Control device, control method, voice processing system, and program

Abstract

Provided is a control device capable of selectively using an emphasis signal and an observation signal, or dynamically controlling the ratio of the emphasis signal and the observation signal, so as to improve the performance of a post-stage task. The control device: acquires an output value that indicates, in accordance with the ratio of an observation signal added to an enhancement signal, whether or not speech enhancement should be performed on an observation signal used for a post-stage task the performance of which changes, or indicates the degree to which the speech enhancement should be performed; determines the ratio of the observation signal and an enhancement signal generated by the speech enhancement under a prescribed condition, using the acquired output value; and determines an input signal to be used for a post-stage task.

Description

Control device, control method, audio processing system, and program

[0001] The present invention relates to a technique for controlling the intensity of speech enhancement.

[0002] Thanks to advances in deep learning technology, the performance of speech enhancement has improved dramatically in recent years. There are two main applications of speech enhancement. One is to use it before a machine learning model, such as speech recognition, to convert inputs containing undesired signals into clean signals, thereby improving the performance of subsequent processing. The other is to use speech enhancement to obtain clean signals for human listening.

[0003] As prior art, when using speech enhancement as a machine learning model, particularly as a pre-stage of a speech recognizer, a technology has been proposed that can further improve the performance of a speech recognizer by adding an enhancement signal output by the speech enhancer and the original speech before enhancement in an appropriate ratio (Patent Document 1, Non-Patent Document 1).

[0004] Non-Patent Document 1 focuses on the fact that the output of a speech enhancer may or may not improve the performance of a speech recognizer, and proposes using a method such as machine learning to determine the proportion of the emphasis signal to be used depending on the speech recognition performance for both the emphasis signal and the unemphasized observed signal.

[0005] International Publication No. 2023 / 105778

[0006] Sato, Hiroshi, et al., "Learning to enhance or not: Neural network-based switching of enhanced and observed signals for overlapping speech recognition", ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022.

[0007] However, Non-Patent Document 1 is a method that assumes the use of a speech recognizer in a subsequent stage, and is therefore not suitable for building a speech enhancer intended for human listening, for example.

[0008] An object of the present invention is to provide a control device, a control method, a speech processing system, and a program that dynamically control the use of an emphasis signal and an observation signal, or the ratio between the emphasis signal and the observation signal, thereby improving the performance of downstream tasks.

[0009] In order to solve the above problem, according to one aspect of the present invention, a control device acquires an output value indicating whether or not speech enhancement should be performed on an observation signal used in a subsequent task whose performance changes depending on the proportion of an observation signal added to an enhancement signal, or the degree to which the speech enhancement should be performed, and uses the acquired output value to determine the proportion between the observation signal and the enhancement signal generated by the speech enhancement under specified conditions, thereby determining an input signal to be used in the subsequent task.

[0010] According to the present invention, it is possible to selectively use an emphasis signal and an observation signal, or to dynamically control the ratio between the emphasis signal and the observation signal, thereby achieving an effect of improving the performance of subsequent tasks.

[0011] FIG. 1 is a diagram showing an example of the configuration of an audio signal processing system according to the first embodiment. FIG. 2 is a functional block diagram of a label generation device according to the first embodiment. FIG. 3 is a diagram showing the processing flow of the label generation device according to the first embodiment. FIG. 4 is a functional block diagram of a model learning device according to the first embodiment. FIG. 5 is a diagram showing the processing flow of the model learning device according to the first embodiment. FIG. 6 is a functional block diagram of an audio signal processing device according to the first embodiment. FIG. 7 is a diagram showing the processing flow of the audio signal processing device according to the first embodiment. FIG. 8 is a diagram showing an example of the configuration of an audio signal processing system according to the second embodiment. FIG. 9 is a functional block diagram of a learning device according to the second embodiment. FIG. 10 is a diagram showing the processing flow of the learning device according to the second embodiment. FIG. 11 is a diagram showing verification results of switching for improving auditory quality. FIG. 12 is a diagram showing an example of the functional configuration of a computer.

