JPH08123480A - Speech recognition device - Google Patents

Abstract

PURPOSE: To provide the speech recognition device which can perform substantially real-time processing, perform speech recognition faster than a conventional example, and facilitate replacement by processing parts and is rich in extensibility. CONSTITUTION: The speech recognition device is equipped with an AD converting means 2, a feature extracting means 3, speech recognition means 4 and 5, and a control means 7 which performs control so that AD conversion, feature extraction, and speech recognition are performed in a frame synchronized state; and the control means 7 and converting means 2, the control means 7 and feature extracting means 3, and the control means 7 and speech recognition means 4 and 5 are connected by control buses 71-74 for sending control signals and status signals, the converting means 2 and feature extracting means 3 are connected by a 1st data bus 11 for transmitting speech data, and the feature extracting means 3 and speech recognizing means 4 and 5 are connected by 2nd data buses 12 and 13 for transmitting data on feature parameters.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device.

[0002]

2. Description of the Related Art FIG. 3 is a block diagram of a conventional voice recognition device. The speech recognition apparatus of FIG. 3 includes an AD converter 2a, a feature extraction unit 3a, a phoneme matching unit 4a, an LR parser 5a, a switch 8 and a voice recognition controller 7a, and the AD converter 2a, the feature extraction unit 3a and a phoneme matching unit. 4a and LR parser 5
a is connected to the voice recognition controller 7a only by control data buses 71a to 74a. Hereinafter, the AD converter 2a, the feature extraction unit 3a, and the phoneme matching unit 4a.
And the LR parser 5a are collectively referred to as a processing unit.

In the conventional voice recognition apparatus shown in FIG. 3, the speaker activates the voice recognition apparatus by pressing a switch 8 such as a foot switch at the same time as speaking. The uttered voice of the speaker is input to the microphone 1 and converted into a voice signal, and then input to the AD converter 2a. On the other hand, when the voice recognition device is activated, the voice recognition controller 7a
Outputs a control signal instructing the start of AD conversion to the AD converter 2a via the control data bus 71a. The AD converter 2a AD-converts the audio signal into audio data that is a digital signal, and then sends the audio data and a status signal indicating the end of the AD conversion process to the control data bus 71.
It is output to the voice recognition controller 7a via a. The voice recognition controller 7a outputs the voice data and a control signal instructing the start of the feature extraction processing of the voice data to the feature extraction unit 3a via the control data bus 72a.
The feature extraction unit 3a converts the input voice data into, for example, L
A PC analysis is performed to extract 34-dimensional characteristic parameters including logarithmic power, 16th-order cepstrum coefficient, Δ logarithmic power, and 16th-order Δcepstrum coefficient, and indicate time series data of the characteristic parameter and the end of the characteristic extraction processing. The status signal is output to the voice recognition controller 7a via the control data bus 72a.

The voice recognition controller 7a outputs the time-series data of the characteristic parameter and a control signal for instructing the start of the phoneme matching process to the phoneme matching unit 4a via the control data bus 73a. The phoneme collation unit 4a collates the time-series data of the extracted feature parameters with reference to information in a hidden Markov network memory (hereinafter, referred to as HM network memory) corresponding to phoneme prediction data described later, and unidentifies. A speaker model is used to calculate the likelihood for the data in the phoneme matching section, and the likelihood value is used as a phoneme matching score together with a status signal indicating the end of the phoneme matching process, and a voice recognition controller via the control data bus 73a. Output to 7a.
The voice recognition controller 7a uses the phoneme matching score and L
A control signal instructing the start of the R purging process is output to the LR parser 5a via the control data bus 74a. L
The R parser 5a refers to the LR table and processes the input phoneme matching score from left to right without backtracking. Here, the LR table is created in advance by converting a predetermined context-free grammar as is known, and stored in the LR table memory in the LR parser 5a. In the case of syntactic ambiguity, the stack is split and parsing of all candidates is processed in parallel. LR Parser 5a
Is the next phoneme predicted from the LR table, the phoneme prediction data and a status signal indicating the end of the phoneme prediction,
The voice recognition controller 7 via the control data bus 74a
output to a. The above operation is sequentially performed, and phonemes are sequentially connected to recognize continuous speech. And
The LR parser 5a outputs the voice recognition result data to an external device. In the above conventional speech recognition device, the AD converter 2
The AD conversion in a and the extraction of the feature parameter in the feature extraction unit 3a are processed for each frame, and the phoneme matching unit 4
The phoneme matching in a and the LR parsing in the LR parser 5a are processed for each phoneme.

