JP2003255971A - Speech extracting method and speech extracting device using the method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech extracting method which extracts a speech in noisy environment, and its device. <P>SOLUTION: The speech extracting method and its device are characterized in that attention is paid to the difference in acoustic feature between background noise and a speech. When an octave band level analysis of a speech including background noise is taken, its level difference becomes remarkable in a specified band. For the purpose, a filter 30 which has such a characteristic that the level difference between the both becomes large is generated through the analysis. This filter part 30 is made to operate on the output of a microphone 10. Consequently, a band where the difference in power between the speech and background noise is large is passed. Namely, the filter part 30 relatively emphasizes the speech. According to the filter output, a threshold determination part 40 determines a threshold. Lastly, a speech extraction part 60 sets the threshold for a speech signal on a time series and regards a part above the threshold as a speech section. In this section, the speech level is relatively large, so the speech can accurately be extracted even in noisy environment. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice extraction method for improving a voice recognition rate in a noise and a voice extraction apparatus using the method. In particular, a voice extraction method and apparatus for applying a filter based on band level analysis corresponding to human hearing to a detection sound including background noise, determining an effective voice section based on the output of the filter, and extracting a voice Regarding INDUSTRIAL APPLICABILITY The present invention can be applied to voice recognition of a vehicle-mounted device that is operated by voice with the vehicle interior noise as background noise.

[0002]

2. Description of the Related Art Conventionally, there are a voice recognition method and a device for recognizing voice under noise. For example, Japanese Patent Laid-Open No. 8-314
There is a voice recognition method and a voice recognition device disclosed in Japanese Patent Publication No. 500. This is a voice recognition method and a voice recognition device for accurately recognizing numbers or the like produced by voice under noise. The threshold for determining the voice section is divided into a first threshold and a second threshold, and the maximum value of the voice signal obtained by the first threshold,
Alternatively, the second threshold value is dynamically determined based on the maximum value of the spectrum, and voice recognition is performed on the cut-out voice signal. That is, the feature is that the second threshold value is changed corresponding to the changing voice signal, and the voice section is determined by the change.

In addition, there is a voice section detection method disclosed in Japanese Patent Application Laid-Open No. 10-254476. This is a method of recognizing a voice / non-voice acoustic model and determining a voice section. Specifically, a speech recognition apparatus prepares a speech acoustic model learned by using all the voices covering the recognition target vocabulary (class) and a non-speech acoustic model learned by using an unvoiced section of the speech. To do. Then, the likelihood ratio between the voice acoustic model and the non-voice acoustic model is calculated for each appropriate section length of the input signal. When a section whose likelihood ratio exceeds a predetermined threshold continues for a certain period of time, it is set as the beginning of a voice section, and thereafter, when a section whose likelihood ratio is below a predetermined threshold continues for a certain period of time, it is set as This is a method of terminating the voice section. As a result, the voice section is accurately determined, and the voice recognition rate is said to be improved.

[0004]

[Patent Document 1] Japanese Unexamined Patent Publication No. Hei 8
In the voice recognition method and the voice recognition device disclosed in Japanese Patent Laid-Open No. 314500, when sudden noise is mixed in the voice signal clipped by the first threshold value, the second threshold value also changes, and the voice interval is changed. There may be an error. That is, even if the threshold value is divided into the first threshold value and the second threshold value, the influence of external noise cannot be reliably removed. That is, it is not a method and an apparatus for determining a voice section without the influence of background noise and always recognizing a voice accurately. Also, JP-A-10-2544
The speech segment detection method disclosed in Japanese Patent Publication No. 76-76 requires a large amount of calculation because pattern recognition processing is performed. Therefore, there is a problem that it is difficult to adopt for voice recognition of a device that requires real-time property, for example, a vehicle-mounted navigation device.

