JP7498231B2 - Information processing device, voice recognition support method, and voice recognition support program - Google Patents

Description

The present invention relates to an information processing device, a voice recognition assistance method, and a voice recognition assistance program.

Conventionally, there are known voice recognition systems in which a voice recognition engine is shared by multiple client terminals. Examples of voice recognition systems include narrow-area voice recognition systems that connect multiple client terminals and a voice recognition server via a home network, and cloud-based voice recognition systems that have a voice recognition engine installed on a cloud server and perform voice recognition via a network such as the Internet.

In such a voice recognition system, a technology has been proposed in which a client terminal is equipped with a voice activity detection (VAD) function that detects speech intervals from input voice data, and only the voice data of the detected speech intervals is sent to the server (see, for example, Patent Document 1).

Patent No. 4425055

Conventional voice activity detection detects whether or not there is sound, so if the output of the voice activity detection is directly input to a speech recognition engine, there is a risk that the speech recognition will not be accurate due to noise or interruptions in speech.

For example, as shown in FIG. 15, when noise occurs, speech and noise cannot be distinguished, so the noise speech signal is detected as a speech section and input to the speech recognition engine. In this case, unnecessary speech recognition will occur. Also, when a user utters "Returning home for summer vacation," if there is a break in the speech between "returning home for summer vacation" and "returning home," the speech data "summer vacation" and "returning home" will be detected separately and input to the speech recognition engine. In this case, it is not possible to capture it as one continuous speech, and the context cannot be understood, so there is a risk of erroneous speech recognition, such as "parasites during summer vacation."

The present invention has been made in consideration of the above circumstances, and aims to provide an information processing device, a voice recognition assistance method, and a voice recognition assistance program that can improve the accuracy of voice recognition.

The first aspect of the present invention is an information processing device that includes a voice section detection unit that detects a voice section where sound is generated from voice data, a smoothing processing unit that smoothes an output signal of the voice section detection unit, a speech determination unit that determines a speech state when a state in which the smoothed signal output from the smoothing processing unit is equal to or greater than a speech start threshold is maintained for a preset speech start determination time in a stopped state, a stop determination unit that determines a stop state when a state in which the smoothed signal is equal to or less than a speech stop threshold is maintained for a preset speech stop determination time in the stopped state, and a voice data management unit that determines a speech section based on the speech start time at which the speech state is determined and the speech stop time at which the stopped state is determined.

A second aspect of the present invention is an information processing device that includes a voice section detection unit that detects voice sections where sound is being generated from voice data, a speech in progress determination unit that determines a speech state when a state in which an output signal of the voice section detection unit is above a speech start threshold is maintained for a predetermined speech start determination time, a stop state determination unit that determines a stop state when a state in which the output signal is below a speech stop threshold is maintained for a predetermined speech stop determination time, and a voice data management unit that determines a speech section based on the speech start time determined to be the speech in progress state and the speech stop time determined to be the stop state, wherein the voice data management unit determines the start time of the speech section based on the speech start time and a predetermined speech start margin time, and determines the end time of the speech section based on the speech stop time and the predetermined speech stop margin time .

The third aspect of the present invention is a speech recognition support method in which a computer executes a speech interval detection step of detecting a speech interval in which sound is generated from speech data, a smoothing process step of smoothing an output signal of the speech interval detection step, a speech determination step of determining an utterance in progress state when a state in which the smoothed signal output from the smoothing process is equal to or greater than a speech start threshold is maintained for a predetermined speech start determination time in a stopped state, a stop determination step of determining an utterance in progress state when a state in which the smoothed signal is equal to or less than a speech stop threshold is maintained for a predetermined speech stop determination time in the stopped state, and a speech data management step of determining an utterance interval based on the speech start time at which the utterance in progress state is determined and the speech stop time at which the utterance in progress state is determined.

A fourth aspect of the present invention is a speech recognition assistance program for causing a computer to execute the above-mentioned speech recognition assistance method.

