JPWO2014049944A1 - Audio processing device, audio processing method, audio processing program, and noise suppression device - Google Patents

Abstract

Provided is a speech processing apparatus capable of performing highly accurate noise suppression even when knowledge of speech recognition cannot be used in advance. The speech processing apparatus includes a teacher signal transmission unit 11 that transmits a plurality of teacher signals to a speech recognition engine, which are speeches whose utterance contents have been grasped in advance and are subjected to noise suppression processing using parameters for noise suppression processing; A recognition result receiving unit 12 that receives a recognition result of speech recognition processing on a plurality of teacher signals by the speech recognition engine, and speech recognition by the speech recognition engine among parameters used for the plurality of teacher signals based on the accuracy of the recognition result And a parameter selection unit 13 that selects parameters for use in noise suppression processing performed before the processing.

Description

  The present invention relates to a voice processing device, a voice processing method, a voice processing program, and a noise suppression device, and more particularly to a voice processing device, a voice processing method, a voice processing program, and a noise suppression device used for voice recognition.

  In recent years, voice recognition technology has been put into practical use. A general apparatus that performs speech recognition performs speech recognition processing on speech input from a microphone. However, in that case, ambient noise or the like may be mixed in addition to the target speech, which is one of the major factors that lower the speech recognition rate.

  As a method for solving this problem, techniques of noise suppression and speech enhancement have been studied for many years. For example, spectral subtraction that suppresses noise by estimating a noise component in an input and subtracting it from the input has already been put into practical use. In recent years, model-based noise suppression and speech enhancement technologies have been developed, and research has been conducted to reduce speech distortion that cannot be eliminated simply by improving speech quality, and to perform conversion suitable for speech recognition. Yes.

  Patent Document 1 describes an example of model-based noise suppression and speech enhancement technology. The technique described in Patent Literature 1 obtains a noise average spectrum, a means for obtaining a temporary estimated speech from an input signal and a noise average spectrum, a standard pattern, and a correction value of the temporary estimated speech using the standard pattern. Use means. Thereby, the technique described in Patent Document 1 can provide a noise suppression system that can remove noise components with high accuracy without missing voice information.

Japanese Patent No. 4765461

  A noise suppression system that performs general model-based noise suppression performs conversion suitable for speech recognition by using a standard pattern, that is, a model having the same quality as the model used by the subsequent speech recognition engine. At this time, the noise suppression system is based on the premise that the knowledge (model) of the subsequent speech recognition engine can be used when building a model for noise suppression.

  However, for example, when the configuration of the previous stage for performing noise suppression processing and the configuration of the subsequent stage for performing speech recognition are independently constructed, knowledge of the subsequent speech recognition engine is not always available. For example, in a client-server type speech recognition system, a client side that performs noise suppression processing and a server side that performs speech recognition may be independent. In such a case, if the client side does not use speech recognition knowledge for noise suppression and there is a mismatch between the noise suppression parameter (model) and the speech recognition parameter, appropriate noise suppression cannot be performed and the voice Recognition accuracy deteriorates.

  An object of the present invention is to provide a speech processing device, a speech processing method, a speech processing program, and a noise suppression device that can perform highly accurate noise suppression even when speech recognition knowledge cannot be used in advance.

  A speech processing apparatus according to the present invention is a teacher signal transmission unit that transmits a plurality of teacher signals, which are speeches whose contents are uttered in advance and subjected to noise suppression processing using parameters for noise suppression processing, to a speech recognition engine A recognition result receiving unit that receives a recognition result of speech recognition processing for a plurality of teacher signals by the speech recognition engine, and a speech by the speech recognition engine from among parameters used for the plurality of teacher signals based on the accuracy of the recognition result And a parameter selection unit that selects a parameter to be used for the noise suppression process performed before the recognition process.

  A noise suppression device according to the present invention is a teacher signal transmission unit that transmits a plurality of teacher signals, which are voices whose utterance contents are grasped in advance and subjected to noise suppression processing using parameters for noise suppression processing, to a speech recognition engine A recognition result receiving unit that receives a recognition result of the speech recognition processing for a plurality of teacher signals by the speech recognition engine, and among the parameters used for the plurality of teacher signals based on the accuracy of the recognition result, the speech by the speech recognition engine A parameter selection unit that selects a parameter to be used for noise suppression processing performed before the recognition processing, and a noise proof processing unit that performs noise suppression processing using the selected parameter are provided.