[0005]

However, in the conventional voice recognition device, the AD converter 2a and the feature extraction unit 3a,
The feature extraction unit 3a, the phoneme matching unit 4a, the phoneme matching unit 4a, and the LR parser 5a are connected by the control data buses 71a to 74a via the voice recognition controller 7a, and each processing in each processing unit is performed by the voice recognition controller. Since it is executed after receiving the data and control signal sent from 7a, real-time processing could not be performed. Therefore, there is a problem that the voice recognition processing takes time. Further, in order to carry out various researches on speech recognition, it is necessary to replace each processing unit with another one and operate it. However, in the conventional speech recognition apparatus, each processing unit has a speech recognition controller 7
Since it is connected only to a and it is necessary to share the interface for data and control signals, there is a problem that it is not easy to replace each processing unit. Therefore, for example, it has been impossible to detect the pitch frequency from the voice signal and compare it with the voice recognition data of the voice signal.

The object of the present invention is to solve the above problems,
It is an object of the present invention to provide a speech recognition apparatus capable of substantially real-time processing, capable of recognizing speech at a higher speed than in the conventional example, and easy to replace each processing unit, and having a high expandability.

[0007]

A voice recognition apparatus according to the present invention includes a conversion means for AD-converting an input voice signal into voice data which is a digital signal and outputting the voice data, and a voice recognition device for voice recognition from the voice data. Feature extracting means for extracting feature parameters and outputting feature parameter data; voice recognition means for recognizing voice of the input voice signal based on the feature parameter data and outputting voice recognition data; The AD conversion of the converting means, the extraction of the characteristic parameter of the characteristic extracting means, and the voice recognition of the voice recognizing means are executed frame-synchronously for each frame of voice data corresponding to a voice signal of a predetermined fixed time. A voice recognition device comprising the conversion means, the feature extraction means, and a control means for controlling the voice recognition means, the control means and the conversion means During the period, a first signal for transmitting a control signal from the control means to the conversion means to start execution of AD conversion and a status signal from the conversion means to the control means indicating the end of AD conversion of one frame. And a control signal from the control means to the feature extraction means to start the execution of the feature parameter extraction processing, and the feature extraction means to perform the control. A second control bus for transmitting a status signal indicating the end of the extraction of the characteristic parameter of one frame to the means, and the control means to the voice recognition means between the control means and the voice recognition means. A control signal for instructing the start of execution of voice recognition and a status signal for indicating the end of voice recognition of one frame from the voice recognition means to the control means. 3 is connected by a control bus, and the converting means and the feature extracting means are connected by a first data bus for transmitting the voice data, and the feature extracting means and the voice recognizing means are connected to each other. It is characterized by being connected by a second data bus for transmitting the data of the characteristic parameters.

[0008]

In the speech recognition apparatus according to the first aspect of the present invention, the control means inputs the control signal for instructing the conversion means to start the AD conversion processing via the first control bus. The conversion means AD-converts the audio signal input in response to the control signal and AD-converts the audio signal for a predetermined fixed time corresponding to one frame into audio data each time While outputting to the feature extraction means via the first data bus, a status signal indicating that the AD conversion processing for one frame is completed is output to the control means via the first control bus. The control means is
Based on the status signal, the feature extraction processing of the input one-frame audio data is performed frame by frame to the feature extraction means via the second control bus for each frame so as to be frame-synchronized with the AD conversion of the conversion means. A control signal for instructing the start of is output. In response to the control signal, the feature extraction means extracts a feature parameter for each voice data of one frame and outputs the feature parameter to the voice recognition means via the second data bus, while extracting the feature of one frame. A status signal indicating that the processing is completed is output to the control means via the second control bus. Based on the status signal, the control means inputs to the voice recognition means frame by frame through the third control bus so as to be frame-synchronized with the feature extraction processing of the feature extraction means. A control signal for instructing the start of voice recognition is output based on the characteristic parameter of the frame. The voice recognition means, in response to the control signal, recognizes voice for each feature parameter of one frame and outputs voice recognition data, while indicating to the control means the end of recognition of one frame of voice data. The status signal is output via the third control bus.