The present invention has the effect of improving the accuracy of voice recognition.

1 is a system configuration diagram illustrating an overall configuration of a voice recognition system according to a first embodiment of the present invention. FIG. 2 is a diagram illustrating an example of a hardware configuration of a client terminal according to the first embodiment of the present invention. FIG. 2 is a functional configuration diagram showing an example of a voice recognition assistance function of the client terminal according to the first embodiment of the present invention. 4 is a diagram showing an example of a VAD signal output from a voice detection unit according to the first embodiment of the present invention; FIG. 5 is a diagram showing an example of a smoothed signal output from a smoothing processing unit according to the first embodiment of the present invention; FIG. 5 is a diagram for explaining the operation of a state management unit according to the first embodiment of the present invention; FIG. FIG. 2 is a speech state transition diagram in a state management unit according to the first embodiment of the present invention. 5A to 5C are diagrams for explaining the operation of a voice data management unit according to the first embodiment of the present invention. FIG. 2 is a diagram illustrating an example of a ring buffer according to the first embodiment of the present invention. FIG. 2 is a functional configuration diagram showing an example of a voice recognition function of the server according to the first embodiment of the present invention. FIG. 11 is a functional configuration diagram showing an example of a voice recognition assistance function of a client terminal according to a second embodiment of the present invention. 13 is a diagram for explaining the operation of a voice recognition support unit according to the second embodiment of the present invention. FIG. FIG. 13 is a functional configuration diagram showing an example of a voice recognition assistance function of a client terminal according to a third embodiment of the present invention. FIG. 13 is a diagram for explaining the operation of a voice recognition support unit according to the third embodiment of the present invention. FIG. 1 is a diagram for explaining a conventional method for determining a speech section.

First Embodiment
Fig. 1 is a system configuration diagram showing an outline of the overall configuration of a voice recognition system according to a first embodiment of the present invention. As shown in Fig. 1, the voice recognition system 1 includes a plurality of client terminals (information processing devices) 10 and a server 50. The plurality of client terminals 10 and the server 50 are configured to be connectable via a communication network 2. Examples of the communication network 2 include the Internet, Bluetooth (registered trademark), Wi-Fi, a wireless LAN, and a wired LAN (Local Area Network). Although two client terminals 10 are shown in Fig. 1, the number of client terminals 10 connected to the server 50 is not particularly limited.

The client terminal 10 is, for example, an information processing device such as a notebook PC, a desktop PC, a tablet terminal, or a smartphone.
Fig. 2 is a diagram showing an example of a schematic configuration of hardware of the client terminal 10 according to this embodiment. As shown in Fig. 2, the client terminal 10 includes, for example, a CPU (Central Processing Unit) 11, a main memory 12, a secondary storage device 13, an external interface 14, a communication interface 15, an input device 16, a display 17, a microphone 18, and a speaker 19. These components are directly or indirectly connected via a bus.

The CPU 11 controls the entire client terminal 10 using, for example, an OS (Operating System) stored in a secondary storage device 13 connected via a bus, and executes various processes, as described below, by executing various programs stored in the secondary storage device 13. One or more CPUs 11 may be provided, and they may work together to realize processes.

The main memory 12 is composed of writable memory such as cache memory and RAM (Random Access Memory), and is used as a working area for reading the execution program of the CPU 11, writing processing data by the execution program, etc.

The secondary storage device 13 is a non-transitory computer readable storage medium. The secondary storage device 13 is, for example, a magnetic disk, a magneto-optical disk, a CD-ROM, a DVD-ROM, a semiconductor memory, etc. Examples of the secondary storage device 13 include a ROM (Read Only Memory), a HDD (Hard Disk Drive), a SSD (Solid State Drive), a flash memory, etc. The secondary storage device 13 stores, for example, an OS for controlling the entire client terminal 10, such as Windows (registered trademark), iOS (registered trademark), Android (registered trademark), a BIOS (Basic Input/Output System), various device drivers for operating the hardware of peripheral devices, various application software, and various data and files. The secondary storage device 13 also stores programs for implementing various processes and various data required for implementing various processes. A plurality of secondary storage devices 13 may be provided, and the above-mentioned programs and data may be divided and stored in each secondary storage device 13.

