, N is a number (N = 20) representing, for example, 20 of 50, 100, 150,. The N numerical values may be given to the closed acoustic model multiple generation unit 110 as constants in advance, or may be given from the outside as indicated by the broken line in FIG. 2 as the learning parameter τ. Detailed description of the closed acoustic model multiple generation unit 110 will be described later.

  The closed acoustic model multiple generation unit 110 repeats the model generation process until N closed acoustic models are obtained. The control unit 170 controls this repetitive process (No in step S1701). If N is increased, the number of closed acoustic models increases, so that the “recognition accuracy upper limit” can be estimated more accurately. However, if N is increased, the processing time of the speech recognition processing in the learning set speech recognition accuracy calculation unit 130 increases.

The learning set speech recognition accuracy calculation unit 130 obtains N speech recognition accuracies obtained by performing speech recognition on the learning set using the N closed acoustic models recorded in the closed acoustic model group recording unit 120 with the learning set as an input. Output (step S130). The learning set is the above-described learning data, and is composed of a set of learning speech data and a correct text for learning that a person hears the learning speech data and transcribes it into characters. For example, Reference 1 (Masayoshi et al., “Free speech recognition technology“ VoiceRex ”for listening to natural conversations with customers”, NTT Technical Journal, Vol. 18, No. 11, pp. 15-18, 2006. .) Is a well-known existing technology described in. The speech recognition accuracy is a value obtained by subtracting the word error rate from 100, and the word error rate is, for example, Reference 2 (X. Huang, A. Acero and H.-W. Hon, “Spoken”). Language Processing, “Prentice Hall, pp.419-421, 2001”). The learning set speech recognition accuracy calculation unit 130 repeats speech recognition processing until N speech recognition accuracy is obtained. Control of this repetitive processing is performed by the control unit 170 (No in step S1702).

  The acoustic model selection unit 140 receives the N speech recognition accuracies output from the learning set speech recognition accuracy calculation unit 130, refers to the closed acoustic model group recording unit 120, and corresponds to the maximum value of the N speech recognition accuracies. The closed acoustic model to be selected is selected and output as the selected closed acoustic model 150 (step S140).

  The speech recognition accuracy calculation unit 160 outputs the speech recognition accuracy obtained by speech recognition of the evaluation set using the selected closed acoustic model selected by the acoustic model selection unit 140 as an estimated upper limit of recognition accuracy (step S160).

  As described above, according to the speech recognition performance estimation device 100, N acoustic models that are models located between the acoustic feature amount of the evaluation speech data and the initial acoustic model are generated, and the N acoustic models are generated. An acoustic model showing the maximum recognition accuracy with respect to the learning set is selected from the models, and the recognition accuracy obtained by voice recognition of the evaluation set using the acoustic model is output as an estimated value of “recognition accuracy upper limit”. That is, the speech recognition performance estimation apparatus 100 determines how close the acoustic model may be to the acoustic feature amount of the evaluation speech data based on the speech recognition accuracy of the learning set, and the speech recognition accuracy of the learning set is maximized. And an “upper recognition accuracy upper limit” is estimated from the acoustic model.

  As a result, when the speech recognition system was introduced, the recognition rate was calculated from the difference between the speech recognition rate obtained using an arbitrary evaluation set used for recognition accuracy evaluation different from the learning set used for learning the acoustic model and the “recognition accuracy upper limit”. You can know the growth allowance. Therefore, it is possible to prevent the unnecessary transcription of the learning correct text.

  Note that any method may be used to create N acoustic models between the initial acoustic model and the acoustic feature amount of the evaluation voice data. A special technical feature of the present invention is a configuration and method in which a plurality of acoustic models are created in advance between the initial acoustic model and the acoustic feature value of the evaluation speech data to obtain the maximum speech recognition accuracy. Hereinafter, the operation of the speech recognition performance estimation apparatus 100 will be described in more detail by showing a more specific functional configuration example of each unit.