  The noise suppression method according to the present invention transmits, to a speech recognition engine, a plurality of teacher signals that are speech whose contents of utterances are grasped in advance and are subjected to noise suppression processing using parameters for noise suppression processing. The noise suppression processing performed before the speech recognition processing by the speech recognition engine among the parameters used for the plurality of teacher signals based on the accuracy of the recognition result based on the recognition result of the speech recognition processing for the plurality of teacher signals by The method is characterized in that parameters for use in the selection are selected.

  The noise suppression program according to the present invention is a teacher that transmits, to a speech recognition engine, a plurality of teacher signals, which are sounds whose contents are uttered in advance and are subjected to noise suppression processing using parameters for noise suppression processing. Based on the signal transmission processing, the recognition result receiving processing for receiving the recognition result of the speech recognition processing for the plurality of teacher signals by the speech recognition engine, and the speech recognition among the parameters used for the plurality of teacher signals based on the accuracy of the recognition result And a parameter selection process for selecting a parameter to be used for the noise suppression process performed before the speech recognition process by the engine.

  According to the present invention, it is possible to perform highly accurate noise suppression even when knowledge of speech recognition cannot be used in advance.

It is a block diagram which shows the structure of 1st Embodiment of the audio processing apparatus by this invention. It is a flowchart which shows operation | movement of 1st Embodiment of the audio processing apparatus by this invention. 1 is a block diagram illustrating a configuration of a noise suppression device according to Embodiment 1. FIG. 3 is a flowchart illustrating the operation of the noise suppression device according to the first embodiment. It is explanatory drawing which shows the example of parameter clustering. It is a block diagram which shows the structure of 2nd Embodiment of the audio processing apparatus by this invention. It is a flowchart which shows operation | movement of 2nd Embodiment of the speech processing unit by this invention. It is a block diagram which shows the structure of the speech recognition system concerning Example 2. FIG. 10 is a flowchart illustrating an operation of the voice recognition system according to the second embodiment. It is a block diagram which shows the structure of the principal part of the speech processing unit by this invention.

  An embodiment of a voice processing device according to the present invention will be described with reference to the drawings.

Embodiment 1. FIG.
FIG. 1 is a diagram illustrating a configuration of a speech processing apparatus according to the first embodiment. The speech processing apparatus 10 shown in FIG. 1 outputs a teacher signal and receives a recognition result, and selects optimum parameters for noise suppression and speech enhancement.

  The speech processing apparatus 10 according to the present embodiment is a general-purpose computer system, and includes a CPU (Central Processing Unit), a RAM (Random Access Memory), a ROM (Read Only Memory), and a nonvolatile storage device, which are not illustrated. Prepare. In the sound processing device 10, the CPU reads an OS (Operation System) and a program stored in a RAM, a ROM, or a non-volatile storage device, and executes sound processing. Note that the voice processing apparatus 10 does not have to be a single computer system, and may be configured by a plurality of computer systems.

  As shown in FIG. 1, the speech processing apparatus 10 according to the present embodiment includes a teacher signal transmission unit 11, a recognition result reception unit 12, a parameter selection unit 13, a parameter storage unit 15, and a teacher signal storage unit 14. . The parameter storage unit 15 and the teacher signal storage unit 14 may be provided outside the speech processing apparatus 10.

  The parameter storage unit 15 stores a plurality of parameters. The parameter is data used for noise suppression processing to be described later. For example, when model-based noise suppression is used, it is a model used for a speech recognition engine. The teacher signal storage unit 14 stores a plurality of teacher signals converted by a plurality of parameters stored in the parameter storage unit 15. This conversion is processing similar to noise suppression processing performed by the noise proof processing unit 102 described later.

  The teacher signal transmission unit 11 sequentially transmits a plurality of teacher signals stored in the teacher signal storage unit 14 to the speech recognition engine 103. The teacher signal is a sound whose utterance content is grasped in advance. Also, the teacher signal transmission unit 11 notifies the recognition result reception unit 12 of information related to the transmitted teacher signal, for example, parameters used for the conversion.

  The recognition result receiving unit 12 sequentially receives the recognition results of the plurality of speech recognition processes obtained from the speech recognition engine 103 and notifies the parameter selecting unit 13 together with the parameter information notified from the teacher signal transmitting unit 11.