As described above, in the voice recognition apparatus according to the present invention, the control means causes the conversion means, the feature extraction means, and the voice recognition means to perform frame synchronization for each frame of voice data. Through the third control bus, it is controlled to execute each process, so that each process in the conversion means, the feature extraction means, and the voice recognition means is executed in substantially real time. Further, the converting means directly transmits only the voice data to the feature extracting means via the first data bus, and the feature extracting means only sends the data of the feature parameter via the second data bus. Is directly transmitted to the voice recognition means.

[0010]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a voice recognition device according to an embodiment of the present invention. The speech recognition apparatus of FIG. 1 includes a microphone 1, an AD converter 2, a feature extraction unit 3,
A phoneme collation unit 4, an LR parser 5, a pitch detection unit 6,
The voice recognition controller 7 and the switch 8 are provided. The features of the voice recognition device are as follows. The voice recognition controller 7 performs the AD conversion of the AD converter 2, the extraction of the characteristic parameters of the characteristic extraction unit 3, the phoneme matching of the phoneme matching unit 4, and the LR purging of the LR parser 5.
The AD is executed so as to be executed in synchronization with each frame of audio data corresponding to an audio signal for a predetermined fixed time.
It controls the converter 2, the feature extraction unit 3, the phoneme matching unit 4, and the LR parser 5. Here, between the AD converter 2 and the voice recognition controller 7, a control signal for instructing the AD converter 2 to start execution of AD conversion and the AD converter 2 to the voice recognition controller are provided. 7 to transmit a status signal indicating the end of AD conversion of one frame, for example, RPC (Remo
te Procedure Calls). Further, between the feature extraction unit 3 and the voice recognition controller 7, a control signal for instructing the feature extraction unit 3 to start execution of a feature parameter extraction process and a voice signal from the feature extraction unit 3 between the feature recognition unit 3 and the voice recognition controller 7. A connection is made to the recognition controller 7 by a control bus 72 such as an RPC for transmitting a status signal indicating the end of extraction of one frame of characteristic parameters. Further, between the phoneme matching unit 4 and the voice recognition controller 7, a control signal for instructing the phoneme matching unit 4 to start the phoneme matching process from the voice recognition controller 7 and the phoneme matching unit 4 to the voice recognition controller 7. For example, a control bus 73 such as an RPC for transmitting a status signal indicating the end of the phoneme collation processing for one frame is connected. In addition, between the LR parser 5 and the voice recognition controller 7, the voice recognition controller 7 transfers the LR to the LR parser 5.
A control signal for instructing the start of purging and one frame of LR from the LR parser 5 to the voice recognition controller 7.
A control bus 74, such as an RPC, for transmitting a status signal indicating the end of the purging process is used for connection. Further, between the pitch detection unit 6 and the voice recognition controller 7, a control signal for instructing the pitch detection unit 6 to start the pitch detection process from the voice recognition controller 7 and the voice detection controller 7 from the pitch detection unit 1 A control bus 75, such as an RPC, for transmitting a status signal indicating the end of the frame pitch detection processing is used for connection. Here, each of the control signals and each of the status signals is a signal having a smaller capacity than voice data, time-series data of characteristic parameters described later, phoneme matching score described later, and phoneme prediction data described later. Further, between the AD converter 2 and the feature extraction unit 3, a data bus 11 for transmitting audio data, which is a large amount of data, capable of high-speed transmission of a large amount of data having a transmission rate of, for example, 240 kbps.
For transmitting the above audio data between the AD converter 2 and the pitch detector 6, for example, 240 kb.
Connection is made by a data bus 14 capable of high-speed transmission of a large amount of data having a transmission rate of ps. Between the feature extraction unit 3 and the phoneme matching unit 4, for example, 240 kb for transmitting time series data of feature parameters, which is a large amount of data.
Connected by a data bus 12 capable of high-speed transmission of a large amount of data having a ps transmission rate, and connected to the phoneme collation unit 4 and L
Between the R parsers 5, for example, 2.4 Mbp for transmitting the phoneme matching score and the phoneme prediction data, which are a large amount of data.
Connection is made by a data bus 13 capable of high-speed transmission of a large amount of data having a transmission rate of s. Hereinafter, the AD converter 2, the feature extraction unit 3, the phoneme matching unit 4, the LR parser 5, and the pitch detection unit 6 will be collectively referred to as a processing unit.