[Closed acoustic model generator]
FIG. 4 shows a functional configuration example of the closed acoustic model multiple generation unit 110. The closed acoustic model multiple generation unit 110 includes an acoustic feature quantity extraction unit 1101, each phoneme section extraction unit 1102, a model parameter update unit 1103, a learning parameter τ 1104, and an output unit 1105.

The acoustic feature quantity extraction unit 1101 extracts an acoustic feature quantity (MFCC: Melt-Frequency Cepstrum Coefficient) x _t from the voice data of the evaluation voice. The acoustic feature amount x _t can be obtained by a known calculation method.

Each phoneme segment extraction means 1102 receives the acoustic feature quantity xt extracted by the acoustic feature quantity extraction means 1101 as input, and refers to the initial acoustic model to represent alphabetic units (a, k, s, t, ..., phoneme sections t _a , t _k , ... corresponding to phonemes substantially corresponding to N) are extracted. Here, “t _a ” with the subscript “a” means a phoneme segment of the phoneme “a”. Hereinafter, a notation such as “a” representing a phoneme is omitted.

The model parameter update unit 1103 receives the initial phone model μ ₀ (average value of feature values) of the same phoneme as the sound feature quantity x _t and the learning parameter τ 1104 as input, and generates a closed sound model μ ^ by the following equation, for example.

  Here, t means a frame, and T means the number of frames.

A closed acoustic model μ ^ is generated for each learning parameter τ for each phoneme. As described above, when the learning parameter τ is 20 numbers in the range of 1 to 1000, for example, 50, 100, 150,..., 1000, the closed acoustic model has μ ₁ ^ to μ ₂₀ ^ for every phoneme. 20 sets are generated. The plurality of closed acoustic models μ ₁ ^ to μ ₂₀ ^ are recorded in the closed acoustic model group recording unit 120 via the output unit 1105. The output unit 1105 transmits the output of the closed acoustic model for each phoneme and each learning parameter τ to the acoustic feature quantity extraction unit 1101 or each phoneme section extraction unit 1102.

The closed acoustic model μ ^ is a weighted average of the initial acoustic model μ ₀ and the acoustic feature amount x _t of the evaluation voice data, as is apparent from the equation (1). The value approaches the initial acoustic model when the learning parameter τ increases, and approaches the acoustic feature amount x _t of the evaluation voice data when the learning parameter τ increases.

The closed acoustic model may be a model in which only the average value of the feature values is changed by Expression (1). In that case, the variance Σ ₀ is left as it is in the initial acoustic model. Or you may make it also change dispersion | distribution (SIGMA) ^ using Formula (2).

  Here, α> k−1, k is the number of dimensions, and ′ means transposition. For the formulas (1) and (2), refer to Reference 3 (Koichi Shinoda “Speaker Adaptation Technology for Speech Recognition Using Stochastic Models”, IEICE Transactions D-II Vol. J87-D-II No .2 pp.371-386 February 2004).

[Learning set speech recognition accuracy calculator]
The learning set speech recognition accuracy calculation unit 130 recognizes speech data for learning in a learning set using a plurality of closed acoustic models recorded in the closed acoustic model group recording unit 120 and a language model, and a plurality of closed acoustic models. A plurality of voice recognition accuracies corresponding to are output. Speech recognition is an existing technology as described above.

  For the speech recognition accuracy, for example, word correct accuracy is used. The word accuracy is obtained by comparing the speech recognition result with the correct text for learning. As described above, the speech recognition accuracy may be obtained from the word error rate (reference document 2).

[Acoustic model selection section]
The acoustic model selection unit 140 receives N speech recognition accuracies corresponding to N acoustic models as input. The acoustic model selection unit 140 refers to the closed acoustic model group recording unit 120 to select a closed acoustic model having a maximum voice recognition accuracy and outputs the selected closed acoustic model 150.