  Based on the accuracy of the recognition result, the parameter selection unit 13 selects a parameter to be used for noise suppression processing performed before the recognition processing by the speech recognition engine 103 from among a plurality of parameters stored in the parameter storage unit 15. select. Specifically, the parameter selection unit 13 compares a plurality of recognition results sequentially received by the recognition result receiving unit 12, and selects the recognition result with the highest recognition accuracy. And the parameter selection part 13 selects the parameter used for the teacher signal which is the recognition object of the selected recognition result. The parameter selection unit 13 selects, for example, the recognition result closest to the utterance content of the teacher signal that has been grasped in advance at the word level or the character level as the recognition result with the highest recognition accuracy. Since the distance comparison method between the word level and the character level is well known as a known technique, the description thereof is omitted here.

  Although the parameter selection unit 13 compares only the recognition accuracy, the parameter selection unit 13 may select an optimal parameter based on the combination of the recognition accuracy of the speech recognition engine 103 and the processing time. In that case, the recognition result receiving unit 12 also obtains the processing time when receiving the recognition result, and notifies the parameter selection unit 13 of it. In general, in speech recognition, the better the matching of parameters or models, the shorter the processing time. Therefore, the parameter selection unit 13 can select a parameter with higher accuracy by considering the processing time.

  Next, the operation of the speech processing apparatus according to this embodiment will be described. FIG. 2 is a flowchart showing the operation of the first embodiment of the speech processing apparatus according to the present invention.

  First, the teacher signal transmission unit 11 transmits a teacher signal (step S200). Specifically, the teacher signal transmission unit 11 sequentially transmits a plurality of teacher signals converted by a plurality of parameters and stored in the teacher signal storage unit 14 to the speech recognition engine 103. The teacher signal transmission unit 11 notifies the recognition result reception unit 12 of information related to the transmitted teacher signal.

  Next, the recognition result receiving unit 12 receives a recognition result from the speech recognition engine 103 (step S201). Specifically, the recognition result receiving unit 12 sequentially receives a plurality of recognition results from the speech recognition engine 103 and notifies the parameter selecting unit 13 together with the parameter information notified from the teacher signal transmitting unit 11.

  Next, the parameter selection unit 13 performs parameter selection (step S202). Specifically, the parameter selection unit 13 compares a plurality of recognition results notified from the recognition result receiving unit 12 and selects a recognition result with the highest recognition accuracy. And the parameter selection part 13 selects the parameter used for the teacher signal which is the recognition object of the selected recognition result. For example, it is assumed that a voice uttered “where is the nearest station” is used as a teacher signal. Then, the speech recognition engine 103 performs speech recognition on the teacher signal converted by the three types of parameters, and 3 such as “where is the nearest station”, “where is the nearest tree”, “where is Chiba station”? Suppose that a type of recognition result is output. In this case, the recognition result with the highest recognition accuracy, that is, the recognition result close to the correct answer is the second recognition result compared at the word level. Therefore, the parameter selection unit 13 selects the second recognition result. This selection result indicates that the parameter used for conversion of the teacher signal that is the object of the second recognition result is optimal for the subsequent speech recognition engine.

  Next, the parameter selection unit 13 outputs the selected parameter (step S203).

  In the present embodiment, the object used for conversion and the object to be selected are expressed as parameters, but the parameter may be, for example, a threshold value of spectral subtraction or a gain of a Wiener filter.

  Note that the teacher signal transmission unit 11 may control the transmission order when sequentially transmitting a plurality of teacher signals converted by a plurality of parameters. For example, when a plurality of parameters are clustered according to acoustic proximity, the teacher signal transmission unit 11 preferentially applies a teacher signal converted by a node (parameter) representing each cluster. Then, the teacher signal transmission unit 11 preferentially applies the parameters in the cluster associated with the representative node selected by the parameter selection unit 13 when the recognition accuracy is high. As a result, the parameter selection unit 13 can efficiently select parameters with a small amount of processing even when there are many parameters.

  FIG. 5 is an explanatory diagram illustrating an example of parameter clustering. As shown in FIG. 5, a large number of parameters (here, each symbol represents a parameter) is clustered in advance, and a representative parameter among them is a representative node. The teacher signal transmitting unit 11 first transmits four representative nodes (a, b, c, d) to perform voice recognition. Then, when the parameter selection unit 13 selects, for example, a as the parameter that provides the best recognition accuracy, the teacher signal transmission unit 11 does not expand the cluster that includes the remaining representative nodes, but within the cluster A that includes a. The teacher signal is recursively transmitted only for the parameters of. This eliminates the processing of parameters that are unlikely to be selected, thereby increasing efficiency. Note that the tree structure shown in FIG. Further, a plurality of representative nodes may be selected in order from the one with the best recognition accuracy, not just the parameter with the best recognition accuracy.

