JP3097764B2 - Voice input device with guidance voice - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice input device with a guidance voice for outputting a guidance voice to a speaker, prompting the speaker to generate a voice, and receiving the voice from the speaker.

[0002]

2. Description of the Related Art A so-called hands-free speech recognition device or the like which uses a speech input device away from a speaker to clarify the timing of speech to a speaker who inputs speech to the device. It is important to add a guidance signal generating device and the like in order to reduce the operation time of the recognition device and the load on the algorithm and to give the speaker a sense of affinity for the device. Since such a guidance signal is generally used as a substitute for a device such as a keyboard in which a voice recognition device is manually input, and because it is desired that the speaker himself / herself feedback the utterance result and immediately understand it, A device that appeals to hearing rather than a device that appeals to sight and touch is desired. That is, as the guidance signal generator, generally, a guidance method using voice is a desirable form.

However, in the voice guidance method, since the guidance voice itself is a signal in the same voice frequency band as the voice emitted by the speaker, the voice is input to the apparatus during the generation of the guidance voice. Considering the case, since the voice signal input device receives the guidance voice together with the voice of the speaker, a voice pattern in which the guidance voice is superimposed is input to the device,
This may cause a recognition error in the recognition device.

[0004] There are four ways to solve this problem. First, there is a method of not accepting a voice input during the generation of the guidance voice. However, this method imposes a time constraint so that the speaker does not speak until the guidance speech is completely extinguished. In addition, if the speaker utters while the guidance voice is being generated, the input voice during the guidance utterance time cannot be analyzed, so that a so-called word truncation phenomenon occurs, which causes a recognition error in the recognition device. Second, there is a method of suppressing the volume of the guidance voice. This method also includes a method of terminating the guidance voice when the voice of the speaker is detected by the algorithm of the voice input device. However, in a recognition device used under high noise, reducing the volume of the guidance voice is not practical for giving clear guidance voice. In addition, when the voice is detected, the guidance voice is cut off. If the algorithm of voice detection is not perfect, the same word truncation as shown in the first method will occur, and the apparatus will be installed in a room with reverberation. When used, it is difficult to completely cancel the reverberation component of the guidance voice. Third, there is a method of acoustically disconnecting a speaker that emits guidance voice from a voice input device. This method separates an audio system, such as a telephone receiver, for allowing a speaker to listen to guidance, and an audio system for inputting the speaker's voice.
However, in this method, a guidance generating device that is in close contact with the speaker is required in order to separate the speaker from the sound input acoustic system, and the speaker is troublesome to wear the device. Fourth, there is a method of removing the superimposed guidance sound from the sound signal input to the device afterwards. As an example of this, an original signal of guidance is stored in advance in the device by an adaptive prediction filter technique, and a guidance audio signal and an audio signal generated accompanying the guidance signal are converted from the audio signal received by the audio input device. There is a method that attempts to estimate and remove based on However, in this method, when the estimation of the guidance signal is not complete, waveform distortion is generated in the voice signal of the speaker, which leads to deterioration of the recognition rate in the recognition device. Further, in order to learn the coefficients of the adaptive prediction filter, a guidance signal for learning the filter and a learning time are required separately from the guidance for prompting the speaker to speak. Therefore, it is difficult to separate the guidance voice from the input voice in any of these four methods without causing any adverse effect on the recognition device and without burdening the speaker.

[0005]

Each of the above-mentioned methods has its own drawbacks. If the voice of the speaker is superimposed on the guidance voice, erroneous recognition will occur if this voice input device is applied to a recognition device. Problem.

[0006] Among the above-mentioned methods, the fourth method requires prior learning using an adaptive prediction filter, and when the estimation of the filter is not perfect, distortion occurs in the voice of the speaker. He pointed out that there was a problem. However, this problem can be solved by limiting the frequency band of the guidance sound to be removed and removing the guidance sound within that band without distorting the sound waveform in other frequency bands. In addition, in this removal method, prior learning is not required if only the band removal filter is used without using the adaptive prediction filter. The voice obtained by removing only the same frequency band as the guidance voice obtained in this way is different in sound quality from the direct voice input by the speaker, but the frequency band to be removed is limited to a band with no auditory problem Then, the sound waveform when reproduced from the sound input device has little change in hearing. Also, when this speech is used in a speech recognition device, if the standard pattern itself used for recognition is also constituted by speech that has passed through such a band elimination filter, a result in which the recognition rate is hardly deteriorated can be obtained. Can be Also, in this method, the quality of the guidance voice itself is reduced, but by increasing the volume of the guidance voice, the phenomenon that the guidance voice itself is difficult to hear can be reduced. In this way, by separating the frequency bands used for the guidance voice and the input voice, it is possible to realize a voice input device with a guidance voice in which even if the guidance voice is superimposed, the input voice itself is hardly affected. By improving the fourth method in this way, the above-mentioned disadvantages can be solved.