  When there are multiple closed acoustic models with the highest speech recognition accuracy, a method of selecting the model with the largest amount of change from the initial acoustic model among the closed acoustic models with the largest speech recognition accuracy, or the earliest acoustic model There is a method of selecting a model with a small amount of change from, and either method may be selected as long as the same method is always selected.

[Voice recognition accuracy calculation section]
The voice recognition accuracy calculation unit 160 outputs the voice recognition accuracy obtained by voice recognition of the evaluation set using the selected closed acoustic model as an estimated upper limit of the recognition accuracy. The speech recognition accuracy is obtained by matching the correct text for evaluation with the speech recognition result, for example, by subtracting the word error rate calculated by the method described in the above-mentioned Reference Document 2 from 100, The estimated value is the upper limit of recognition accuracy.

FIG. 5 shows a functional configuration example of the speech recognition performance estimation apparatus 200 of the present invention. The speech recognition performance estimation device 200 is different only in that the learning set speech recognition accuracy calculation unit 130 of the speech recognition performance estimation device 100 (FIG. 2) described above is replaced with a learning set speech recognition accuracy calculation unit 230. The learning set speech recognition accuracy calculation unit 230 receives the learning speech data as an input, and recognizes the learning speech data using the N closed acoustic models and language models recorded in the closed acoustic model group recording unit. The N speech recognition accuracies respectively corresponding to the N closed acoustic models are output, and the speech recognition accuracies are output without using the learning correct text.

  FIG. 6 shows a functional configuration example of the learning set speech recognition accuracy calculation unit 230. The learning set speech recognition accuracy calculation unit 230 includes a word confusion network creation unit 2301, a language model 2302, a word alignment network conversion unit 2303, and a word correct accuracy calculation unit 2304.

  The word confusion network creation unit 2301 uses the N acoustic models and the language model 2302 recorded in the closed acoustic model group recording unit 120 to recognize speech speech for learning and convert the recognition result word string to word confusion. Convert to a network (WCN: Word Confusion Network) and output. The word confusion network efficiently represents a plurality of recognition result word strings for a certain utterance.

  The word alignment network conversion unit 2303 converts the word confusion network created by the word confusion network creation unit 2301 into a word alignment network. The word confusion network and the word alignment network can be found in, for example, Reference 4 (Atsunori Ogawa, Takaaki Hori, Atsushi Nakamura, “Estimation of recognition accuracy using the word alignment network and discriminative type classification”, Acoustical Society of Japan 2012 Autumn Meeting, pp. 67-68, 2012.).

  The word correct accuracy calculation means 2304 receives the word alignment network converted by the word alignment network conversion means 2303 and outputs an estimated value WACC of word correct accuracy. The estimated value WACC of the word correct accuracy is calculated so as to correspond to N acoustic models recorded in the closed acoustic model group recording unit 120, and becomes N speech recognition accuracy.

  The estimated value WACC of word correct accuracy is obtained by the following equation.

  Here, E (#C) is the estimated number of correct words obtained from the word alignment network, E (#S) is the estimated number of replacement errors, E (#I) is the estimated number of insertion errors, and E (# #D) is an estimated value of the number of deletion errors.

  The speech recognition performance estimation apparatus 200 can output N speech recognition accuracies even if there is no correct text for learning corresponding to the speech data for learning, and thus has an effect of suppressing the cost required for the creation. The idea of using only the learning speech data as the learning set can be applied to the evaluation set. In this case, the speech recognition performance estimation apparatus 200 ′ (FIG. 5) uses a closed acoustic model multiple generation unit 310, a speech recognition accuracy calculation unit 360, and a learning speech data, which receive evaluation speech data instead of an evaluation set. The learning set speech recognition accuracy calculation unit 130 that receives the learning set is included. The operations of the closed acoustic model multiple generation unit 310 and the speech recognition accuracy calculation unit 360 will be described in a third embodiment.