  As described above, the speech processing apparatus 10 according to the present embodiment can perform highly accurate noise suppression and speech enhancement without knowing in advance the knowledge of the subsequent speech recognition engine when performing noise suppression processing. In particular, when using a model-based noise suppression method, if there is a mismatch between the model at the time of noise suppression and the model at the time of speech recognition, it will cause a large deterioration in accuracy, so the effect of improving accuracy by reducing this mismatch Is expensive.

  In addition, the speech processing apparatus 10 according to the present embodiment reduces the amount of processing by efficiently applying a highly accurate one using a tree structure or the like when there are a large number of parameters for creating a teacher signal. Can do.

<Example 1>
Hereinafter, examples of the speech processing apparatus according to the present embodiment will be described. FIG. 3 is a block diagram of the configuration of the noise suppression device according to the first embodiment. The noise suppression device 100 illustrated in FIG. 3 includes a voice processing device 10b and a noise proof processing unit 102. The noise suppression apparatus 100 is connected to the input unit 101 and the speech recognition engine 103. The noise suppression apparatus 100 operates as a noise suppression engine, and outputs a voice suitable for voice recognition by performing a noise suppression process on the input voice.

  The noise suppression apparatus 100 uses a general-purpose computer system, and includes a CPU, a RAM, a ROM, and a non-volatile storage device as components not shown. In the noise suppression device 100, the CPU reads the OS and the noise suppression program stored in the RAM, ROM, or nonvolatile storage device, and executes noise suppression processing. Thereby, the noise suppression apparatus 100 can make the input voice suitable for voice recognition. Note that the noise suppression apparatus 100 does not have to be a single computer system, and may be configured by a plurality of computer systems.

  The input unit 101 inputs voice and outputs it to the noise proof processing unit 102. The input unit 101 is a microphone, for example.

  The speech processing apparatus 10b transmits a teacher signal to the speech recognition engine, receives a recognition result, and notifies the noise immunity processing unit 102 of parameters suitable for noise suppression processing before the input unit 101 inputs speech. . The voice processing device 10b has the same function as the voice processing device 10 shown in FIG.

  The noise proof processing unit 102 performs noise suppression processing on the voice input from the input unit 101 using the parameters notified from the voice processing device 10 b and outputs the result to the voice recognition engine 103. The noise proof processing unit 102 performs, for example, model-based noise suppression as the noise suppression processing.

  Next, the operation of the noise suppression apparatus 100 of the present embodiment will be described. FIG. 4 is a flowchart of the operation of the noise suppression apparatus according to the first embodiment. Since the voice processing device 10b has the same function as the voice processing device 10 of FIG. 1, detailed description of the same operations as those of the voice processing device 10 shown in the flowchart of FIG. 2 is omitted.

  First, the teacher signal transmission unit 11 transmits a teacher signal (step S400). Specifically, the teacher signal transmission unit 11 sequentially transmits a plurality of teacher signals to the speech recognition engine 103.

  Next, the recognition result receiving unit 12 receives the recognition result (step S401). Specifically, the recognition result receiving unit 12 receives a plurality of recognition results sequentially obtained from the speech recognition engine 103.

  Next, the parameter selection unit 13 performs parameter selection (step S402). Specifically, the parameter selection unit 13 selects the one with the highest recognition accuracy among the plurality of recognition results sequentially received from the recognition result receiving unit 12. And the parameter selection part 13 selects the parameter used for the teacher signal which is the recognition object of the selected recognition result.

  Next, the parameter selection unit 13 outputs a parameter (step S403). Specifically, the parameter selection unit 13 notifies the noise proof processing unit 102 of the selected parameter.

  Next, the input unit 101 inputs sound (step S404).

  Next, the noise proof processing unit 102 performs noise suppression processing (step S405). Specifically, the noise proof processing unit 102 performs noise suppression processing on the voice input from the input unit 101 using the parameters notified from the voice processing device 10b, and converts the input voice after noise suppression into the voice recognition engine 103. Output to.

  In the present embodiment, the noise suppression apparatus 100 performs parameter selection by transmission of a teacher signal and noise suppression processing for input speech once each. However, once parameter selection is performed, under the same conditions, You may continue to use the selected parameters.

Embodiment 2. FIG.
The speech processing apparatus of the present embodiment described below is for efficiently selecting parameters even when parameters are changed following an environment that changes from moment to moment.

  FIG. 6 is a block diagram showing the configuration of the second embodiment of the speech processing apparatus according to the present invention. The speech processing apparatus 10c shown in FIG. 6 outputs a teacher signal and receives a recognition result, and selects an optimal model parameter for noise suppression and speech enhancement.