The present invention has been made to solve the above-mentioned problems, and an object thereof is to apply a filter based on an analysis method suitable for human hearing to a detection sound including background noise, This is to set a threshold value based on the filter output and accurately determine the voice section. And
Accordingly, it is an object of the present invention to provide a method and an apparatus for accurately extracting voice. Another object of the present invention is to apply the voice extraction method and device of the present invention to a voice recognition device mounted on a vehicle so that the voice of the driver can be clearly recognized to improve the safety of driving.

[0006]

According to a first aspect of the present invention, there is provided a voice extraction method in a noisy state, wherein a filter based on a band level analysis result is applied to a detection sound including background noise, and the filter is applied. A feature is that a threshold value is determined based on the output, and a section of a level of the filter output that is equal to or higher than the threshold value is set as a voice section and voice is extracted from the voice section.

The speech extraction method according to a second aspect is the speech extraction method according to the first aspect, wherein the band level analysis is an octave band level analysis. The voice extraction method according to claim 3 is the voice extraction method according to claim 1 or 2, wherein the function of the filter is weighting by adding a predetermined weight to a predetermined band by band level analysis and adding the weight. It is characterized by addition.

The voice extraction method according to a fourth aspect is the voice extraction method according to the first or second aspect, wherein the function of the filter is that the level of each predetermined band obtained by band level analysis is perceived to be large. It is characterized in that it is a weighted addition in which a scale conversion is performed, and a higher weight is added to the converted value in a higher frequency band to be added. Further, the voice extraction method according to claim 5 is the voice extraction method according to any one of claims 1 to 4, wherein the voice extraction is performed based on the frequency component of the calculated voice section and the frequency of the background noise. It is characterized in that it is obtained by subtracting the components.

Further, the voice extraction device according to claim 6 is a voice extraction device in a noisy state, wherein voice detection means for detecting a voice including background noise and band level analysis of the detected sound of the voice detection means. A filter means for applying a filter characteristic based on the threshold value, a threshold value determining means for determining a threshold value based on the output of the filter means, and a voice section calculating means for determining a section having a level equal to or higher than the threshold value by the threshold value determining means as a voice section; And a voice extraction unit for extracting a voice from the voice section.

Further, the voice extraction device according to claim 7 is the voice extraction device according to claim 6, wherein the band level analysis used for determining the filter characteristic is an octave band level analysis. To do. Further, the voice extraction device according to claim 8 is the voice extraction device according to claim 6 or 7, wherein the filter means adds a predetermined weight to a predetermined band of the detected sound to which the filter characteristic is applied. It is characterized in that weighted addition is performed.

Further, the voice extraction device according to claim 9 is the voice extraction device according to claim 6 or 7, wherein the filter means applies the detected sound to which filter characteristics (band level analysis result) are applied. It is characterized in that the level of a predetermined band is scale-converted to a perceptual size, and the converted value is weighted and added by adding a greater weight to a higher frequency band. The voice extraction device according to claim 10 is the voice extraction device according to any one of claims 6 or 9, wherein the voice extraction means detects the voice section calculated by the voice section calculation means. It is characterized in that the frequency component of background noise is subtracted from the sound frequency component. The voice extraction device according to claim 11 is the voice extraction device according to any one of claims 6 to 10, wherein the background noise is noise in the vehicle interior and the voice is mounted on the vehicle of the driver. It is characterized by being an instruction to the device.

[0012]

According to the voice extracting method of the first aspect, the filter based on the band level analysis is applied to the detected sound. The band level analysis is an analysis for selecting a frequency band (band) from the detected sound so as to correspond to human hearing frequencies exponentially distributed and obtaining the power. This is because the acoustic characteristics of background noise and human voice differ depending on the frequency band as shown in FIG.
Background noise is random noise and periodic noise such as engine sound, but the frequency of human voice is exponentially distributed. Therefore, the detected sound is analyzed in an exponential band focusing on the characteristic. For example, if the 1.25 kHz band is extracted from the band level analysis result of FIG. 4, the background noise can be reduced and the voice can be detected greatly.
That is, the filter characteristic based on the band level analysis is applied to the detection sound including the background noise, and the band having a large difference between the voice information and the background noise is passed to emphasize the voice.