  FIG. 7 shows a functional configuration example of the speech recognition performance estimation apparatus 300 of the present invention. In the speech recognition performance estimation device 300, the closed acoustic model multiple generation unit 110 of the speech recognition performance estimation device 200 (FIG. 5) described above is the closed acoustic model multiple generation unit 310, and the speech recognition accuracy calculation unit 160 is the speech recognition accuracy calculation unit. 360 is different in that each is replaced.

  The closed acoustic model multiple generation unit 310 receives the initial acoustic model and the evaluation voice data, and extracts an acoustic feature amount from the voice recognition result obtained by voice recognition of the evaluation voice data using the initial acoustic model and the evaluation voice data. Then, N (N ≧ 2) closed acoustic models that are models between the acoustic feature quantity and the initial acoustic model are generated and output to the closed acoustic model group recording unit 120. The closed acoustic model multiple generation unit 310 is different from the closed acoustic model multiple generation unit 110 in that the speech recognition result obtained by performing speech recognition using the initial acoustic model is handled and processed in the same manner as the above-described correct text for evaluation.

  The speech recognition accuracy calculation unit 360 receives the speech data for evaluation, and the speech recognition accuracy obtained by performing speech recognition on the speech data for evaluation using the selected closed acoustic model 150 and the language model, as an estimated upper limit of the recognition accuracy. Output. The language model 2302 described above can be used as the language model. Similar to the learning set speech recognition accuracy calculation unit 230 described above, the speech recognition accuracy calculation unit 360 outputs speech recognition accuracy without using correct text.

  The speech recognition performance estimation apparatus 300 can output an estimated value of the upper limit of recognition accuracy without the learning correct text and the evaluation correct text, and thus has an effect of suppressing the cost required to create them.

  FIG. 8 shows a functional configuration example of the speech recognition performance estimation apparatus 400 of the present invention. The speech recognition performance estimation device 400 includes the speech recognition performance estimation devices 100, 200, 200 ′, and 300 described above, a data set selection unit 480, and an averaging unit 490.

  The data set selection unit 480 receives a plurality of data sets composed of a set of voice data and correct text and a plurality of data sets composed only of the voice data, and inputs two data sets other than the selected data set of the plurality of data sets. When one set is selected and correct text is given to both of the two selected data sets, one data set is output to the speech recognition performance estimation apparatus 100 as a learning set and the other data set as an evaluation set. If the correct text is given to one of the two selected data sets, the data set to which the correct text is given is used as the evaluation set or learning set, and the other is used as the learning voice data or evaluation voice data. Output to the recognition performance estimation devices 200 and 200 ′, and correct text for any of the two selected data sets. There performs count operation a predetermined output to the speech recognition performance estimator 300 one if not granted training speech data, as voice data for evaluation of the other. The predetermined number of times may be one. Or it sets to about 3-10 according to calculation time. Alternatively, the maximum value M of the number of combinations of data sets may be used.

  For example, assume that there are two data sets that contain correct text and two data sets that do not have correct text. In this case, there are two cases in the speech recognition performance estimation apparatus 100: one is a learning set and the other is an evaluation set, and the other is an evaluation set and the other is a learning set. In this case, there are two cases in the speech recognition performance estimation apparatus 300 as in the speech recognition performance estimation apparatus 100. Thus, the predetermined number of times is the number of times that may differ between the speech recognition performance estimation devices 100, 200, 200 ′, and 300.

  The averaging unit 490 receives the estimated values of the recognition accuracy upper limit output from the speech recognition performance estimation apparatuses 100, 200, and 300, respectively, and calculates and outputs the average value of all the recognition accuracy upper limit estimated values.

  According to the speech recognition performance estimation apparatus 400, an upper limit of recognition accuracy can be obtained from a data set in which a plurality of data sets without correct text and a data set with correct text are mixed. By averaging the estimated values obtained from the combinations, it is possible to increase the estimation accuracy as compared with the case where each of the speech recognition performance estimation devices 100, 200, 200 ′, 300 is operated once. Further, in a development environment in which a data set composed of a set of speech data and correct text is not always prepared, the speech recognition performance estimation device 400 estimates an upper limit of recognition accuracy from a data set of only speech data. The value can be determined. That is, the speech recognition performance estimation apparatus 400 also has an excellent effect of operating regardless of the data set format.