  The speech processing apparatus 10c of the present embodiment uses a general-purpose computer system, and includes a CPU, a RAM, a ROM, and a nonvolatile storage device as components not shown. In the sound processing device 10c, the CPU reads the OS and the sound processing program stored in the RAM, ROM, or nonvolatile storage device, and executes sound processing. Thereby, appropriate noise suppression can be performed efficiently. The voice processing device 10c does not have to be a single computer system, and may be configured using a plurality of computer systems.

  As illustrated in FIG. 6, the speech processing apparatus 10 c according to the present embodiment includes a teacher signal transmission unit 11, a recognition result reception unit 12, a parameter selection unit 13, and a parameter recording unit 16. The voice processing device 10c further includes a parameter storage unit 15 that stores a plurality of parameters, and a teacher signal storage unit 14 that stores a plurality of teacher signals converted by the plurality of parameters. Hereinafter, only the difference of the audio processing device 10c from the audio processing device 10 shown in FIG. 1 will be described.

  The parameter recording unit 16 records the parameter notified from the parameter selection unit 13 and notifies the teacher signal transmission unit 11 of the parameter recorded in the next voice processing. At this time, the number of parameters to be recorded is not limited to one, and history information of a plurality of parameters selected in the sound processing so far may be recorded. Further, the parameter recording unit 16 may notify the teacher signal transmission unit 11 of a plurality of parameters.

  The teacher signal transmission unit 11 transmits a plurality of teacher signals converted by a plurality of parameters to the speech recognition engine 103. The teacher signal transmission unit 11 preferentially uses the teacher signal converted by the parameters recorded in the parameter recording unit 16.

  The parameter selection unit 13 compares a plurality of recognition results sequentially received from the recognition result reception unit 12, selects the one with the highest recognition accuracy, and sets the parameters used for the teacher signal that is the recognition target of the selected recognition result. The selected information is output to the noise proof processing unit 102 and simultaneously notified to the parameter recording unit 16.

  Next, the operation of the speech processing apparatus according to this embodiment will be described. FIG. 7 is a flowchart showing the operation of the speech processing apparatus according to this embodiment.

  First, the teacher signal transmission unit 11 transmits a teacher signal (step S700). Specifically, the teacher signal transmission unit 11 of the speech processing device 10 sequentially transmits a plurality of teacher signals converted by a plurality of parameters to the speech recognition engine 103. The teacher signal transmission unit 11 notifies the recognition result reception unit 12 of information related to the transmitted teacher signal. At this time, the teacher signal transmission unit 11 preferentially uses a plurality of parameters notified from the parameter recording unit 16, that is, teacher signals converted by parameters selected in the past. The teacher signal transmission unit 11 preferentially uses, for example, the teacher signal converted by the parameter selected immediately before among the parameters selected in the past recorded in the parameter recording unit 16.

  Next, the recognition result receiving unit 12 receives the recognition result (step S701). Specifically, the recognition result receiving unit 12 of the speech processing apparatus 10 sequentially receives a plurality of recognition results obtained from the speech recognition engine 103 and selects parameters together with parameter information notified from the teacher signal transmitting unit 11. Notification to the unit 13.

  Next, the parameter selection unit 13 performs parameter selection (step S702). Specifically, the parameter selection unit 13 of the speech processing device 10 compares a plurality of recognition results sequentially received by the recognition result receiving unit 12 and selects a recognition result with the highest recognition accuracy. Then, the parameter selection unit 13 selects a parameter used for the teacher signal that is a recognition target of the selected recognition result.

  Next, the parameter selection unit 13 outputs a parameter (step S703). Specifically, the parameter selection unit 13 of the voice processing device 10 notifies the parameter recording unit 16 of the selected parameter.

  Next, the parameter recording unit 16 records the parameters (step S704). Specifically, the parameter recording unit 16 of the voice processing device 10 records the parameter notified from the parameter selecting unit 13 and notifies the teacher signal transmitting unit 11 of the parameter during the next voice processing.

  Further, the speech processing apparatus 10c according to the present embodiment may control the order of parameters for performing speech recognition by hierarchizing parameters as shown in FIG. In this case, for example, the parameter recorded in the parameter recording unit 16 is set as the representative node. Then, the teacher signal transmission unit 11 preferentially transmits the teacher signal noise-processed by the parameter and the teacher signal noise-processed by the parameter included in the lower cluster of the parameter to the speech recognition engine 103.

  As described above, the speech processing apparatus 10c according to the present embodiment can reduce the processing amount by efficiently selecting the parameters following the information up to immediately before.