[0012] An embodiment of the present invention will be described in detail below. Components having the same function are assigned the same number, and duplicate explanations will be omitted. In the following description, symbols such as "^" and "~" used in the text should normally be written directly above the character immediately following them, but due to limitations in text notation, they are written immediately before the character in question. In formulas, these symbols are written in their original positions. Furthermore, unless otherwise specified, processing performed on each element of a vector or matrix is ​​assumed to apply to all elements of that vector or matrix.

[0013] <Key Points of First Embodiment> Based on the performance of the emphasis signal, the observation signal, and the signal obtained by weighting and adding them in a target task, an appropriate weighting ratio between the emphasis signal and the observation signal is obtained as a correct answer.

[0014] The weighting ratio is determined by machine learning or rules depending on the observed signal, the emphasized signal, the application domain, the SNR estimation result, etc.

[0015] By solving the subsequent task (hereinafter simply referred to as the subsequent task) using a signal obtained by appropriately weighting and adding the observed signal and the emphasized signal, it is possible to maximize the performance improvement due to the addition of the original sound.

[0016] The subsequent tasks are not limited to those executed by a device. For example, by constructing a method for estimating the weighting ratio using the human auditory quality estimated by the well-known DNSMOS (see Reference 1), it is possible to assign weights appropriate for human hearing.

[0017] (Reference 1) Reddy, Chandan KA, Vishak Gopal, and Ross Cutler, "DNSMOS P. 835: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors," ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022. <First embodiment> FIG. 1 shows an example configuration of a speech signal processing system according to a first embodiment. Processing in the speech signal processing system consists of a learning stage and a speech processing stage, and the learning stage consists of two stages: a label generation stage and a model learning stage.

[0018] The speech signal processing system includes a learning device 100 that performs processing in a learning stage and a speech signal processing device 200 that performs processing in a speech processing stage. Furthermore, the learning device 100 includes a label generation device 110 that performs a label generation stage and a model learning device 120 that performs a model learning stage.

[0019] <Label Generation Apparatus 110> FIG. 2 is a functional block diagram of the label generation apparatus 110, and FIG. 3 shows the processing flow thereof.

[0020] The label generation device 110 receives training observation signals as input, generates emphasis signals (observation signals that have undergone speech enhancement processing) from the training observation signals, and generates and outputs switching labels using the observation signals and the emphasis signals. Note that the switching labels indicate the proportion of appropriate observation signals in a signal (hereinafter also referred to as a weighted sum signal) obtained by adding the observation signals to the emphasis signals, corresponding to the subsequent task to be solved.

[0021] The label generating device 110 includes a speech enhancing unit 111 , a performance calculating unit 113 , and a switching label generating unit 114 .

[0022] <Speech Enhancement Unit 111> In step S111, the speech enhancement unit 111 performs speech enhancement processing. That is, the speech enhancement unit 111 acquires an observed signal as input, extracts only desired speech from the acquired observed signal using a known speech enhancement technique, and executes speech enhancement processing. A known target speaker extraction technique, for example, can be used as a method for extracting the desired speech. As shown in FIG. 2, the target speaker extraction technique is a technique in which the speech enhancement unit 111 acquires auxiliary information about the target speaker in addition to acquiring the observed signal, thereby extracting only the speech of the target speaker from the observed signal. The auxiliary information about the target speaker can be, for example, an utterance registered in advance by the target speaker. Furthermore, known noise suppression technology, echo suppression technology, dereverberation suppression technology, or a combination thereof can also be used as the speech enhancement technology.

[0023] The speech enhancement unit 111 may use the speech waveform itself obtained from the observed signal, or may use features extracted from the observed signal, as the input signal. The speech enhancement unit 111 outputs the speech signal that has been subjected to speech enhancement processing (hereinafter also referred to as the “enhanced signal”) to the performance calculation unit 113.

[0024] <Performance Calculation Unit 113> In step S113, the performance calculation unit 113 receives the observed signal and the emphasis signal as input. The performance calculation unit 113 also receives the correct label for the subsequent task as input, if necessary.