[0007] The present invention has been made in view of the above,
An object of the present invention is to provide a voice input device with guidance voice that can take out only a voice signal that does not include a component of the guidance voice even when a voice of a speaker is superimposed on the guidance voice.

[0008]

In order to achieve the above object, a voice input device with guidance voice according to the present invention comprises a signal generation means for generating a guidance signal for prompting a speaker to speak, and a signal generation means for the signal generation means. A band-pass filter that extracts a signal of a specific frequency band from the output signal, a loudspeaker that generates the output signal as an audio signal so that a speaker can hear the output signal of the band-pass filter, The gist of the present invention is to have a signal input unit that receives a voice output from a speaker and outputs the signal as an electric signal, and a band removal filter that extracts a signal other than the specific frequency band from an output signal of the signal input unit. .

[0009]

In the voice input device with guidance voice according to the present invention, the guidance signal is limited to a signal in a specific frequency band by passing the guidance signal through a band-pass filter, and the voice output from the speaker in response to the guidance voice is provided. On the other hand, a signal in the same frequency band as the guidance signal is removed using a band elimination filter, so that the guidance signal is not included even if the guidance voice and the voice signal of the speaker are temporally superimposed. An audio signal can be obtained.

[0010]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a voice input device with guidance voice according to one embodiment of the present invention. In the figure, a guidance signal source 11 is a speaker 1
9 is a signal source that outputs a guidance voice prompting utterance, for example, a voice recorded by a tape recorder, or an electrically synthesized voice, for example. Bandpass filter 1
2 is a filter for passing only a signal in a specific frequency band, and the pass band is, for example,
1 kHz to 2 kHz. Also, its pass band is
It is not necessary to limit to one band. The amplifier 13 amplifies the signal after passing through the band-pass filter 12 without distortion. The speaker 14 emits the output signal of the amplifier 13 to a space and causes the speaker 19 to listen. The microphone 15 is a microphone that picks up the voice of the speaker 19. The band elimination filter 16 removes a signal in a frequency band to be passed by a band-pass filter,
This is a filter for passing signals of other frequency bands, and the pass band is, for example, 1
00Hz to 1kHz and 2kHz to 5kHz
And The audio output unit 17 outputs an audio signal that has passed through the band elimination filter 16, and is, for example, an electrical terminal, a tape recorder, a magnetic memory, or the like. The switch 18 is a switch for a speaker to request an operation of the apparatus of the present embodiment, and is, for example, an electrical contact or an optical sensor.

The operation of the embodiment shown in FIG. 1 will be described below.

When the speaker 19 inputs a voice to the apparatus, first, the speaker operates the switch 18 to give an operation request of the apparatus. In response to the operation request, the apparatus generates a voice signal such as “Please input voice” from the guidance signal source 11 to make the speaker 19 confirm that the request has been accepted. This signal is applied to the bandpass filter 1
2, the signal becomes only a specific frequency band of 1 kHz to 2 kHz, and is emitted from the speaker 14 through the amplifier 13. Further, in response to the operation request of the switch 18, the microphone 15, the band elimination filter 16, and the audio output unit 17 also start operating, and wait for a voice input by the speaker. The time to wait for input is, for example, a finite time of several seconds to several tens of seconds. Here, when there is a voice from the speaker 19 during or after the guidance signal is emitted from the speaker 14, the voice is input from the microphone 15 and passes through the band elimination filter 16. The audio signal is removed from a specific frequency band of 1 kHz to 2 kHz, and the audio output unit 17 outputs an audio signal that does not include the guidance signal band.

FIG. 2 is a block diagram of another embodiment of the present invention. The voice input device of this embodiment is applied to a voice recognition device.