<Example 2>
Hereinafter, examples of the speech processing apparatus according to the present embodiment will be described. FIG. 8 is a diagram illustrating the configuration of the speech recognition system according to the second embodiment. The speech recognition system 800 in FIG. 8 includes a speech processing device 10d, a noise proof processing unit 102, and a speech recognition engine 103. The voice recognition system 800 is connected to the input unit 101 and the output unit 801.

  The speech recognition system 800 uses a general-purpose computer system, and includes a CPU, a RAM, a ROM, and a nonvolatile storage device as components not shown. In the speech recognition system 800, the CPU reads the OS and speech recognition program stored in the RAM, ROM, or nonvolatile storage device, and executes speech recognition processing. Thereby, the speech recognition system 800 can realize speech recognition that operates robustly even under noise. Note that the speech recognition system 800 does not have to be a single computer system, and may be configured using a plurality of computer systems. The input unit 101 accepts input audio and inputs it to the noise proof processing unit 102. The input unit 101 is, for example, a microphone.

  The speech processing apparatus 10d transmits a teacher signal to the speech recognition engine 103 before the input unit 101 accepts the input speech, receives the recognition result, and notifies the noise proof processing unit 102 of a parameter suitable for noise suppression processing. . The voice processing device 10d has a function similar to that of the voice processing device 10c in FIG.

  The noise proof processing unit 102 performs noise proof processing on the voice input from the input unit 101 using the parameter notified from the voice processing device 10 d and outputs it to the voice recognition engine 103. The speech recognition engine 103 performs speech recognition processing on the input speech after noise suppression input from the noise proof processing unit 102 and notifies the output unit 801 of the speech recognition result.

  The output unit 801 outputs a recognition result. For example, the output unit 801 may be a display that displays text on the screen, or may be a speaker that incorporates a speech synthesizer for outputting the recognition result as speech.

  Next, the overall flow of the speech recognition processing according to the second embodiment will be described. FIG. 9 is a flowchart illustrating the operation of the speech recognition system according to the second embodiment. Here, since the speech processing apparatus 10d of the speech recognition system 800 has the same function as the speech processing apparatus 10c of FIG. 6, detailed description of the same operations as those in the flowchart of FIG. 7 is omitted.

  First, the teacher signal transmission unit 11 transmits a teacher signal (step S900). Specifically, the teacher signal transmission unit 11 sequentially transmits a plurality of teacher signals to the speech recognition engine 103. At this time, the teacher signal converted by a plurality of past parameters notified from the parameter recording unit 16 is preferentially used.

  Next, the recognition result receiving unit 12 receives the recognition result (step S901). Specifically, the recognition result receiving unit 12 receives a plurality of recognition results sequentially obtained from the speech recognition engine 103.

  Next, the parameter selection unit 13 performs parameter selection (step S902). Specifically, the parameter selection unit 13 compares a plurality of recognition results sequentially received by the recognition result receiving unit 12, and selects the recognition result with the highest recognition accuracy. And the parameter selection part 13 selects the parameter used for the teacher signal which is the recognition object of the recognition result.

  Next, the parameter selection unit 13 outputs a parameter (step S903). Specifically, the parameter selection unit 13 of the voice processing device 10d notifies the selected parameter to the noise proof processing unit 102 and the parameter recording unit 16 of the voice processing device 10d.

  Next, the parameter recording unit 16 records the parameters (step S904). Specifically, the parameter recording unit 16 of the voice processing device 10d records the notified parameter and makes it available for the next voice processing.

  Next, the input unit 101 inputs sound (step S905).

  Next, the noise proof processing unit 102 performs noise suppression processing (step S906). Specifically, the noise proof processing unit 102 performs noise suppression processing on the voice input from the input unit 101 using parameters notified from the voice processing device 10d, and the input voice after noise suppression is input to the voice recognition engine 103. input.

  Next, the speech recognition engine 103 recognizes speech (step S907). Specifically, the speech recognition engine 103 performs speech recognition processing on the speech after noise suppression input from the noise proof processing unit 102 and notifies the output unit 801 of the recognition result.

  Next, the speech recognition engine 103 outputs a recognition result (step S908). Specifically, the output unit 801 displays the recognition result notified from the speech recognition engine 103 on, for example, a display.

  As described above, since the speech recognition system according to the second embodiment can select parameters efficiently by using information on parameters selected in the past, the amount of processing can be increased even if the noise suppression processing is repeatedly performed. It becomes possible to suppress.