[0025] The performance calculation unit 113 generates a weighted sum signal by adding the observed signal to the emphasized signal at various ratios. For example, if the ratio j is between 0 and 1, the emphasized signal is ^S, the observed signal is Y, and the weighted sum signal ~S, then the weighted sum signal ~S can be expressed by the following equation: ~S=(1-j)^S+jY When the ratio at which the observed signal Y is added to the emphasized signal ^S is 0, the weighted sum signal ~S matches the emphasized signal ^S, and when the ratio at which the observed signal Y is added to the emphasized signal ^S is 1, the weighted sum signal ~S matches the observed signal Y. For example, six ratios (0, 0.2, 0.4, 0.6, 0.8, 1) are prepared in increments of 0.2 from 0 to 1, and six types of weighted sum signals are generated that include the emphasized signal itself and the observed signal itself.

[0026] The performance calculation unit 113 executes a subsequent task on the generated weighted sum signal. The subsequent task is a task that is performed after the process of switching between the observed signal and the emphasis signal using the switching label or the process of generating the weighted sum signal, and may be any task whose performance changes depending on the proportion of the observed signal added to the emphasis signal, such as speaker identification, sound source direction estimation, sound source position estimation, human listening, etc.

[0027] The performance calculation unit 113 calculates an index indicating the performance of the subsequent task, and outputs it to the switching label generation unit 114 .

[0028] For example, when speaker identification is a subsequent task, the performance calculation unit 113 receives as input a correct label indicating the correct speaker, and calculates and outputs an index indicating speaker identification performance using the speaker identification results of each of the multiple weighted sum signals and the correct label. In this case, the posterior probability for the correct speaker class or the like can be used as an index indicating speaker identification performance. For sound source direction estimation and sound source position estimation, an index indicating the performance of the subsequent task can be calculated in a similar manner using a correct label indicating the correct direction or position.

[0029] Furthermore, when the subsequent task is human listening, for example, the performance calculation unit 113 calculates and outputs the DNSMOS (see Reference 1) value of each of the multiple weighted sum signals as an index indicating human listening performance. In this case, a correct answer label is not required. Note that, since there is a correlation between DNSMOS and ease of human listening, DNSMOS can be used as an index indicating human listening performance.

[0030] <Switching Label Generation Unit 114>In step S114, the switching label generation unit 114 takes as input the indicators representing the performance of the subsequent task for each of the plurality of weighted addition signals acquired from the performance calculation unit 113, and based on these values, calculates and outputs a switching label k corresponding to the subsequent task. As described above, the switching label k indicates the ratio of the appropriate observation signal in the weighted addition signal corresponding to the subsequent task. The switching label generated here is used as a teacher label for the optimization unit 122 described later to optimize the parameters of the switching model described later.

[0031] (Switching Label Generation Method 1) For example, the switching label generation unit 114 sets the weight corresponding to the one with the best performance as the switching label k based on the indicators representing the performance of the subsequent task for each of the plurality of weighted addition signals. For example, when speaker identification is the subsequent task and the posterior probability for the correct speaker class is used as the indicator representing the speaker identification performance, the weight (the ratio of adding the observation signal to the emphasis signal) corresponding to the weighted addition signal with the highest posterior probability is set as the switching label k. Also, for example, when human listening is the subsequent task and DNSMOS is used as the indicator representing the performance of human listening, the weight corresponding to the weighted addition signal with the highest DNSMOS is set as the switching label k.

[0032] (Switching Label Generation Method 2) Also, it can be determined based on rules from the performance of each of the plurality of weighted addition signals including the emphasis signal itself and the observation signal itself. For example, the SNR (Signal to Noise Ratio) of the observation signal can be calculated, and the switching label k can be calculated according to the magnitude relationship between the SNR and the threshold. When the SNR of the observation signal is small, it is assumed that the performance of the subsequent task is good even when using the observation signal itself, and rules are set such that the ratio of the observation signal increases as the SNR of the observation signal decreases. For example, L - 1 thresholds T(1), T(2), …, T(L - 1) (where T(1) < T(2) < … < T(L - 1)) are prepared, and C representing the SNR of the observation signal obsIf T(i) is less than T(i), the ratio of observed signals is set to the maximum of 1, and T(i)≦C obs If T(i+1), the ratio of the observed signal is the i+1th (i=1, 2, ..., L-2) largest, and C obs If is equal to or greater than T(L-1), the ratio of observed signals is set to the smallest value of 0.