In the voice recognition apparatus of FIG. 1, the switch 8
Is composed of, for example, a foot petal switch or the like, and every time it is turned on, the terminal of the voice recognition controller 7 is grounded to notify the start or end of utterance. The voice recognition controller 7 includes a ROM storing a voice recognition control program, a CPU that executes voice recognition control processing according to the voice recognition control program, and a RAM used as a working area for the processing. When the voice recognition controller 7 is notified by the switch 8 that the vocalization has started, the voice recognition controller 7 causes the AD converter 2 to
A control signal for instructing the start of AD conversion processing is output via the control bus 71, and AD conversion of the audio signal for a predetermined fixed time corresponding to one frame is completed from the AD converter 2 via the control bus 71. Each time, the status signal indicating that the processing for one frame is completed is input.

Next, the voice recognition controller 7 uses the status signal from the AD converter 2 to extract the feature extraction unit 3
A control signal for instructing the start of the feature extraction processing for each frame via the control bus 72, and a control signal for instructing the pitch detection section 6 for the start of the pitch detection processing for each frame via the control bus 75. Is output. The voice recognition controller 7 receives a status signal from the feature extraction unit 3 via the control bus 72 each time the extraction of the feature parameters of one frame is finished, which indicates that the processing of one frame is finished.

The voice recognition controller 7 includes a feature extraction unit 3
A control signal for instructing the start of the phoneme matching processing for each frame is output to the phoneme matching unit 4 via the control bus 73 based on the status signal from the phoneme matching unit 4.
1 each time one frame of phoneme matching processing is completed via
A status signal indicating that the phoneme matching process of the frame is completed is input. The voice recognition controller 7 uses the LR parser 5 based on the status signal from the phoneme matching unit 4.
The control signal for instructing the start of the LR purging process is output to each of the frames via the control bus 74, and the processing of one frame is completed each time the processing of the one frame is completed from the LR parser via the control bus 74. A status signal indicating that this has been done is input.

The microphone 1 converts the input uttered voice of the speaker into a voice signal and outputs the voice signal to the AD converter 2. The AD converter 2 includes a ROM that stores an AD conversion processing program, a CPU that executes AD conversion processing according to the AD conversion program, a RAM used as a working area for processing, an input socket 21, and an output socket 22. , 23, 24. A
The D converter 2 includes the voice recognition controller 7 to the control bus 7
In response to a control signal instructing to start the AD conversion processing input via 1, the audio signal input from the microphone 1 is AD converted into audio data which is a digital signal in a time of less than 20 milliseconds, The audio data is output from the output sockets 22, 23, 24 frame by frame. Here, the voice data output from the output socket 23 is input to the input socket 31 of the feature extraction unit 3 via the data bus 11.
Audio data input to the input socket 61 of the pitch detection unit 6 via the data bus 14 is input to the input socket 61 of the pitch detection unit 6. The AD converter 2 sends a status signal indicating the end of AD conversion processing of the frame for each frame via the control bus 71 to the voice recognition controller 7.
Output to.