Then, the threshold value is determined based on the output of the filter. For example, a value based on the maximum power in the frequency spectrum of the filter output, for example, -20 dB of the maximum power is set as the threshold value. Then, a section having a level equal to or higher than the threshold is set as a voice section. The voice section is a section on the time axis where the voice is emitted. Then, the voice is extracted from the voice section. Since the background noise is low and the voice power is high in this section, the voice can be extracted efficiently and reliably. As a result, even a voice buried in background noise can be accurately extracted.

The voice extraction method according to a second aspect of the present invention is the voice extraction method according to the first aspect, wherein the band level analysis is octave band level analysis. Octave band level analysis is a method of selecting and analyzing a band that conforms to human exponential hearing. For example, in the case of 1/3 octave band analysis, the center frequency of the initial value is 10
If it is 00 Hz, the ratio with the next stage is 2 ^1/3 , so the center frequencies are 1250 Hz, 1600 Hz, 200
0 Hz, ... In the band above 1000Hz,
These almost agree with the critical band, which is an effective way of explaining human auditory phenomena. Background noise does not conform to the octave band, but human voice and hearing sensation conform to the octave band. Therefore, band analysis in this critical band enables efficient voice acquisition. That is, the voice section can be efficiently determined.

A voice extraction method according to a third aspect of the present invention is the voice extraction method according to the first or second aspect, wherein a predetermined band obtained by band level analysis is added with a predetermined weight by a filter action and added. To be done. This is a band where the level difference between voice and background noise is remarkable after the band level analysis, for example, 1 kHz, 1.25 kHz, 4 kHz.
This is because the speech spectrum centered on is obtained and they are effectively emphasized. In this band, the level difference between the voice and the background noise is remarkable, so if these bands are increased according to the maximum level of the band and added,
The acoustic features of the voice are emphasized more. For example, weighting addition is performed on the center of the spectrum with a ratio of 3: 2: 1. In this way, the background noise and the acoustic features of the voice are further emphasized. Then, after weighted addition, a section that is equal to or more than a predetermined threshold is set as a voice section. Since the background noise is relatively reduced, the voice section is more accurately determined.

The voice extraction method according to a fourth aspect is the voice extraction method according to the first or second aspect, wherein, in filtering the detected sound, first, each predetermined band (critical value) by band level analysis is detected. The level of the band is scaled to an auditory level. Scale conversion is conversion from a dB value of a frequency spectrum value to a sone value. The dB value is a unit of sound pressure by the electrostatic microphone, and the sone value is a scale corresponding to the perception of the size of human hearing. When the sound value is S and the sound pressure is P, Log ₁₀ S = 0.0 for both.
3 (P-40). This is a relationship in which the S value exponentially increases when the P value is plotted on the horizontal axis and the S value is plotted on the vertical axis. That is, the conversion into the sound value means matching with the loudness perception of sound in human hearing. And
Further, the filtering action adds a greater weight to the converted value in the higher frequency band and adds the weight. This is a transformation that matches the perception of sound acuity in human hearing and increases the sound acuity of speech. This is because the high-frequency component is originally included in the pronunciation of consonants and the like, so that the high-frequency band enhancement is effective for voice recognition. As a result, the acoustic characteristics of the voice can be emphasized in the same manner as the processing in human hearing. As a result, the voice section can be set more accurately.

A voice extraction method according to a fifth aspect of the present invention is the voice extraction method according to any one of the first to fourth aspects, wherein the voice extraction is performed from the frequency components of the calculated voice section to the background. It is calculated by subtracting the frequency component of noise. Here, the frequency component of the background noise may be the frequency component of the detected sound immediately before the calculated voice section, or may be the frequency component previously measured in a predetermined band when there is no voice. Then, at the time of voice extraction, the frequency component of the background noise is subtracted from the detected sound frequency component of the voice section calculated (spectral subtraction). Thereby, the voice is effectively extracted.