[0033] Furthermore, for example, if human listening is the subsequent task, the DNSMOS of the observed signal may be calculated, and the switching label k may be calculated based on the relationship between the magnitude of the DNSMOS and a threshold. If the DNSMOS of the observed signal is high, it is assumed that the performance of the subsequent task will be good even if the observed signal itself is used, and a rule is set so that the higher the DNSMOS of the observed signal, the higher the proportion of the observed signal.

[0034] Through the above process, the label generation device 110 generates and outputs a combination of three types of information: the observed signal, the emphasis signal, and the switching label. Note that this combination may include auxiliary information about the target speaker, if necessary.

[0035] Next, the model learning stage will be described.

[0036] <Model Learning Device 120> FIG. 4 is a functional block diagram of the model learning device 120, and FIG. 5 shows the processing flow thereof.

[0037] The model learning device 120 has a switching model unit 121 and an optimization unit 122. The model learning device 120 performs learning by subjecting a model created by the switching model unit 121 to optimization processing in the optimization unit 122. After learning by the model learning device 120, the switching model unit 121 is used by the switching model unit 205 as a trained model to be used in the audio signal processing device 200. The model learning device 120 performs processing of each step illustrated in Fig. 5, thereby realizing the learning process of the switching model. Hereinafter, a learning method for the switching model of the embodiment will be described with reference to Fig. 5.

[0038] <Switching Model Unit 121> In step S121, the switching model unit 121 receives as input the training observation signals, emphasis signals, meta information, or a combination thereof, estimates a switching label using the switching model being trained, and outputs the estimated value ^k to the optimization unit 122. The switching model is a model constructed by machine learning, and receives as input the training observation signals, emphasis signals, meta information, or a combination thereof, estimates a switching label, and outputs the estimated value. The switching model estimates the switching label using initial parameters or parameters optimized by the optimization unit 122 described below. The switching model constructed by machine learning is a model trained using, for example, a well-known technique such as a deep neural network. Possible meta information includes, for example, the domain and SNR estimation result related to the training observation signals and emphasis signals. The estimated value ^k can be a scalar value ranging from 0 to 1, as defined by the following equation:

[0039]

[0040] The switching model unit 121 may be configured to calculate the output ^k as a time-series vector. By calculating the output ^k as a time-series vector, it is possible to adopt a different ratio for each time, allowing for more detailed settings.

[0041] <Optimization Unit 122> In step S122, the optimization unit 122 receives the estimated value ^k from the switching model unit 121 and the switching label created by the label generation device 110, optimizes the parameters of the switching model, and outputs the optimized parameters to the switching model unit 121. The processing between the switching label estimation processing by the switching model unit 121 and the parameter optimization by the optimization unit 122 may be configured to be a loop processing, and the optimization may be completed by repeating these processing. In either case, when the optimization is completed and the parameters are determined, the contents are reflected in the switching model unit 121, and the switching model is completed.

[0042] The specific method of optimization by the optimization unit 122 is as follows: The optimization unit 122 calculates a loss function between the switching label k generated by the label generation device 110 and the estimated value ^k calculated by the switching model unit 121, and optimizes the model parameters included in the switching model unit 121 by minimizing the loss function. As the loss function, for example, the well-known cross-entropy loss used in Patent Document 1 and the like can be used.

[0043] The above has described the learning method of the switching model unit 121 through the processing of the switching model unit 121 and the optimization unit 122. The learned switching model is used as a switching model in the switching model unit 205 of the audio signal processing device 200.