The external interface 14 is an interface for connecting to an external device. Examples of external devices include an external monitor, a USB memory, an external HDD, and an external camera. Note that in the example shown in FIG. 2, only one external interface 14 is shown, but multiple external interfaces 14 may be provided. The external interfaces 14 each have appropriate input/output terminals and interfaces depending on the device to be connected, for example.

The communication interface 15 functions as an interface for connecting to a network to communicate with other devices and to send and receive information. For example, the communication interface 15 communicates with other devices by wire or wirelessly. Examples of wireless communication include communication through lines such as Bluetooth (registered trademark), Wi-Fi, mobile communication systems (3G, 4G, 5G, 6G, LTE, etc.), and wireless LAN. An example of wired communication is communication through lines such as a wired LAN (Local Area Network).

The input device 16 is a user interface, such as a keyboard, mouse, or touchpad, that allows the user to give instructions to the client terminal 10.

The display 17 is, for example, a liquid crystal display, an organic electroluminescence (EL) display, etc. Furthermore, the display 17 may be a touch panel display on which a touch panel is superimposed.
The microphone 18 converts the sound into an analog signal and outputs it as an audio signal.
The speaker outputs the audio signal as sound.

The server 50 is an information processing device and includes a CPU, main memory, secondary storage device, etc. Note that the main hardware configuration is almost the same as that of the client terminal 10 described above and is publicly known, so a detailed description will be omitted.

Figure 3 is a functional configuration diagram showing an example of a voice recognition support function possessed by the client terminal 10 according to this embodiment.

As an example, a series of processes for realizing the various functions described below is stored in the secondary storage device 13 in the form of a program (e.g., a voice recognition assistance program), and the CPU 11 reads this program into the main memory 12 and executes information processing and arithmetic processing to realize the various functions. Note that the program may be pre-installed in the secondary storage device 13, provided in a state stored in another computer-readable storage medium, or distributed via wired or wireless communication means. Examples of computer-readable storage media include magnetic disks, magneto-optical disks, CD-ROMs, DVD-ROMs, and semiconductor memories.

As shown in FIG. 3, the client terminal 10 includes a voice recognition support unit 20. The voice recognition support unit 20 includes, for example, a voice input unit 21, a voice segment detection unit 22, and a correction filter unit 23.

The audio input unit 21 divides the audio signal input from the microphone 18 into certain intervals (e.g., several tens of ms) to generate audio data for each frame, and outputs the data together with time information. Note that, although the present embodiment illustrates a case in which an audio signal is input from the microphone 18, the audio signal is not limited to this. The audio signal may be, for example, an audio signal received from an external device, or an audio signal read from a recording medium, etc.

The voice data is output to the voice activity detection unit 22 and the voice data management unit 33 of the correction filter unit 23 .
The voice activity detection unit (VAD) 22 detects a voice activity in which sound is generated from the voice data output from the voice input unit 21. For example, the voice activity detection unit 22 determines the presence or absence of voice for each frame of voice data, and outputs an output signal of "1" when voice is present and "0" when voice is not present. Note that various methods have been proposed for voice activity detection (VAD), and any of these known techniques may be appropriately adopted.

The output signal of the voice activity detection unit 22 is associated with time information and output to the correction filter unit 23. Hereinafter, the signal output from the voice activity detection unit 22 is referred to as the "VAD signal." Figure 4 shows an example of a VAD signal. In Figure 4, the horizontal axis represents time and the vertical axis represents output. The output is "1" for sections where voice is detected, and "0" for sections where voice is not detected.