According to the sixth aspect of the voice extraction device, the voice detection means first detects the voice including the background noise. The filter means then act on the filter characteristics based on the band level analysis. That is, it is processed into a size that is audible. The band level analysis is an analysis for selecting, for example, an exponentially distributed band (frequency band) from the detected sound and obtaining its power. Since the background noise and the voice have different acoustic characteristics, the background noise can be reduced and the voice can be emphasized as a result if the band level analysis is performed exponentially, for example. Then, the threshold value determining means determines the threshold value based on the output of the filter means. For example, a value based on the maximum power in the frequency spectrum of the filter output, for example, −20d of the maximum power.
Let B be the threshold. Then, the voice section calculating means sets a section having a level equal to or higher than the threshold as a voice section. What is a voice section?
It is a section on the time axis where the voice is emitted. Finally, the voice extraction means extracts voice from the voice section.
The extraction is obtained on the frequency axis by subtracting the spectrum of background noise, for example. As described above, according to the voice extraction device of the present invention, the voice is emphasized by the filter means based on the band level analysis to determine the voice section, so that the voice can be extracted reliably and efficiently even if there is background noise. can do.

Further, the voice extraction apparatus according to claim 7 claims.
Item 6. The voice extraction device according to Item 6, comprising a filter unit
Octave band level analysis is used to determine the filter characteristics.
Is used. Octave band level analysis
A method of selecting and analyzing a band that matches the human hearing
is there. For example, with 1/3 octave band analysis,
If the center frequency of the initial value is 1000 Hz,
The ratio to the center frequency is 2 ^1/3Therefore, 1250 sequentially
Hz, 1600 Hz, 2000 Hz, ... Man
Voice and hearing conform to the octave band, background noise
It doesn't follow that, so in bands centered around these
If analyzed, human voice can be emphasized efficiently.
That is, the band level analysis is changed to the octave band level analysis.
As a result, if a filter means based on it is used,
The voice section can be accurately determined. That is,
It is possible to extract voice frequently.

Further, the voice extraction device according to claim 8 is the voice extraction device according to claim 6 or 7, wherein the filter means detects the detection sound to which filter characteristics (band level analysis) are applied. A predetermined weight is added to a predetermined band and added. After the band level analysis, for example, 1 kHz, 1.2
A band centered at 5 kHz and 4 kHz is obtained. In this band, the voice is concentrated and the difference from the background noise is remarkable.
Here, in order to emphasize the voice more effectively, those bands are increased, for example, according to the maximum level of the band, and then added. For example, weighting addition is performed on the center of the spectrum with a ratio of 3: 2: 1. In this way, the background noise and the acoustic features of the voice are further emphasized.
If the voice section is determined based on this, the voice can be more accurately extracted. That is, it is possible to realize the voice extraction device according to claim 6 or 7, which performs voice extraction more accurately.

Further, the voice extraction device according to claim 9 is the voice extraction device according to claim 6 or 7, wherein the filter means applies the filter characteristics (band level analysis) to each detected sound. The level of a predetermined band is scaled to an auditory level. Scale conversion is conversion from the dB value of the frequency spectrum to the sone value. The sound pressure by the electrostatic microphone is detected by the dB value, but the human auditory sense is measured by the dB value.
Do not perceive the value linearly. Human hearing perceives sound pressure with a sone value. Here, when the sound value is S and the sound pressure is P, both have a relationship of Log ₁₀ S = 0.03 (P-40). This is a relationship in which the S value exponentially increases with respect to the P value when the P value is plotted on the horizontal axis and the S value is plotted on the vertical axis. That is, the conversion into the sound value means matching with the perception of the loudness of sound in human hearing. Then, the filter means performs weighted addition in which a greater weight is added to the converted value in the higher frequency band and added. this is,
It is the conversion of sound into sharpness. This is because human voice originally contains high-frequency components in pronunciation of consonants and the like, so that enhancement of a high-frequency band is effective for voice recognition. Thereby, the acoustic feature of the voice can be further emphasized. That is, it is possible to realize a voice extraction device that sets a voice segment more accurately and more reliably extracts a voice.