The pitch detection unit 6 has a ROM storing a pitch detection processing program, a CPU for executing the pitch detection processing according to the pitch detection processing program, a RAM used as a working area for the processing, and an input socket 61. , And an output socket 62. The pitch detection unit 6 responds to a control signal input from the voice recognition controller 7 via the control bus 75 and instructing the start of pitch detection processing for each frame, and outputs pitch data from the voice data input from the input socket 61. The frequency is detected in less than 20 milliseconds and the pitch frequency is output from the output socket 62. The pitch detection unit 6 is 1
For each frame, a status signal indicating the end of the pitch detection processing of the frame is output to the voice recognition controller 7 via the control bus 75.

The feature extraction unit 3 stores a ROM storing a feature extraction processing program, a CPU that executes the feature extraction processing according to the feature extraction program, a RAM used as a working area for the processing, and an input socket 3.
1 and output sockets 32, 33, 34. The feature extraction unit 3 responds to a control signal, which is input from the voice recognition controller 7 via the control bus 72, for instructing the start of feature extraction processing for each frame, from the voice data input from the input socket 31, for example. LPC analysis is performed to extract 34-dimensional characteristic parameters including logarithmic power, 16th-order cepstrum coefficient, Δlogarithmic power, and 16th-order Δcepstrum coefficient, and time-series data of the characteristic parameters are output to sockets 32, 33, 34. Output from. Here, the time-series data of the characteristic parameter output from the output socket 33 is input to the input socket 41 of the phoneme matching unit 4 via the data bus 12. The feature extraction unit 3 executes the above feature extraction processing in a time of less than 20 milliseconds. The feature extraction unit 3 sends a status signal indicating the end of the feature extraction processing of the frame for each frame to the control bus 7.
It outputs to the voice recognition controller 7 via 2.

The phoneme matching unit 4 stores a ROM storing a phoneme matching processing program, a CPU for executing the phoneme matching processing according to the phoneme matching program, a RAM used as a working area for the processing, and each state as a node. HM storing hidden Markov network data (hereinafter referred to as HM network) represented as a plurality of networks
Network memory, input socket 41, output socket 42,
44 and the input / output socket 43. With the above configuration, the phoneme matching unit 4 responds to the control signal input from the voice recognition controller 7 via the control bus 73 and instructing the start of the phoneme matching process for each frame, to generate phoneme prediction data described later. Matching is performed by referring to the corresponding HM network, the likelihood for the phoneme matching section data is calculated using the unspecified speaker model, and the likelihood value is used as the phoneme matching score in the input / output socket 43 and the output socket 42. , 44. Here, the phoneme matching score output from the input / output socket 43 is input to the input / output socket 51 of the LR parser 5 via the data bus 13. Phoneme verification unit 4
Performs the phoneme matching process described above in a time period of less than 20 milliseconds. The phoneme collation unit 4 sends a status signal indicating the end of the phoneme collation processing of the frame for each frame to the control bus 7.
3 to the voice recognition controller 7.

The LR parser 5 includes a ROM storing an LR purging processing program and a CPU for executing the LR purging processing according to the LR purging processing program.
And R used as a working area for processing
A phoneme context-dependent LR parser including an AM, an LR table memory storing an LR table created in advance by converting a predetermined context-free grammar as known, an input / output socket 51, and an output socket 52. Is.
The LR parser 5 responds to the control signal instructing the start of the LR purging process for each frame, which is input from the voice recognition controller 7 via the control bus 74, with the phoneme matching score input from the input / output socket 51. , LR table, processing from left to right without backtracking.
If there is syntactic ambiguity, split the stack and process parsing all candidates in parallel. Further, the LR parser 5 predicts the next phoneme from the LR table and outputs phoneme prediction data from the input / output socket 51. Here, the phoneme prediction data output from the input / output socket 51 is input to the input / output socket 43 of the phoneme matching unit 4 via the data bus 13. And the LR parser 5 is 1
For each frame, a status signal indicating the end of the LR purging process of the frame is output to the voice recognition controller 7 via the control bus 74. The LR parser 5 is the above-mentioned L
The R purging process and the phoneme prediction are executed together in less than 20 milliseconds. As described above, the phoneme matching unit 4 and the LR
The parser 5 successively recognizes continuous voices by sequentially connecting the phonemes and outputs final voice recognition result data from the output socket 52.