[0044] <Audio Signal Processing Device 200> FIG. 6 is a functional block diagram of the audio signal processing device 200, and FIG. 7 shows the processing flow thereof.

[0045] The speech signal processing device 200 receives an observation signal to be processed as input, generates an enhancement signal (an observation signal that has been subjected to speech enhancement processing) from the observation signal to be processed, calculates an estimate of a switching label using a trained switching model, calculates a weighted sum signal from the calculated estimate, the observation signal, and the enhancement signal, and outputs the weighted sum signal. With this configuration, it is possible to switch between the observation signal and the enhancement signal to be used in a subsequent task using the trained switching model, or to calculate a switching label (the proportion of appropriate observation signals in the weighted sum signal that corresponds to the subsequent task) using the trained switching model, thereby improving the performance of the subsequent task.

[0046] The audio signal processing device 200 includes an audio enhancement unit 201 , a switching model unit 205 , and a signal integration unit 207 .

[0047] <Speech Emphasis Unit 201 > The speech emphasizing unit 201 performs the same processing as the speech emphasizing unit 111 .

[0048] In step S201 , the speech enhancement unit 201 receives an observation signal to be processed as an input, performs speech enhancement processing, and outputs the enhancement signal to the switching model unit 205 .

[0049] <Switching Model Unit 205> The switching model unit 205 performs the same processing as the switching model unit 121.

[0050] In step S205, the switching model unit 205 receives as input the observed signal to be processed, the emphasis signal, the meta information, or a combination thereof, estimates the switching label using the trained switching model, and outputs the estimated value ^k to the signal integration unit 207.

[0051] <Signal Combining Unit 207> In S207, the signal combining unit 207 receives the observed signal to be processed, the enhanced signal, and the estimated value ^k as input, generates a weighted sum signal using the estimated value ^k, and outputs it.

[0052] When using an emphasis signal and an observation signal selectively, there are two types of weighted addition signals: an emphasis signal and an observation signal.

[0053] Here, if the weighted sum signal is ~S, then the weighted sum signal ~S is determined to be either the emphasis signal ^S or the observation signal Y, as defined by the following equation. In equation (2), λ is a preset value in the range of 0 < λ < 1, such as 0.5. In this embodiment, this method of determining either the emphasis signal ^S or the observation signal Y as the weighted sum signal ~S is referred to as a "hardware method."

[0054]

[0055]

[0056] The weighted sum signal ~S may be determined by weighting and adding the emphasis signal ^S and the observation signal Y using the estimated value ^k of the switching model unit 205, as defined by the following equation. In this embodiment, the method of determining the weighted sum signal ~S by weighting and adding the emphasis signal ^S and the observation signal Y using the estimated value ^k is referred to as the "software method."

[0057]

[0058] The signal combining unit 207 outputs the combined signal S by a hardware or software method.

[0059] <Effects> With this configuration, it is possible to selectively use the emphasis signal and the observation signal, or to dynamically control the ratio between the emphasis signal and the observation signal, thereby improving the performance of subsequent tasks.

[0060] In this embodiment, when the subsequent task is human listening, DNSMOS is used as an index showing human listening performance. However, the two indices used in calculating DNSMOS (an index showing the degree of signal distortion and an index showing the degree of noise cancellation) may be used separately. For example, the switching label generation unit 114 may use these two indices when determining switching labels based on performance on a rule basis, and by setting separate rules for each index, more detailed rules can be set.

[0061] A device including the speech enhancement unit 201 and the switching model unit 205 can be said to be a device that dynamically controls the use of an enhancement signal and an observation signal, or the ratio between the enhancement signal and the observation signal, and is therefore also called a control device. The switching label, which is the output value of the control device, can be used to improve the performance of subsequent tasks. <Second Embodiment> The following description will focus on the differences from the first embodiment.

[0062] In this embodiment, the switching model is trained directly on the basis of the desired downstream task.

[0063] 8 shows an example of the configuration of an audio signal processing system according to the second embodiment. The processing in the audio signal processing system consists of a learning stage and an audio processing stage.