  As the speech recognition performance estimation device 400, the configuration of the speech recognition performance estimation devices 200 and 200 ′ may be provided with either one. If either one of the configurations is provided, even if two data sets are selected, a data set consisting of a set of voice data and correct text and a data set consisting only of voice data, the estimated upper limit of recognition accuracy is set. It is possible to ask.

FIG. 9 shows a functional configuration example of the speech recognition performance estimation apparatus 500 of the present invention. The operation flow is shown in FIG. The speech recognition performance estimation device 500 includes a closed acoustic model multiple generation unit 510, a closed acoustic model group recording unit 120, a speech recognition accuracy calculation unit 560, a pinch out control unit 580, a recognition accuracy upper limit output unit 540, and a control. Part 570.

  Similar to the above-described speech recognition performance estimation device 100 and the like, the speech recognition performance estimation device 500 reads a predetermined program into a computer including, for example, a ROM, a RAM, and a CPU, and the CPU executes the program. It is realized with.

  The closed acoustic model multiple generation unit 510 receives the initial acoustic model, the evaluation set, the sandwiched control signal, and the speech recognition accuracy, extracts an acoustic feature amount from the evaluation speech data constituting the evaluation set, and extracts the acoustic feature. A closed acoustic model having an upper limit on the amount and a lower limit on the initial acoustic model is generated according to the sandwiching control signal and output to the closed acoustic model group recording unit 120, and the speech recognition accuracy calculation unit 560 is configured to output the closed acoustic model. The voice recognition accuracy obtained using the acoustic model is assigned to the corresponding closed acoustic model that has been output (step S510). The pinch-in control signal is a control signal corresponding to the learning parameter τ described above, and is a control signal that changes based on the pinch-out method of a well-known numerical calculation method and is set so that the value range is sequentially narrowed.

The speech recognition accuracy calculation unit 560 outputs the speech recognition accuracy of speech recognition using the learning set as an input, using the closed acoustic model in the order recorded in the closed acoustic model group recording unit 120 (step S560). .

  The pinch-out controller 580 receives the voice recognition accuracy as an input, variably outputs the pinch-in control signal so that the voice recognition accuracy is maximized, and uses the variable pinch-out control signal to generate a plurality of closed acoustic model generation units 510. The process of operating the voice recognition accuracy calculation unit 560 is repeated a predetermined number of times (step S580).

  The recognition accuracy upper limit value output unit 540 searches the closed acoustic model recorded in the closed acoustic model group recording unit 120 for the maximum speech recognition accuracy, and uses the maximum speech recognition accuracy value as an estimated recognition accuracy upper limit value. (Step S540).

  Unlike the speech recognition performance estimation apparatus 100 or the like that calculates an estimated value of the upper limit of recognition accuracy after generating N closed acoustic models in advance, the speech recognition performance estimation apparatus 500 is based on the sandwiching method of numerical calculation methods. This method searches for a closed acoustic model with high speech recognition accuracy, and can reduce the amount of calculation.

  As described above, the speech recognition performance estimation device according to the present invention determines how close the acoustic model can be to the acoustic feature amount of the evaluation speech data, and maximizes the speech recognition accuracy of the learning set. Is generated. Then, the speech recognition accuracy obtained by performing speech recognition on the evaluation set using the acoustic model is used as an estimated value of the recognition accuracy upper limit. As a result, it is possible to solve the problem that the “recognition accuracy upper limit” obtained by the closed evaluation of the prior art is higher than the actual “recognition accuracy upper limit”.