The correction filter unit 23 includes, for example, a smoothing processing unit 31 , a state management unit 32 , and a voice data management unit 33 .
The smoothing processing unit 31 smoothes, for example, the VAD signal from the voice activity detection unit 22. One example of the smoothing is moving average (for example, simple moving average, weighted moving average, exponential moving average, etc.). In this embodiment, a case where a simple moving average is used as the smoothing process will be described as an example.

The smoothing section of the moving average can be appropriately determined, for example, based on the time length of one frame set by the voice input unit 21. As an example, when the time length of one frame is 20 ms, the smoothing section of the moving average is set in the range of 100 ms to 300 ms.
Fig. 5 shows an example of a smoothed signal corresponding to the VAD signal shown in Fig. 4. In this way, by performing smoothing processing, it is possible to express discrete signals of "1" and "0" as a continuous signal.

The state management unit 32 manages the speech state, such as the speech in progress state and the paused state, based on the smoothed signal output from the smoothing processing unit 31 .
The state management unit 32 includes an utterance determination unit 34 and a stoppage determination unit 35 .
The speech initiation determination unit 34 determines that the state is a speech initiation state when, in the stopped state, the smoothed signal output from the smoothing processing unit 31 remains equal to or greater than the speech start threshold for a predetermined period (speech start determination time).
When the smoothed signal is maintained at or below the speech stop threshold for a predetermined time (speech stop determination time) during speech, the stop state determination unit 35 determines that the speech state is in a stop state.
Here, the speech stop threshold is set to a value higher than the speech start threshold.
The speech start determination time and speech stop determination time can be set appropriately, but it is preferable to set both in the range of 200 ms or more and less than 500 ms.

For example, as shown in FIG. 6, at times t1, t2, and t3, the VAD signal indicates "1", but the smoothed signal is below the speech start threshold, so it is not determined that speech is occurring. At time t4, the smoothed signal becomes above the speech start threshold, and when this state is maintained for the speech start determination time, the speech determination unit 34 determines that the speech is occurring (time t5).

In addition, in FIG. 6, at time t6, the smoothed signal is below the speech stop threshold, but this state is not maintained for the speech stop determination time, so it is not determined to be a stopped state. On the other hand, at time t7, when the smoothed signal becomes below the speech stop threshold and this state is maintained for the speech stop determination time, the stopped state determination unit 35 determines that it is a stopped state (time t8).

The state management unit 32 outputs the speech start time, which is the time when it is determined that the speech has started, and the speech stop time, which is the time when it is determined that the speech has stopped, to the voice data management unit 33. For example, for the smoothed signal illustrated in FIG. 6, time t5 is output as the speech start time, and time t8 is output as the speech stop time.

Figure 7 is a speech state transition diagram in the state management unit 32. As shown in Figure 7, when the smoothing signal goes from a state below the speech start threshold (S1) to a state above the speech start threshold (S2) and this state is maintained for the speech start determination time, the state transitions from the stopped state to the speaking state. On the other hand, if the smoothing signal goes above the speech start threshold but the state is not maintained for the speech start determination time, the state returns to S1 again.

In the speaking state, if the smoothing signal goes from a state exceeding the speech stop threshold (S3) to a state below the speech stop threshold (S4) and this state is maintained for the speech stop determination time, the state transitions from the speaking state to the stopped state. On the other hand, if the smoothing signal goes below the speech stop threshold but this state is not maintained for the speech stop determination time, the state returns to S3 again.

Returning to FIG. 3, the voice data management unit 33 determines an utterance section based on an utterance start time and an utterance stop time, and transmits voice data in the utterance section to the server 50.
For example, the voice data management unit 33 determines the start time of the speech section based on the speech start time and a preset speech start margin. Specifically, the voice data management unit 33 determines the start time of the speech section to be a time that is earlier than the speech start time by the speech start margin.

The voice data management unit 33 also determines the end of the speech section based on, for example, the speech stop time and a predetermined speech stop margin. Specifically, the voice data management unit 33 determines the end of the speech section to be a time that is earlier than the speech stop time by the speech stop margin.