  FIG. 10 is a block diagram showing the configuration of the main part of the speech processing apparatus according to the present invention. As shown in FIG. 10, the speech processing apparatus according to the present invention has, as a main configuration, a plurality of teachers that are speeches whose utterance contents have been grasped in advance and are subjected to noise suppression processing using parameters for noise suppression processing. Based on the accuracy of the recognition results, a teacher signal transmitting unit 11 that transmits signals to the speech recognition engine, a recognition result receiving unit 12 that receives recognition results of speech recognition processing for a plurality of teacher signals by the speech recognition engine, and a plurality of recognition results. A parameter selection unit 13 is provided for selecting a parameter to be used for noise suppression processing performed before speech recognition processing by the speech recognition engine among parameters used for the teacher signal.

  Moreover, in said embodiment, the audio processing apparatus and noise suppression apparatus which are shown to the following (1)-(6) are also disclosed.

(1) The noise suppression process is a model-based noise suppression process, and the parameter is a speech processing apparatus that is a model for the noise suppression process. According to such a speech processing apparatus, the accuracy of noise suppression can be greatly improved. When using a model-based noise suppression method, if there is a mismatch between the model at the time of noise suppression and the model at the time of speech recognition, it will cause a large deterioration in accuracy, so the effect of improving accuracy by reducing this mismatch is high. Because.

(2) In the speech processing apparatus, the recognition result receiving unit acquires the processing time of speech recognition processing for a plurality of teacher signals performed by the speech recognition engine, and the parameter selection unit determines the accuracy of the recognition result and the processing time. Based on the parameters used for the plurality of teacher signals, a parameter for use in noise suppression processing performed before speech recognition processing by the speech recognition engine may be selected. According to such a voice processing device, it is possible to select a parameter with higher accuracy by taking the processing time into consideration. This is because, generally speaking, the better the parameter or model matching in speech recognition, the shorter the processing time.

(3) In the speech processing apparatus, a plurality of parameters are expressed as a tree structure classified into a cluster of a plurality of hierarchies according to acoustic proximity, a plurality of parameters are expressed as representative nodes of the cluster, and a teacher signal is transmitted. The unit preferentially transmits a teacher signal using a parameter expressed as a representative node of the tree structure to the speech recognition engine, and selects a parameter used for the teacher signal having the highest recognition accuracy among the teacher signals. The teacher signal using the parameter belonging to the cluster including the parameter may be configured to be transmitted to the speech recognition engine with priority next. According to such a speech processing apparatus, even when there are a large number of parameters, it is possible to efficiently select parameters with a small processing amount.

(4) The speech processing apparatus includes a parameter recording unit (for example, parameter recording unit 16) that records a parameter selected in the past, and the teacher signal transmission unit uses a parameter recorded in the parameter recording unit. The signal may be configured to be transmitted to the speech recognition engine with priority. According to such a speech processing device, the amount of processing can be reduced by efficiently selecting a parameter following the information up to immediately before.

(5) The speech processing apparatus includes a parameter recording unit that records a parameter selected in the past, and the plurality of parameters are expressed as a tree structure classified into a cluster of a plurality of layers according to acoustic proximity. Are represented as representative nodes of the cluster, and the teacher signal transmission unit selects one representative node based on the parameters recorded in the parameter recording unit, and belongs to the cluster including the parameter represented as the representative node. The teacher signal using the parameter may be preferentially transmitted to the speech recognition engine. According to such an audio processing device, even when there are a large number of parameters, it is possible to efficiently select parameters with a small amount of processing, and it is possible to efficiently select parameters following the information up to immediately before. , Can reduce the processing amount.

(6) A teacher signal transmission unit that transmits a plurality of teacher signals, which are voices whose utterance contents have been grasped in advance and have been subjected to noise suppression processing using parameters for noise suppression processing, to the speech recognition engine, and the speech recognition engine A recognition result receiving unit for receiving recognition results of speech recognition processing for a plurality of teacher signals by the voice, and parameters used for the plurality of teacher signals based on the accuracy of the recognition results before the speech recognition processing by the speech recognition engine Noise suppression apparatus comprising: a parameter selection unit that selects parameters for use in noise suppression processing to be performed; and a noise resistance processing unit (for example, noise resistance processing unit 102) that performs noise suppression processing using the selected parameters .

  This application claims the priority on the basis of Japanese application Japanese Patent Application No. 2012-213864 for which it applied on September 27, 2012, and takes in those the indications of all here.