In FIG. 2, the guidance signal source 21 is
Similar to the guidance signal source of FIG. 1, the signal source is a signal source that outputs a guidance voice prompting the speaker 33 to utter. For example, based on an output result of a label output unit 31 described later, a label of the recognition result is output. It has a function of outputting to the speaker 33 for confirmation. A band-pass filter 22, an amplifier 23,
The speaker 24, the microphone 25, and the band elimination filter 26 are respectively the band-pass filter 1 in FIG.
2, amplifier 13, speaker 14, microphone 15,
It has the same function as the band elimination filter 16.
The voice analysis unit 27 calculates a parameter representing a characteristic of a spectral region from the voice signal output from the band elimination filter 26, and uses, for example, a short-time LPC cepstrum analysis. The feature storage unit 28 stores the time series of the parameters calculated by the speech analysis unit 27 in units of time intervals, assigns a label, and configures a standard pattern used for speech recognition. For example, if this embodiment is used as a word speech recognition device, it is assumed that one word section is stored. The label input unit 29 receives, for example, for each one-hour section accumulated in the feature storage unit 28, by manual input of the speaker 33, or
For example, labels are assigned in an automatic order, and the feature storage unit 2
8 is stored. The similarity calculation unit 30 performs similarity calculation between a parameter string input from the voice analysis unit 27 as a voice whose label is unknown and the standard pattern stored in the feature storage unit 28. The method uses, for example, DP matching. The label output unit 31 outputs the label of the standard pattern determined to have the highest similarity to the input speech whose label is unknown from the result of the similarity calculation unit 30, and the output is the same. A guidance signal source 21 is provided. Switch 3
Reference numeral 2 denotes a switch for operating the voice recognition device.

The operation of the block diagram shown in FIG. 2 will be described below.

In this embodiment, a registration mode for storing a standard pattern used for speech recognition, a recognition mode for inputting speech to calculate a similarity to the standard pattern, and outputting a label of the standard pattern having the highest similarity. Two operations of the mode are performed.

First, the operation in the registration mode will be described.

First, the speaker 33 operates the switch 32 to the recognition mode to give an operation request of the recognition device. In response to the operation request, the recognition device sends the guidance signal source 21 to the speaker 33 to confirm that the request has been accepted.
For example, a guidance voice such as “Please say the recognized word” is generated. This signal is converted into a signal only in a specific frequency band of 1 kHz to 2 kHz by the band-pass filter 22, and is emitted from the speaker 24 through the amplifier 23. At this time, the operation of the microphone 25, the band elimination filter 26, and the voice analysis unit 27 also starts, and waits for the input of voice by the speaker. The time to wait for an input is, for example, within a few seconds to a few tens of seconds after the operation of the switch 32 is started. Here, during or after the guidance signal is emitted from the speaker 24,
When there is an utterance from the utterer 33, this sound is input from the microphone 25 and passes through the band elimination filter 26 to become an audio signal from which a specific frequency band of 1 kHz to 2 kHz has been removed. It is analyzed as an audio signal that does not include a guidance signal.
At this time, the voice analysis unit 27 selects only a voice section and outputs an analysis result. The voice section detection method is realized by, for example, observing a short-time signal power at the output of the voice analysis unit 27 and detecting a section in which the signal power exceeds a certain threshold for a predetermined time or more. The output of the voice analysis unit 27 is labeled from the label input unit 29 as a standard pattern used for recognition, and stored in the feature storage unit 28. Next, the operation in the recognition mode will be described.

The speaker 33 operates the switch 32 to the recognition mode to give an operation request of the recognition device. In response to the operation request, the recognition device generates a guidance voice such as “Please input voice” from the guidance signal source 21 to make the speaker 33 confirm that the request has been accepted. This signal becomes a signal only in a specific frequency band of 1 kHz to 2 kHz by the band-pass filter 22 as in the case of the registration mode, and is emitted from the speaker 24 through the amplifier 23. At this time, the microphone 25, the band elimination filter 26, and the
The operation is also started for 7 and the input of a voice by the speaker is awaited. The time to wait for an input is, for example, within a few seconds to a few tens of seconds after the operation of the switch 32 is started. Here, when the speaker 33 utters while the guidance signal is being emitted from the speaker 24 or after the guidance signal is emitted, the utterance is transmitted to the microphone 2 as in the case of the registration mode.
5 and 1k via the band elimination filter 26.
The audio signal is obtained by removing a specific frequency band from 2 Hz to 2 kHz, and only the audio section is selected by the audio analysis unit 27 and output. The output of the voice analysis unit 27 is subjected to pattern matching in the similarity calculation unit 30 with a plurality of standard patterns stored in the feature storage unit 28 as an input voice pattern whose label is unknown. As a result, the label of the standard pattern determined to have the highest similarity is output to the label output unit 3.
1 is output. The output label can be fed back to the guidance signal source 21 and used for confirming the recognition result of the speaker 33, for example. For example, if the voice labels registered as standard patterns are "Suzuki", "Tanaka", and "Sato", and the label of the recognized standard pattern is "Suzuki", the recognition word is "Suzuki". Is generated from the guidance signal source 21 so that the speaker 33 can listen to and confirm the synthesized voice such as "?", And input the confirmation voice such as "Yes" or "No" to the speaker 33. Can be.