The voice extraction device according to claim 10 is the voice extraction device according to any one of claims 6 and 9, wherein the voice extraction means is a voice calculated by the voice section calculation means. The background noise frequency component is subtracted from the detected sound frequency component of the section. Here, the frequency component of the background noise may be the frequency component of the detected sound immediately before the calculated voice section, or if there is no voice in advance prior to voice extraction, a predetermined band, for example, octave band level (critical band). It may be a frequency component measured in advance. Then, at the time of voice extraction, the voice extraction means subtracts the frequency component of the background noise from the detected sound frequency component in the voice section (spectral subtraction). Thereby, the voice is effectively extracted.

Further, the voice extraction device according to claim 11 is the voice extraction device according to any one of claims 6 to 10, wherein the background noise is noise in the vehicle interior and the voice is the driver. This is an instruction to the on-vehicle equipment of. Since the voice extraction device according to any one of claims 6 to 10 can efficiently and reliably extract voice even in the presence of background noise, it can be applied to a vehicle-mounted device that requires voice recognition. That is, the driver's voice (instruction) can be extracted efficiently and reliably even if there is vehicle interior noise as background noise. Therefore,
The voice recognition rate of a vehicle-mounted device such as a navigation device is improved, which can contribute to safe driving.

[0024]

BEST MODE FOR CARRYING OUT THE INVENTION A voice extraction method and apparatus according to the present invention will be described below with reference to the drawings. The voice extraction method and apparatus of the present invention are effective under general noise. For example, it is effective under traffic noise, vehicle noise, and noise in factories. Here, a method and a device for extracting a voice in a vehicle interior noise will be described. The vehicle interior noise is composed of periodic noise caused by engine noise and random noise caused by running (road noise, wind noise, etc.). Since the power of these vehicle interior noises is concentrated in the low frequency band, preprocessing for removing them using a high-pass filter is effective during voice recognition. However, at present, even if a high-pass filter is used, the recognition interval cannot be accurately specified, and therefore the recognition rate has not been further improved. Therefore, the inventor pays attention to the fact that the difference between the acoustic characteristics of voice and noise, that is, the difference between the two becomes significant in the audible band (critical band), and emphasizes this band to calculate the voice section. I tried.

FIG. 1 shows an embodiment of the voice extraction device of the present invention. The figure is a system configuration diagram. The voice extraction device of the present invention includes a microphone 10 that is a voice detection unit, a filter unit 30 that is a filter unit whose filter characteristics are determined based on band level analysis, a threshold determination unit 40 that is a threshold determination unit, and a voice section calculation. It is composed of a voice section calculation unit 50 which is a unit and a voice extraction unit 60 which is a voice extraction unit. Note that each of the above-mentioned elements is specifically a CPU 110, a ROM 120, a RAM 130, an A / D as shown in FIG.
The computer device 100 includes a conversion device 140 and an I / O interface device 150. For example, the filter unit 30 includes the A / D conversion device 140, the ROM
Processing program stored in 120, C for executing it
It includes a PU 110 and a RAM 130 that is an execution area.