  Note that the idea of generating N closed acoustic models as the weighted average value corresponding to the learning parameter τ described in the first embodiment can be applied to the second and third embodiments. Further, the idea of the speech recognition performance estimation apparatus 500 using the numerical calculation method of the pinch-out method can be applied to the second and third embodiments.

  When the processing means in the above apparatus is realized by a computer, the processing contents of the functions that each apparatus should have are described by a program. Then, by executing this program on the computer, the processing means in each apparatus is realized on the computer.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

Claims (8)

The initial acoustic model and the evaluation set are input, the acoustic feature quantity is extracted from the evaluation speech data constituting the evaluation set, and N (N ≧ 2) that is a model between the acoustic feature quantity and the initial acoustic model Closed acoustic model multiple generation unit that generates a single closed acoustic model and outputs the closed acoustic model group recording unit,
A learning set speech recognition accuracy calculation unit that outputs N speech recognition accuracies obtained by performing speech recognition using the N closed acoustic models recorded in the closed acoustic model group recording unit with the learning set as an input; ,
The sound to be output as a selected closed acoustic model by selecting the closed acoustic model corresponding to the maximum value of the N speech recognition accuracy by referring to the closed acoustic model group recording unit with the N speech recognition accuracy as an input A model selector,
A speech recognition accuracy calculation unit that outputs speech recognition accuracy obtained by speech recognition of the evaluation set using the selected closed acoustic model as an estimated upper limit of recognition accuracy;
A speech recognition performance estimation device comprising: The initial acoustic model and the evaluation set are input, the acoustic feature quantity is extracted from the evaluation speech data constituting the evaluation set, and N (N ≧ 2) that is a model between the acoustic feature quantity and the initial acoustic model Closed acoustic model multiple generation unit that generates a single closed acoustic model and outputs the closed acoustic model group recording unit,
Using the learning speech data as an input, the learning speech data is speech-recognized using the N closed acoustic models and the language model recorded in the closed acoustic model group recording unit , and the N closed acoustic models are obtained. A learning set speech recognition accuracy calculation unit for outputting N speech recognition accuracy corresponding to each,
The sound to be output as a selected closed acoustic model by selecting the closed acoustic model corresponding to the maximum value of the N speech recognition accuracy by referring to the closed acoustic model group recording unit with the N speech recognition accuracy as an input A model selector,
A speech recognition accuracy calculation unit that outputs speech recognition accuracy obtained by speech recognition of the evaluation set using the selected closed acoustic model as an estimated upper limit of recognition accuracy;
A speech recognition performance estimation device comprising: Using the initial acoustic model and the evaluation voice data as input, the acoustic feature value is extracted from the voice recognition result obtained by voice recognition of the evaluation voice data using the initial acoustic model and the evaluation voice data. A closed acoustic model multiple generation unit that generates N (N ≧ 2) closed acoustic models that are models between the models and outputs the closed acoustic model group recording unit;
Using the learning speech data as an input, the learning speech data is speech-recognized using the N closed acoustic models and the language model recorded in the closed acoustic model group recording unit , and the N closed acoustic models are obtained. A learning set speech recognition accuracy calculation unit for outputting N speech recognition accuracy corresponding to each,
The sound to be output as a selected closed acoustic model by selecting the closed acoustic model corresponding to the maximum value of the N speech recognition accuracy by referring to the closed acoustic model group recording unit with the N speech recognition accuracy as an input A model selector,
Using the selected closed acoustic model, the speech recognition accuracy for speech recognition of the evaluation set using the evaluation correct text output from the closed acoustic model multiple generation unit and the evaluation speech data input from the outside as an evaluation set, A speech recognition accuracy calculation unit that outputs an estimated value of the recognition accuracy upper limit;
A speech recognition performance estimation device comprising: The speech recognition performance estimation device according to claim 1 is a first speech recognition performance estimation device,
The speech recognition performance estimation device according to claim 2 is replaced with a second speech recognition performance estimation device,
The speech recognition performance estimation device according to claim 3 as a third speech recognition performance estimation device,