[0064] The audio signal processing system includes a learning device 300 that performs processing in the learning stage and an audio signal processing device 200 that performs processing in the audio processing stage. The audio signal processing device 200 has the same configuration as in the first embodiment. The learning device 300 will be described below.

[0065] <Learning Device 300> FIG. 9 is a functional block diagram of the learning device 300, and FIG. 10 shows the processing flow thereof.

[0066] The learning device 300 includes a speech enhancement unit 301 , a switching model unit 302 , a signal integration unit 303 , a loss calculation unit 304 , and an optimization unit 305 .

[0067] <Speech Emphasis Unit 301 > The speech emphasizing unit 301 performs the same processing as the speech emphasizing unit 111 .

[0068] In step S301 , the speech enhancement unit 301 acquires an observed signal for learning as an input, performs speech enhancement processing, and outputs the enhancement signal to the switching model unit 302 .

[0069] <Switching Model Unit 302 > The switching model unit 302 performs the same processing as the switching model unit 121 .

[0070] In step S302, the switching model unit 302 receives as input the observed signal of the learning amount, the emphasis signal, the meta information, or a combination thereof, estimates the switching label using the switching model being trained, and outputs the estimated value ^k to the signal integration unit 303.

[0071] <Signal Combining Unit 303 > The signal combining unit 303 performs the same processing as the signal combining unit 207 .

[0072] In S303, the signal integration unit 303 receives the observed signal for learning, the enhancement signal, and the estimated value ^k as input, generates a weighted sum signal using the estimated value ^k, and outputs it.

[0073] <Loss Calculation Unit 304> In step S304, the loss calculation unit 304 receives the weighted sum signal generated by the signal integration unit 303. Furthermore, the loss calculation unit 304 receives, as input, a correct label for a subsequent task, as necessary.

[0074] The loss calculation unit 304 calculates how superior the weighted sum signal is in terms of the subsequent task, and calculates and outputs a loss based on the calculation. For example, the loss calculation unit 304 executes a subsequent task on the weighted sum signal and calculates the loss between the result of the subsequent task and the correct label. For example, if speaker identification is the subsequent task, the loss calculation unit 304 receives as input a correct label indicating the correct speaker and calculates a loss function between the speaker identification result of the weighted sum signal and the correct label. As the loss function, for example, the well-known cross-entropy loss described in the optimization unit 122 of the first embodiment can be used. For example, if human listening is the subsequent task, the loss calculation unit 304 calculates the DNSMOS of the weighted sum signal. As in the first embodiment, the subsequent task may be any task whose performance changes depending on the proportion of observed signals added to the enhancement signal, such as speaker identification, sound source direction estimation, or sound source location estimation.

[0075] <Optimization Unit 305> In S305, the optimization unit 305 receives the loss calculated by the loss calculation unit 304, optimizes the parameters of the switching model, and outputs the optimized parameters to the switching model unit 302. For example, the optimization unit 305 optimizes the model parameters included in the switching model unit 302 by minimizing a loss function such as cross-entropy loss. Furthermore, for example, the optimization unit 305 optimizes the model parameters included in the switching model unit 302 by maximizing DNSMOS.

[0076] The process from the switching label estimation process by the switching model unit 302 to the parameter optimization process by the optimization unit 305 may be configured to be a loop process, and the optimization may be completed by repeating these processes. In either case, when the optimization is completed and the parameters are determined, the contents are reflected in the switching model unit 302, and the switching model is completed.

[0077] The above has described the learning method of the switching model unit 302 through the processing from the switching model unit 302 to the optimization unit 305. The learned switching model is used as a switching model in the switching model unit 205 of the audio signal processing device 200.

[0078] <Effects> With this configuration, the same effects as those of the first embodiment can be obtained.