The operation of each of the above components will be described with reference to the system configuration diagrams of FIGS. 1 and 2 and the flowchart of FIG.
Prior to that, the filter unit 3 is analyzed by band level analysis.
A method for determining the filter characteristic of 0 will be described.
The band level analysis is an analysis method that exponentially selects a frequency band so as to obtain the power of the detected sound from the microphone 10 so as to correspond to the human hearing frequency. It is, for example, an octave level analysis. Background noise in the passenger compartment is a frequency structure in which random noise is added to periodic noise related to the vehicle. On the other hand, the acoustic characteristic of voice is a frequency structure in which the reference frequency is exponentially increased, as is well known. For example, the frequency structure has a reference frequency (for example, 200 Hz) of 2 ⁿ when n is an integer.
Doubled octave structure, or, for example, 2 ^{n / 3} times 1 /
It can be represented by a 3-octave structure. Therefore, if the detected sound including the background noise is subjected to, for example, 1/3 octave level analysis, the background noise is reduced, and as a result, the voice is emphasized. Along with that, a band where the difference between the two is remarkable is found. FIG. 4 shows a graph obtained by performing a 1/3 octave band level analysis of vehicle interior noise and 6 person's utterances. The horizontal axis is the 1/3 octave band, and the vertical axis is the level. As can be seen from the figure, 1kHz band, 1.25kHz band, 2k
The difference is remarkable in each band of Hz to 4 kHz. Therefore, the filter characteristic of the filter unit 30 is set in advance so as to pass the band.

Then, at the time of voice extraction, the process starts from step s100 of the flowchart shown in FIG. Step s
In 100, first, the microphone 10 detects a voice including background noise. Then, the A / D converter 140
Is converted into a digital signal in step s110, and is passed through the filter unit 30 based on the band level analysis in step s110. As described above, the filter unit 30 is a filter that passes the frequency band based on the result of the band level analysis in advance (band pass). As a result, for example, in the above 1 kHz band where the difference between the audio information and the background noise is remarkable, 1.2
The 5 kHz band and the 2 kHz to 4 kHz bands are passed. That is, the voice is relatively emphasized. Then, in step s120, the threshold determination unit 40 determines the threshold based on the power of the selected band. For example, the maximum level of −20 dB is set as the threshold value.

Next, in step s130, the voice section calculator 50 calculates the voice section. That is, RA in time series
From the detection signal stored in M120, a section equal to or greater than the calculated threshold value is set as a voice section. Next, the frequency component of the background noise is measured immediately before the voice section or in the section after the end of the voice section. This is the next step (step s15
This is because it is used for spectral subtraction in 0). Finally, in step s150, voice is extracted. The voice extraction is obtained by subtracting the frequency component of the background noise obtained in step s140 from the frequency component of the detected sound signal in the voice section by the voice extraction unit 60 (spectral subtraction). Since the difference between the background noise and the voice level is large in this voice section, the voice can be more effectively emphasized by the spectral subtraction calculation. That is, the sound can be accurately extracted even in the presence of noise. Therefore, if this is applied to voice recognition, the voice recognition rate under noise can be improved.

(Second Embodiment) In the voice section determination method of the first embodiment, the filter section 30 based on the band level analysis is used to relatively emphasize the voice, and the maximum power of the emphasized critical band is set. The threshold value is calculated based on the above, and the voice section is determined by the threshold value. However, for example, if the maximum power is projected by entering a critical band having a sudden background noise, the threshold level also becomes large, which may cause an error in the voice section. The present embodiment is intended to accurately extract a voice section even in such a case by emphasizing a voice in a critical band arbitrarily.

FIG. 5 shows a voice extraction device of the present invention. The voice extraction device of the present embodiment is characterized in that the filter unit 30 of the first embodiment is provided with a weighting addition unit 31 for weighting each critical band and adding the weights. In addition, accordingly, the weighted addition of step s111 shown in FIG. 6 is added between step s110 and step s120 of the flowchart of the first embodiment, that is, after passing through the filter unit 30 based on the band level analysis of step s110. And It should be noted that the specific configuration is configured by the computer device 100 shown in FIG. 2, and the weighted addition unit 31 also includes the CPU 1.
10, program stored in ROM120, RAM1
It is composed of 30 and others. After passing through the filter unit 30,
As shown in FIG. 4, for example, 1 kHz, 1.25 kHz, 4
A spectrum in the band centered on kHz is obtained. In this band, the level difference between voice and background noise is remarkable. Therefore, in order to effectively emphasize them, the weighting addition unit 31 of the filter unit 30 is used, that is, in step s111, the bands are weighted and added according to the maximum level of the band. For example, weighting addition is performed on the center of the spectrum with a ratio of 3: 2: 1. In this way, even if sudden noise enters and the threshold value rises, the levels of other critical bands are also increased by weighting and are not affected by the influence. That is, the voice segment can be calculated more accurately, and the voice can be extracted more accurately.