Select two data sets other than the selected data set of multiple data sets, with multiple data sets consisting of a set of voice data and correct text and multiple data sets consisting only of voice data as inputs, When correct text is given to both of the two selected data sets, one data set is output as a learning set and the other data set is output as an evaluation set to the first speech recognition performance estimation device, and the selected 2 When the correct text is given to one of the data sets, the data set to which the correct text is given is output to the second speech recognition performance estimation device as the evaluation set, and the other is used as the speech data for learning. If the correct text is not assigned to either of the two data sets, one is used for learning speech data and the other is used for evaluation speech data. A data set selection unit that performs count operation a predetermined output to the third speech recognition performance estimator as data,
The estimation value upper limit of recognition accuracy output from the first speech recognition performance estimation device, the estimation value upper limit of recognition accuracy output from the second speech recognition performance estimation device, and the recognition accuracy upper limit output from the third speech recognition performance estimation device. And an average unit for calculating and outputting an average value of all the recognition accuracy upper limit estimated values,
A speech recognition performance estimation device comprising: The speech recognition performance estimation apparatus according to any one of claims 1 to 3,
The closed acoustic model multiple generator is
N closed acoustic models are generated with a weighted average value corresponding to N (N ≧ 2) learning parameters τ in a range where the initial acoustic model parameter is the lower limit and the acoustic feature parameter is the upper limit. An apparatus for estimating speech recognition performance, wherein the apparatus is for outputting. The initial acoustic model, the evaluation set, the sandwiched control signal, and the voice recognition accuracy are input, and the acoustic feature quantity is extracted from the evaluation voice data constituting the evaluation set. The closed acoustic model of the lower limit range is generated according to the pinching control signal and output to the closed acoustic model group recording unit, and the speech recognition accuracy calculated by the speech recognition accuracy calculation unit using the closed acoustic model A closed acoustic model multiple generation unit for assigning to the corresponding closed acoustic model that has been output,
A speech recognition accuracy calculation unit that outputs speech recognition accuracy obtained by performing speech recognition using the closed acoustic models in the order recorded in the closed acoustic model group recording unit with the learning set as an input;
With the voice recognition accuracy as an input, the pinch control signal is variably output so that the voice recognition accuracy is maximized, and the closed acoustic model multiple generation unit and the voice recognition accuracy calculation unit are output with the variable pinch control signal. A sandwiching control unit that repeats the process of operating a predetermined number of times,
A recognition accuracy upper limit output unit that searches for the maximum speech recognition accuracy from among the closed acoustic models recorded in the closed acoustic model group recording unit and outputs the maximum speech recognition accuracy value as an estimated upper limit of recognition accuracy When,
A speech recognition performance estimation device comprising: A closed acoustic model multiple generation unit receives an initial acoustic model and an evaluation set, extracts an acoustic feature amount from audio data for evaluation constituting the evaluation set, and a model between the acoustic feature amount and the initial acoustic model A closed acoustic model multiple generation process for generating N (N ≧ 2) closed acoustic models and outputting them to the closed acoustic model group recording unit;
The learning set speech recognition accuracy calculation unit outputs N speech recognition accuracy obtained by performing speech recognition on the learning set using the N closed acoustic models recorded in the closed acoustic model group recording unit. Learning set speech recognition accuracy calculation process,
The acoustic model selection unit selects the closed acoustic model corresponding to the maximum value of the N speech recognition accuracies by referring to the closed acoustic model group recording unit using the N speech recognition accuracies as an input. Acoustic model selection process to be output as an acoustic model;
A speech recognition accuracy calculation process in which a speech recognition accuracy calculation unit outputs speech recognition accuracy obtained by performing speech recognition of the evaluation set using the selected closed acoustic model as an estimated upper limit of recognition accuracy;
A speech recognition performance estimation method comprising:   A program for causing a computer to operate as the speech recognition performance estimation device according to any one of claims 1 to 6.

Images