For example, as shown in FIG. 8, when time t5 is output as the speech start time and time t8 is output as the speech stop time from the state management unit 32, the voice data management unit 33 determines time Ts as the start time of the speech section and time Te as the end time of the speech section. As a result, the speech section is determined to be the period from Ts to Te. Note that the speech section shown in FIG. 8 illustrates a case where the speech stop margin time is a positive value.

The speech start margin and speech stop margin may be set according to the specifications of a voice recognition unit mounted on the server 50, which will be described later, and are determined by the length of the buffer required before and after the voice data to be recognized.
The speech start margin may be set, for example, in the range of 400 ms to 1000 ms.
The speech stop margin may be set, for example, in the range of -400 ms to 50 ms. If the speech stop margin is a negative value, it means that the end of the speech section is later than the speech stop time.

For example, when in a stopped state, the audio data management unit 33 manages the audio data output from the audio input unit 21 in a fixed ring buffer. Figure 9 shows a conceptual diagram of a ring buffer. Each time new audio data is input, the audio data management unit 33 moves the pointer of the ring buffer and stores the new audio data in a specified memory area.

When the voice data management unit 33 transitions from the stopped state to the speaking state, it stores the voice data in the ring buffer from the start of the speech section onwards in the variable length buffer. In the speaking state, the voice data management unit 33 adds new voice data to the variable length buffer each time the data is input. When the voice data management unit 33 transitions from the speaking state to the stopped state, it reads out the voice data from the beginning of the variable length buffer to the end of the speech section as voice data for the speech section, and transmits it to the server 50 (see Figure 1) via the communication network 2.

Fig. 10 is a functional configuration diagram showing an example of functions provided by the server 50. As shown in Fig. 10, the server 50 includes a request management unit 51 and a voice recognition unit 52. The request management unit 51 manages the voice data received via the communication network 2 in the order of reception. The voice recognition unit 52 performs voice recognition on the voice data in the order of reception, and transmits text data of the voice result to the corresponding client terminal 10. Note that these voice recognition functions provided by the server 50 may be appropriately implemented using publicly known technologies.

Next, the operation of the voice recognition support unit 20 according to this embodiment will be briefly described.
First, the audio signal from the microphone 18 or the like is divided at a predetermined time interval to obtain audio data for each frame, and voice activity detection (VAD) is performed for each frame. Then, the VAD signal is smoothed to obtain a smoothed signal.

Then, the state of speech is determined using the smoothed signal. For example, if the state in which the smoothed signal is equal to or greater than the speech start threshold is maintained for the speech start determination time, it is determined to be in a speech state. Also, if the state in which the smoothed signal is equal to or less than the speech stop threshold is maintained for the speech stop determination time, it is determined to be in a stopped state. Then, the speech section is determined using the speech start time determined to be the speech start state, the speech stop time determined to be the speech stop state, the speech start margin time, and the speech stop margin time, and the voice data of the speech section is transmitted to the server 50.

For example, the server 50 performs voice recognition on the speech section voice data received from each client terminal 10 in the order in which it is received, and transmits the voice recognition results to the corresponding client terminal 10.

As described above, according to the information processing device, speech recognition assistance method, and speech recognition assistance program of this embodiment, the speech section is determined using a smoothed signal obtained by smoothing the VAD signal, so that it is possible to determine an appropriate speech section even when noise is input or when an interruption occurs during the user's speech. This makes it possible to reduce wasteful speech recognition, such as performing speech recognition on noisy speech data, and to improve the efficiency of speech recognition. In addition, speech data related to a series of utterances can be treated as a single block of speech data for speech recognition. As a result, it is possible to improve the accuracy of speech recognition.

In the above-described first embodiment, the smoothing processing unit 31 obtains a smoothed signal using a moving average, but the smoothing processing is not limited to the moving average. For example, the smoothed signal may be obtained using a regression model. Specifically, a regression model is obtained by performing regression analysis based on output signals from the voice activity detection unit for a certain number of times in the past from the current time. In this case, the regression model is preferably a linear model. The current value is then calculated using the obtained regression model.