In the voice recognition device configured as described above, the operator turns on the switch 8 to
The terminal of the voice recognition controller 7 is grounded, the voice recognition controller 7 is activated, and the voice recognition processing of the voice uttered by the speaker is started. The voice recognition controller 7 inputs a control signal for instructing the AD converter 2 to start the AD conversion process. On the other hand, the voice uttered by the speaker is input to the microphone 1 and converted into a voice signal, and then continuously input to the AD converter 2. The AD converter 2 AD-inputs the audio signal
Every time the audio signal is converted and AD-converted into a voice signal for 20 milliseconds corresponding to one frame, the voice data is output to the output socket 23, the data bus 11, and the input socket 31.
Output to the feature extraction unit 3 via
A status signal indicating that the D conversion process is completed is output to the voice recognition controller 7. Based on the status signal, the voice recognition controller 7 starts processing of the input one frame of voice data to the feature extraction unit 3 for each frame so as to be frame-synchronized with the AD conversion process of the AD converter 2. Outputs a control signal to instruct. In response to the control signal, the feature extraction unit 3 extracts the above-mentioned feature parameter for each frame of voice data, and outputs the feature parameter to the phoneme collation unit 4 via the output socket 33, the data bus 12, and the input socket 41. On the other hand, a status signal indicating that the feature extraction processing for one frame is completed is output to the voice recognition controller 7.

The voice recognition controller 7 includes a feature extraction unit 3
Based on the status signal output from the above, the phoneme collation unit 4 is instructed to start the phoneme collation of the input one-frame feature parameter for each frame so as to be frame-synchronized with the feature extraction processing of the feature extraction unit 3. Output a control signal. In response to the control signal, the phoneme matching unit 4 refers to the HM network corresponding to the phoneme prediction data from the LR parser 5 for each feature parameter of one frame to perform phoneme matching, and obtains the phoneme matching score of the frame. While outputting to the LR parser 5, it outputs to the voice recognition controller 7 a status signal indicating that the phoneme matching process of the frame has ended. Based on the status signal from the phoneme matching unit 4, the voice recognition controller 7 inputs one frame of phoneme matching into the LR parser 5 for each frame so as to be frame-synchronized with the phoneme matching process of the phoneme matching unit 4. LR of score
A control signal for instructing the start of purging is output. LR
In response to the control signal from the voice recognition controller 7, the parser 5 refers to the LR table and processes the input phoneme matching score from left to right without backtracking.
If there is syntactic ambiguity, split the stack and process parsing all candidates in parallel. LR Parser 5
Outputs the phoneme prediction data to the phoneme collation unit 4 via the input / output socket 51, the data bus 13, and the input / output socket 43 by predicting the next phoneme from the LR table. After the above processing, the LR parser causes the voice recognition controller 7 to
A status signal indicating the end of the LR purging process is output. By sequentially connecting the phonemes as described above, continuous speech is recognized and the speech recognition result data is output. After the speaker finishes speaking, the operator turns on the switch 8 again to ground the terminal of the voice recognition controller 7. After this, the voice recognition controller 7
The voice recognition result data of the voice input before the end of utterance is L
After being output from the R parser 5, the status signal from the LR parser is received and the processing of the voice recognition device is terminated. As described above, in the speech recognition apparatus of this embodiment, each processing unit executes processing for each frame corresponding to a speech signal of a predetermined fixed time, and the speech recognition controller 7 is
Since each processing unit is controlled to execute the processing in frame synchronization, each processing unit executes each processing substantially in real time.

FIG. 2 is a graph showing the start time and the end time when the uttered voice is processed by the feature extraction unit 3, the phoneme matching unit 4, and the LR parser 5. Figure 2
As is clear from the above, after the voice signal of the vocalized voice is AD-converted, the feature extraction unit 3 starts the feature extraction processing. Next, after the feature extraction for one frame is completed, the phoneme matching unit 4 starts the phoneme matching process. continue,
After the phoneme matching for one frame is completed, the LR parser 5 starts the LR purging process. The feature extraction unit 3, the phoneme matching unit 4, and the LR parser 5 are performed as described above.
Is controlled to start the processing by shifting the time by 20 milliseconds, that is, by one frame, and is controlled to execute the processing in frame synchronization.