(Third Embodiment) In the voice section determination method of the first embodiment, the critical band in which the voice is emphasized by the filter unit 30 based on the band level analysis is selected, and the maximum power of the critical band is selected. This is an example in which the threshold value is calculated and the voice section is determined based on the threshold value. In addition, in the voice section determination method of the second embodiment, each critical band of the output of the filter unit 30 is weighted and added to enhance the voice more effectively, so that the voice section is more accurate. Was an example of calculating. The present embodiment is an example in which attention is paid to the fact that the size and sharpness in the auditory sense enhance the voice more effectively, and the voice segment is calculated more accurately by using them.

FIG. 7 shows a voice extraction device of the present invention. The voice extraction device of this embodiment has a scale conversion unit 32 for converting a sound pressure level into a perceptual perceptual level in the filter unit 30 of the first embodiment and each critical band after scale conversion in proportion to its frequency. It is characterized in that a weighting addition unit 33 for weighting is provided. Further, as a result, as shown in FIG. 8, steps s110 and s12 of the flowchart of the first embodiment.
During 0, the sound pressure level (dB) is set to the human hearing level (s).
One) is added with the scale conversion in step s112 and the weighted addition in step s113.

Scale conversion is the dB of the frequency spectrum value.
It is a conversion from a value to a sone value. Microphone 10
Although the sound pressure due to is detected by the dB value, human auditory sense does not perceive the sound pressure linearly but perceives the sound pressure with an exponential sone value. When the tone value is S and the sound pressure is P, both are L
There is a relationship of og ₁₀ S = 0.03 (P-40). That is,
The son value has an exponentially increasing relationship with the dB value. That is, the conversion into the sound value means enhancement of the voice band. Therefore, if scale conversion is performed in step s112 using the scale conversion unit 32, the difference between voice and background noise is further emphasized.

Then, in step s113, the weighted addition section 33 is further used to perform weighted addition in which a greater weight is added to the scale-converted value in the higher frequency band. This is an emphasis on the sharpness of the sound. Human voice includes high-frequency components in the pronunciation of consonants, and background noise has few high-frequency components. Therefore,
Sharpness is emphasized by weighting the higher the frequency band. That is, it is possible to further emphasize the acoustic characteristics of the voice. If the threshold is determined and the voice section is calculated from the spectrum converted to the perceptual size and the sharpness is emphasized in the same manner as in the first and second embodiments,
The voice can be extracted more accurately.

FIG. 9 and FIG. 10 show the voice extraction result by the conventional method and the voice extraction result by the method of this embodiment. Figure 9
9A is a voice signal, and FIG. 9B is a noise superimposed signal. Further, A of FIG. 10 is the extraction result of the audio signal according to the present embodiment, and B is the extraction result of the audio according to the conventional method. As shown in FIG. 10, if a filter based on band level analysis is used to extract a critical band in which speech is dominant, and then scale conversion and weighted addition are performed, it can be understood that speech can be extracted more accurately than before. Therefore. If the device of the present embodiment is applied to a vehicle-mounted device such as a navigation device that requires voice recognition, the voice can be accurately and accurately extracted, so that, for example, the voice of the driver can be accurately recognized.
That is, the method and apparatus of this embodiment can contribute to safe driving of the vehicle.