In order to confirm that the recognition performance of the speech recognition apparatus according to the present embodiment is unlikely to be degraded due to the superimposition of the guidance speech signal, an experiment simulating the embodiment was performed. The experiment consisted of five adult men with guidance voice superimposed.
It recognizes 100 words of the Japanese city name from which the name was uttered. The guidance voice was produced by an adult female in her twenties saying "Please utter the command name." This is equivalent to the 1k equivalent to the band-pass filter 22 described in the above description.
The signal was passed through a filter that passes only the frequency band from Hz to 2 kHz, and was superimposed at the “command name” portion so that the power had the same level as the short-time power of the original voice of the speaker. With respect to the guidance superimposed speech, a change in the speech recognition rate between when the guidance was removed by the band removal filter and when the guidance was not removed was observed. The used band elimination filter is 1 k, which is equivalent to the band elimination filter 26 described above.
A filter that removes frequencies from 2 Hz to 2 kHz,
With respect to the guidance superimposed voice, when removing the guidance, the filter is passed through this filter, and when the guidance is not removed, the filter is not passed through this filter.
A PC cepstrum analysis was performed, a speech pattern was created, and a recognition experiment was performed. In addition, the standard pattern was created by setting so that the standard pattern and the input voice pattern were analyzed under the same acoustic conditions, including the presence / absence of passing through a filter. According to the experimental results averaged for the five speakers, when there is no superimposition of the guidance voice, the guidance voice is superimposed on the voice having the recognition rate of 98.6%, and when the voice is not removed by the band elimination filter, Recognition rate 84.8%
And a decrease of about 14%. On the other hand, it was confirmed that the speech removed by the band elimination filter had a recognition rate of 93.6%, and a decrease in the recognition rate could be suppressed to 5%. As described above, in the voice recognition apparatus with guidance, when the guidance voice whose band is limited is used, even if the guidance voice is superimposed, the superposition of the guidance voice is removed by the band elimination filter, so that the deterioration of the recognition rate is suppressed to a small extent. Can be.

[0022]

As described above, according to the present invention,
Pass the guidance signal through a band-pass filter to limit the signal to a specific frequency band, and use the band elimination filter for the voice output from the speaker in response to this guidance voice and use the same as the guidance signal. Therefore, even if the guidance voice and the voice signal of the speaker are temporally superimposed on each other, a voice signal containing no guidance signal component can be obtained. It is not necessary to arrange the microphone and the microphone for voice input acoustically, which makes it easy to configure a hands-free voice input device, and makes it possible to produce an utterance at any time after the guidance signal is started. Acceptable, compatible voice input devices can be configured, and conventional guidance using adaptive prediction filters, etc. Learning algorithm filters required in the voice removal method, it is possible to eliminate the guidance voice signal and the learning time for learning. Further, when the present voice input device is applied to a voice recognition device, the recognition rate does not decrease even if the guidance voice is superimposed on the voice or the guidance voice is insufficiently removed.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a voice input device with guidance voice according to one embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration when a voice input device with guidance voice according to another embodiment of the present invention is applied to a voice recognition device.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 11 Guidance signal source 12 Band pass filter 14 Speaker 15 Microphone 16 Band elimination filter 17 Audio output part

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G06F 3/16 G10L 13/00 G10L 15/20 G10L 21/02

Claims (1)

(57) [Claims] 1. A signal generating means for generating a guidance signal for encouraging a speaker to utter, a band-pass filter for extracting a signal of a specific frequency band from an output signal of the signal generating means, and an output of the band-pass filter A loudspeaker for generating the output signal as an audio signal so as to allow the speaker to listen to the signal, signal input means for receiving a voice output from the speaker in response to the audio signal, and outputting the signal as an electric signal; A voice input device with guidance voice, comprising: a band elimination filter for extracting a signal other than the specific frequency band from an output signal of a signal input unit.