Second Embodiment
Next, a speech recognition system according to a second embodiment of the present invention will be described with reference to the drawings. In the first embodiment described above, fixed values are used as the speech start threshold and speech stop threshold, but in this embodiment, the speech start threshold and speech stop threshold are dynamically set based on the voice data.
Hereinafter, configurations common to the first embodiment described above will be denoted by the same reference numerals and description thereof will be omitted, and differences will be mainly described.

11 is a functional configuration diagram showing an example of a voice recognition support function of the client terminal 10a according to the present embodiment. As shown in FIG. 11, the correction filter unit 23a of the voice recognition support unit 20a includes a threshold setting unit 36.
The threshold setting unit 36 obtains a second smoothed signal by smoothing the VAD signal output from the voice activity detection unit 22 in a smoothing section (second smoothing section) longer than the smoothing section (first smoothing section) of the smoothing processing unit 31, and sets the speech start threshold and the speech stop threshold using this second smoothed signal. Here, the speech start threshold and the speech stop threshold may be set to the same value or different values. For example, the second smoothed signal may be used as the speech start threshold and the speech stop threshold. Also, the second smoothed signal may be set as the speech start threshold, and a signal obtained by increasing the second smoothed signal by a predetermined value may be set as the speech stop threshold.
The second smoothing section is set, for example, in the range of 300 ms to 900 ms.

When the speech start threshold and speech stop threshold are set by the threshold setting unit 36 in this manner, these thresholds are output to the state management unit 32. As a result, the speech start threshold and speech stop threshold set by the threshold setting unit 36 are used by the state management unit 32 to manage the speech state.

Figure 12 shows an example of a speech section determined by the voice recognition support unit 20a according to this embodiment. Figure 12 illustrates a case in which the second smoothed signal is used as the speech start threshold and speech stop threshold. In other words, the speech start threshold and speech stop threshold are set to the same value.

In this way, according to the information processing device, speech recognition assistance method, and speech recognition assistance program of this embodiment, the speech start threshold and speech stop threshold are dynamically set based on the VAD signal, making it possible to determine the speech section according to the user environment (frequency of noise and speech interruptions). This makes it possible to improve the accuracy of estimating the speech section, and is expected to improve the accuracy of speech recognition.

Third Embodiment
Next, a speech recognition system according to a third embodiment of the present invention will be described with reference to the drawings. In the first and second embodiments described above, the speech state is managed using a smoothed signal obtained by smoothing the VAD signal output from the voice activity detector 22. However, in this embodiment, the speech state is managed directly using the VAD signal without smoothing.
Hereinafter, configurations common to the first embodiment described above will be denoted by the same reference numerals and description thereof will be omitted, and differences will be mainly described.

13 is a functional configuration diagram showing an example of a voice recognition support function of the client terminal 10b according to the present embodiment. As shown in FIG. 13, the correction filter unit 23b of the voice recognition support unit 20b is configured such that the smoothing processing unit 31 is omitted.
According to such a configuration, the state management unit 32 judges the speaking state and the stopped state using the VAD signal output from the voice section detection unit 22 and the preset speech start threshold and speech stop threshold. Specifically, the speaking determination unit 34 of the state management unit 32 judges that the state is speaking when the state in which the VAD signal is equal to or greater than the speech start threshold is maintained for a predetermined speech start determination time, in other words, when the state in which the output of the VAD signal is "1" is maintained for the speech start determination time. Moreover, the stopped state determination unit 35 judges that the state is stopped when the state in which the VAD signal is equal to or less than the speech stop threshold is maintained for a predetermined speech stop determination time, in other words, when the state in which the output of the VAD signal is "0" is maintained for the speech stop determination time.