  While the present invention has been described with reference to the embodiments (and examples), the present invention is not limited to the above embodiments (and examples). Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

Industrial applicability

  INDUSTRIAL APPLICABILITY According to the present invention, the speech recognition system can be applied to an application such as a noise-resistant processing system in order to execute it accurately under noise.

10, 10b, 10c, 10d Speech processing device 11 Teacher signal transmission unit 12 Recognition result reception unit 13 Parameter selection unit 14 Teacher signal storage unit 15 Parameter storage unit 16 Parameter recording unit 100 Noise suppression device 101 Input unit 102 Noise resistance processing unit 103 Speech recognition engine 800 Speech recognition system 801 Output unit

Claims (9)

A teacher signal transmitting unit that transmits a plurality of teacher signals that are noises that have been preliminarily grasped and are subjected to noise suppression processing using parameters for noise suppression processing;
A recognition result receiving unit that receives a recognition result of voice recognition processing for the plurality of teacher signals by the voice recognition engine;
A parameter selection unit for selecting a parameter to be used for noise suppression processing performed before speech recognition processing by the speech recognition engine from parameters used for the plurality of teacher signals based on the accuracy of the recognition result; An audio processing device comprising: Noise suppression processing is model-based noise suppression processing,
The speech processing apparatus according to claim 1, wherein the parameter is a model for noise suppression processing. The recognition result receiver
Acquire the processing time of speech recognition processing for multiple teacher signals performed by the speech recognition engine,
The parameter selector
Based on the accuracy of the recognition result and the processing time, a parameter to be used for noise suppression processing performed before speech recognition processing by the speech recognition engine is selected from the parameters used for the plurality of teacher signals. The speech processing apparatus according to claim 1 or 2. The plurality of parameters are represented as a tree structure classified into a cluster of a plurality of layers according to acoustic proximity, and the plurality of parameters are represented as representative nodes of the cluster,
The teacher signal transmitter
The teacher signal using the parameter expressed as the representative node of the tree structure is transmitted to the speech recognition engine with priority, and the parameter used for the teacher signal with the highest recognition accuracy is selected from among the teacher signals, The speech processing apparatus according to any one of claims 1 to 3, wherein a teacher signal in which a parameter belonging to a cluster including a parameter is used is transmitted with priority to the speech recognition engine. A parameter recording unit for recording parameters selected in the past is provided.
The teacher signal transmitter
The speech processing apparatus according to any one of claims 1 to 3, wherein a teacher signal using the parameter recorded in the parameter recording unit is preferentially transmitted to a speech recognition engine. A parameter recording unit for recording parameters selected in the past is provided.
The plurality of parameters are represented as a tree structure classified into a cluster of a plurality of layers according to acoustic proximity, and the plurality of parameters are represented as representative nodes of the cluster,
The teacher signal transmitter
One representative node is selected based on the parameter recorded in the parameter recording unit, and a teacher signal using a parameter belonging to a cluster including the parameter expressed as the representative node is preferentially transmitted to the speech recognition engine. The speech processing apparatus according to any one of claims 1 to 3. A teacher signal transmitting unit that transmits a plurality of teacher signals that are noises that have been preliminarily grasped and are subjected to noise suppression processing using parameters for noise suppression processing;
A recognition result receiving unit that receives a recognition result of voice recognition processing for the plurality of teacher signals by the voice recognition engine;
A parameter selection unit for selecting a parameter to be used for noise suppression processing performed before speech recognition processing by the speech recognition engine from parameters used for the plurality of teacher signals based on the accuracy of the recognition result; ,
A noise suppression apparatus, comprising: a noise proof processing unit that performs noise suppression processing using the selected parameter. Sending a plurality of teacher signals, which are speeches whose utterance contents have been grasped in advance and subjected to noise suppression processing using parameters for noise suppression processing, to the speech recognition engine,
Receiving recognition results of voice recognition processing for the plurality of teacher signals by the voice recognition engine;
Based on the accuracy of the recognition result, a parameter used for noise suppression processing performed before speech recognition processing by the speech recognition engine is selected from parameters used for the plurality of teacher signals. Voice processing method. On the computer,
A teacher signal transmission process for transmitting a plurality of teacher signals, which are voices whose utterance contents have been grasped in advance and subjected to noise suppression processing using parameters for noise suppression processing, to a speech recognition engine;
A recognition result receiving process for receiving a recognition result of a voice recognition process for the plurality of teacher signals by the voice recognition engine;
A parameter selection process for selecting a parameter to be used for a noise suppression process performed before a voice recognition process by the voice recognition engine from parameters used for the plurality of teacher signals based on the accuracy of the recognition result; Voice processing program to execute.

Images