[0079] <Simulation Results> Figure 11 shows the results of testing switching to improve auditory quality in the audio signal processing system of the first embodiment, demonstrating the effectiveness of the present invention. It has been shown that optimal weighting from the perspective of DNSMOS can achieve even higher auditory quality than using the emphasis signal as is. [Modifications] Furthermore, a device (terminal) for using the device, system, or method of the present invention via a network (telecommunications line) may also be provided. The "device (terminal) for use" may be equipped with functions (e.g., control functions, decoding functions, restoration functions, input / output functions, etc.) necessary to achieve the effects of implementing the device, system, or method of the present invention. [Processor, Program, Recording Medium] The functions realized by the components described herein may be implemented in circuitry or processing circuitry, including general-purpose processors, application-specific processors, integrated circuits, ASICs (Application Specific Integrated Circuits), CPUs (Central Processing Units), conventional circuits, and / or combinations thereof, programmed to realize the described functions. A processor includes transistors and other circuits and is considered to be circuitry or processing circuitry. A processor may also be a programmed processor that executes a program stored in memory.

[0080] In this specification, a circuitry, unit, or means is hardware that is programmed to realize or performs the described functions, which may be any hardware disclosed herein or any hardware known to be programmed to realize or perform the described functions.

[0081] If the hardware is a processor considered to be a type of circuitry, the circuitry, means, or unit is a combination of the hardware and software used to configure the hardware and / or processor.

[0082] The various processes described above can be implemented by loading a program that executes each step of the above method into the recording unit 2020 of the computer 2000 shown in Figure 12, and operating the control unit 2010, input unit 2030, output unit 2040, display unit 2050, etc.

[0083] The program describing the processing contents can be recorded on a computer-readable recording medium, which may be, for example, a magnetic recording device, an optical disk, a magneto-optical recording medium, a semiconductor memory, or any other suitable recording medium.

[0084] The program may be distributed by, for example, selling, transferring, lending, etc. portable recording media such as DVDs and CD-ROMs on which the program is recorded. Furthermore, the program may be stored in a storage device of a server computer, and then transferred from the server computer to other computers via a network, thereby distributing the program.

[0085] A computer that executes such a program may first temporarily store the program recorded on a portable recording medium or transferred from a server computer in its own storage device. Then, when executing a process, the computer reads the program stored on its own recording medium and executes the process in accordance with the read program. Alternatively, the computer may read the program directly from a portable recording medium and execute the process in accordance with the program. Furthermore, the computer may execute the process in accordance with the program each time a program is transferred from a server computer to the computer. Alternatively, the server computer may not transfer the program to the computer, but may instead execute the process through a so-called ASP (Application Service Provider) service, which realizes the processing function by issuing an execution instruction and obtaining the results. Furthermore, the server computer may execute the process at the terminal using a so-called SaaS (Software as a Service) service, which allows users to use part of a server computer along with the program. In this embodiment, the program includes information used for processing by an electronic computer that is equivalent to a program (such as data that is not a direct instruction to a computer but has properties that dictate computer processing).

[0086] Furthermore, in this embodiment, the device is configured by executing a predetermined program on a computer, but at least a part of the processing contents may be realized by hardware.

Claims

1. A control device that acquires an output value indicating whether or not speech enhancement should be performed on an observed signal used in a subsequent task whose performance changes depending on the ratio of an observed signal added to an enhancement signal, or the degree to which said speech enhancement should be performed, and determines an input signal to be used in the subsequent task by using the acquired output value to determine the ratio between said observed signal and the enhancement signal generated by said speech enhancement under specified conditions.

2. The control device according to claim 1, wherein the subsequent task is speaker identification, sound source direction estimation, sound source position estimation, or human listening.

3. A control method comprising: acquiring an output value indicating whether or not speech enhancement should be performed on an observed signal used in a subsequent task, the performance of which changes depending on the ratio of an observed signal added to an enhancement signal, or the degree to which said speech enhancement should be performed; and using said acquired output value to determine the ratio between said observed signal and the enhancement signal generated by said speech enhancement under specified conditions, thereby determining the input signal to be used in the subsequent task.

4. A voice processing system comprising: a control device according to claim 1; and a device for using said control device via a network.

5. A program for causing a computer to function as the control device according to claim 1.

Images