As described above, in the speech recognition apparatus of this embodiment, the AD conversion unit 2, the feature extraction unit 3, the phoneme collation unit 4, and the LR.
Each processing unit of the parser 5 should be shifted by one frame,
Since each process is executed in frame synchronization, each process in each processing unit can be executed substantially in real time, and as a result, the conventional speech recognition processing time is processed for each phoneme. It can be shorter than.

Control signals and status signals (hereinafter referred to as control signals), which are smaller in capacity than the voice data, the time-series data of characteristic parameters, the phoneme matching score, and the phoneme prediction data, are transmitted via the control buses 71 to 75. Etc.) is transmitted. On the other hand, a large amount of voice data is transmitted via the data buses 11 and 14 capable of high-speed transmission of large-capacity data, and a data bus 12 capable of high-speed transmission of large-capacity data.
A time-series data of large-capacity characteristic parameters is transmitted via the data bus 13 and high-speed transmission of large-volume data is possible.
A large-capacity phoneme matching score and phoneme prediction data are transmitted via. Therefore, the voice data, the time-series data of the characteristic parameters, the data including the phoneme matching score and the phoneme prediction data, the control signal, and the like are simultaneously transmitted using each bus provided according to the data amount. Therefore, signal transmission can be performed at high speed. As a result, the voice recognition processing time can be shortened as compared with the conventional voice recognition device that transmits the above-described data, control signals and the like via the same control data buses 71a to 74a.

Furthermore, the AD converter 2 and the feature extraction unit 3
The processing units of the voice collation unit 4, the LR parser 5, and the pitch detection unit 6 each have an input / output socket for each data, and the connection between the processing units can be easily performed using these sockets. Moreover, since the data buses 11 to 14 for transmitting each data and the control buses 71 to 75 for transmitting the control signal and the status signal are separately provided, the connection between each processing unit and the voice recognition controller is made. It is possible only by matching the interfaces between the processing units and the voice recognition controller 7, and the connections between the respective processing units are possible by matching the interfaces only between the respective processing units to be connected. That is, in each processing unit, it is not necessary to match the interface for control signals and status signals with the interface for each data, and as a result, each processing unit has each interface for each processing unit as compared with the conventional voice recognition device. Since it can be easily replaced with a new processing unit that performs different processing, it can be easily expanded in response to research on a new speech recognition device. Here, the new processing unit is, for example, a processing unit that performs voice recognition including prosodic information necessary for natural voice recognition.

In the present embodiment, speech is recognized using the phoneme collation unit 4 and the LR parser 5, but the present invention is not limited to this.
For example, a voice recognition circuit using a One Pass DP voice recognition method or the like may be used to recognize voice.

In this embodiment, one frame is set to 20 milliseconds for each processing, but the present invention is not limited to this.

Since the voice recognition apparatus of this embodiment is provided with the pitch detection section 6 and the pitch detection processing is simultaneously executed based on the same voice signal as the voice recognition, the same voice signal is used. It is possible to simultaneously compare the voice recognition data detected from the pitch frequency with the voice recognition data.

In the voice recognition apparatus of this embodiment, each of the voice recognition controller 7, the AD converter 2, the feature extraction unit 3, the phoneme collation unit 4, the LR parser 5, and the pitch detection unit 6 is C.
Although the present invention is configured to include the PU, the present invention is not limited to this, and for example, a UNIX system of one CPU may be used to perform time division multiplex pipeline processing so as to control all processing units. .