Then, the speech data management unit 33 determines the speech section based on the speech start time and speech stop time detected by the state management unit 32. By performing such speech recognition support processing, the speech section is determined as shown in FIG. 14.

In this way, according to the information processing device, speech recognition assistance method, and speech recognition assistance program of this embodiment, the VAD signal is not smoothed, but the VAD signal is directly used to manage the speech state, making it possible to estimate the speech section with a simple configuration. Furthermore, even if the input voice signal contains noise or if a break occurs midway through the speech, it is possible to determine the appropriate speech section.

The present invention has been described above using embodiments, but the technical scope of the present invention is not limited to the scope described in the above embodiments. Various modifications or improvements can be made to the above embodiments without departing from the gist of the invention, and forms with such modifications or improvements are also included in the technical scope of the present invention. In addition, the above embodiments may be combined as appropriate.

For example, in each of the above-described embodiments, the correction filter units 23, 23a, and 23b are provided in the client terminals 10, 10a, and 10b, but this is not limiting. For example, the correction filter units 23, 23a, and 23b may be provided in the server 50.

1: Speech recognition system 2: Communication network 10: Client terminal (information processing device)
10a: Client terminal (information processing device)
10b: Client terminal (information processing device)
11: CPU
12: Main memory 13: Secondary storage device 14: External interface 15: Communication interface 16: Input device 17: Display 18: Microphone (microphone)
19: Speaker 20: Voice recognition support unit 20a: Voice recognition support unit 20b: Voice recognition support unit 21: Voice input unit 22: Voice section detection unit 23: Correction filter unit 23a: Correction filter unit 23b: Correction filter unit 31: Smoothing processing unit 32: State management unit 33: Voice data management unit 34: Speaking determination unit 35: Stop determination unit 36: Threshold setting unit 50: Server 51: Request management unit 52: Voice recognition unit

Claims (7)

a voice activity detection unit for detecting a voice activity in which sound is generated from the voice data;
a smoothing processor that smoothes an output signal of the voice activity detector;
a speech-on-call determination unit that determines the state as being in a speech-on-call state when a state in which the smoothed signal output from the smoothing processing unit is equal to or greater than a speech-on-call threshold is maintained for a preset speech-on-call determination time in a stopped state;
a stop state determination unit that determines the state as a stop state when a state in which the smoothed signal is equal to or less than an utterance stop threshold is maintained for a preset utterance stop determination time in the utterance state;
and a voice data management unit that determines a speech section based on an utterance start time determined to be an utterance in progress and an utterance stop time determined to be an utterance stopped state. The information processing device according to claim 1, wherein the voice data management unit determines the start time of a speech section based on the speech start time and a preset speech start margin, and determines the end time of a speech section based on the speech stop time and a preset speech stop margin. The information processing device according to claim 1, further comprising a threshold setting unit that sets the speech start threshold and the speech stop threshold using the smoothed signal. The information processing device according to claim 3, wherein the threshold setting unit smoothes the output signal of the voice activity detection unit over a section longer than the smoothing section of the smoothing processing unit, and uses this smoothed signal to set the speech start threshold and the speech stop threshold. The information processing device according to claim 1, wherein the smoothing processing unit smoothes the output signal of the voice activity detection unit using a moving average. a voice activity detection step of detecting a voice activity in which sound is generated from the voice data;
a smoothing process step of smoothing an output signal of the voice activity detection process;
a speech-on-call determination step of determining, in a stopped state, that the state is one in which the smoothed signal output from the smoothing processing step is equal to or greater than a speech-on-call threshold for a preset speech-on-call determination time;
a stop state determination step of determining, in the speaking state, that the state is a stop state when the state in which the smoothed signal is equal to or less than the speech stop threshold is maintained for a preset speech stop determination time;
and a voice data management step of determining a speech section based on an utterance start time determined to be an utterance in progress and an utterance stop time determined to be an utterance stop state. A speech recognition support program for causing a computer to execute the speech recognition support method according to claim 6 .