(Modification) The above embodiment is only one example, and various modifications can be considered. For example, the first to third embodiments
In the embodiment, the band level analysis is described as the 1/3 octave band level analysis, but the octave band level analysis can of course be adopted. Further, the voice extraction method and its device of the present invention are effective not only in the vehicle interior noise of the automobile but also in other general noises. It is effective for the voice extraction device under the noise environment in the factory, the indoor noise in the railway, the ship, the airplane, etc.

[Brief description of drawings]

FIG. 1 is a system configuration diagram of a voice extraction device according to a first embodiment of the present invention.

FIG. 2 is a system configuration diagram using a computer device of the voice extraction device according to the first embodiment of the present invention.

FIG. 3 is a flowchart showing the operation of the voice extraction device of the first embodiment of the present invention.

FIG. 4 is a graph showing a 1/3 octave band level analysis result of voice and background noise according to the first exemplary embodiment of the present invention.

FIG. 5 is a system configuration diagram of a voice extraction device according to a second embodiment of the present invention.

FIG. 6 is a part of a flowchart showing the operation of the voice extraction device according to the second embodiment of the present invention.

FIG. 7 is a system configuration diagram of a voice extraction device according to a third embodiment of the present invention.

FIG. 8 is a part of a flowchart showing the operation of the voice extraction device of the third embodiment of the present invention.

FIG. 9 is an explanatory diagram of an audio signal (a) and a background noise superimposed signal (b) including the audio signal.

FIG. 10 is a comparison diagram of voice extraction results by the voice extraction device according to the third embodiment of the present invention and a conventional device.

[Explanation of symbols]

10 ... Microphone 30 ... Filter section 40 ... Threshold value determining unit 50 ... Voice section calculator 60 ... voice extraction unit

Claims (11)

[Claims] 1. A sound extraction method under noise, wherein a filter based on a band level analysis result is applied to a detected sound including background noise, a threshold value is determined based on the filter output, and the filter output is A voice extraction method, characterized in that a section having a level equal to or higher than a threshold is set as a voice section, and voice is extracted from the voice section. 2. The voice extraction method according to claim 1, wherein the band level analysis is an octave band level analysis. 3. The voice extraction method according to claim 1, wherein the action of the filter is weighted addition in which a predetermined weight is added to a predetermined band by the band level analysis and added. . 4. The function of the filter is that the level of each predetermined band obtained by the band level analysis is scale-converted into an audible level, and the converted value is added with a greater weight in a higher frequency band. The voice extraction method according to claim 1 or 2, wherein the weighted addition is performed. 5. The voice extraction according to claim 1, wherein the voice extraction is obtained by subtracting the frequency component of the background noise from the calculated frequency component of the voice section. Audio extraction method. 6. A voice extraction device under noise, comprising voice detection means for detecting voice including background noise, and filter means for applying a filter characteristic based on band level analysis to the detection sound of the voice detection means. A threshold value determining unit that determines a threshold value based on the output of the filter unit; a voice section calculating unit that sets a section of a level equal to or higher than the threshold value as a voice section by the threshold value determining unit; And a voice extraction device. 7. The voice extraction apparatus according to claim 6, wherein the band level analysis used for determining the filter characteristic is an octave band level analysis. 8. The weighted addition for adding a predetermined weight to a predetermined band of the detected sound to which the filter characteristic is applied and adding the weighted addition, according to claim 6 or claim 7. The voice extraction device described. 9. The filter means scale-converts the level of a predetermined band of the detected sound to which the filter characteristic is applied into a audible level, and weights the converted value in a higher frequency band. The voice extraction device according to claim 6 or 7, wherein weighted addition is performed to add and add. 10. The voice extraction means subtracts the frequency component of the background noise from the detected sound frequency component of the voice section calculated by the voice section calculation means. The voice extraction device according to item 1. 11. The background noise is noise inside a vehicle,
11. The voice extraction device according to claim 6, wherein the voice is an instruction to a vehicle-mounted device by a driver.