[0029]

As described above, in the voice recognition apparatus according to the present invention, the AD conversion of the conversion means, the extraction of the characteristic parameter of the characteristic extraction means, and the voice recognition of the voice recognition means are performed in a predetermined fixed time. Since the conversion means, the feature extraction means, and the control means for controlling the voice recognition means are provided so as to be executed in frame synchronization for each frame of voice data corresponding to the voice signal, the conversion is performed. The respective processes in the means, the feature extracting means, and the voice recognizing means are executed substantially in real time. In addition, the first and second data buses capable of high-speed transmission of large-capacity data are transmitted via the first, second, and third control buses, in which control signals and status signals, which are small-capacity signals, are transmitted. Through this, a large amount of voice data and time series data of a large amount of characteristic parameters are transmitted. That is, since the respective data and the control signal and the like are simultaneously transmitted through the buses provided separately according to the data amount, the control signal and the data are transmitted at high speed. From the above, according to the present invention, it is possible to provide a voice recognition device capable of performing voice recognition at a higher speed than in the conventional example.

Since the data bus for transmitting data and the control bus for transmitting the control signal and the status signal are separately configured, the interface for data and the interface for the control signal and the status signal are provided. It is not necessary to use a common processing unit, and it is possible to easily replace it with a new processing unit that performs different processing for each processing unit of the conversion unit, the feature extraction unit, and the voice recognition unit.
This makes it possible to provide a highly expandable speech recognition device that can be used for research of the speech recognition device.

[Brief description of drawings]

FIG. 1 is a block diagram of a voice recognition device according to an embodiment of the present invention.

FIG. 2 is a block diagram of the feature extraction unit 3 in the voice recognition device of FIG.
3 is a graph showing the start time and the end time of the processing when the processing is performed by the voice matching unit 4 and the LR parser 5.

FIG. 3 is a block diagram of a conventional voice recognition device.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Microphone, 2 ... AD converter, 3 ... Feature extraction part, 4 ... Phoneme collation part, 5 ... LR parser, 6 ... Pitch detection part, 7 ... Voice recognition controller, 11, 12, 13, 14 ... Data bus, 21, 31, 41, 61, ... Input sockets, 22, 23, 24, 32, 33, 34, 42, 44, 5
2, 62 ... Output sockets, 43, 51 ... Input / output sockets, 71, 72, 73, 74, 75 ... Control buses.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Tomohiko Beppu Tomohiko Beppu No.5 Mihiraya, Seiji-cho, Seika-cho, Soraku-gun, Kyoto Prefectural ATR Co., Ltd. Speech Translation Laboratory (72) Yoshinori Kosaka, Soraku Kyoto Gunma Seika-cho, Osamu Osamu, Osamu Osamu, 5 Hiratani, A-T Co., Ltd.

Claims (1)

[Claims] 1. A conversion unit that AD-converts an input voice signal into voice data which is a digital signal and outputs the voice data, and a feature parameter for voice recognition is extracted from the voice data to output feature parameter data. Feature extraction means, voice recognition means for recognizing the voice of the input voice signal based on the feature parameter data and outputting voice recognition data, AD conversion of the conversion means, and feature parameters of the feature extraction means And the voice recognition of the voice recognition means are performed in frame synchronization for each frame of voice data corresponding to a voice signal of a predetermined fixed time, the conversion means, the feature extraction means, and the voice recognition. A voice recognition device comprising: a control means for controlling the means, wherein a space between the control means and the conversion means is changed from the control means to the conversion means. From the control signal and the converting means for instructing to start execution of D conversion of one frame to the control means AD
First for transmitting a status signal indicating the end of conversion
And a control signal from the control means to the feature extraction means for instructing the feature extraction means to start execution of feature parameter extraction processing, and the feature extraction means to perform control between the control means and the feature extraction means. A second control bus for transmitting a status signal indicating the end of extraction of one frame of characteristic parameters is connected to the means, and the control means and the voice recognition means are connected from the control means to the voice recognition means. A control signal for instructing to start the execution of voice recognition and the voice recognition means to the control means 1
A third control bus for transmitting a status signal indicating the end of voice recognition of a frame, and a first data bus for transmitting the voice data between the conversion means and the feature extraction means. The voice recognition device is characterized in that the feature extraction means and the voice recognition means are connected by a second data bus for transmitting the data of the feature parameters.