JP5450298B2 - Voice detection device - Google Patents

Description

  The present invention relates to a voice detection device, and more particularly to a voice detection device that detects a voice component from an input signal in which a voice component and a noise component are mixed.

  As this type of voice detection device, there is one conventionally applied to a voice response switch disclosed in Patent Document 1, for example. According to this prior art, when the input level of the audio signal is equal to or higher than a predetermined value, at least the first formant F1 and the second formant F2 are extracted from the audio signal for a predetermined time. Then, a change in vowels is obtained from the extracted first formant F1 and second formant F2, and this change includes a predetermined start including one of the two vowels “a” and “o”. When the above condition is satisfied, it is determined that the voice signal matches the preset control voice, and the switch element is turned on.

JP-A-61-246800

  As described above, in the above-described prior art, voice detection is performed based on the first formant F1 and the second formant F2. In particular, since the extraction of the first formant F1 is essential, the following problems are caused. There is. That is, in an everyday environment, for example, there are many noises having a large power in a frequency band near 1 kHz including road traffic noise. On the other hand, the frequency band near 1 kHz overlaps with the frequency band of the first formant F1. For this reason, in the conventional technique in which the extraction of the first formant F1 is indispensable, there is a problem that it is easily affected by daily noise such as road traffic noise, and therefore the usable environment is extremely limited. is there. In addition, in the prior art, each formant including the first formant F1 is extracted based on the signal level for each frequency band divided into a plurality of frequencies. Therefore, noise of a certain level or more in each frequency band. When a component exists, the noise component is erroneously detected as a formant. Therefore, not only road traffic noise but also other noises are easily affected.

  Therefore, the object of the present invention is to provide a voice detection device that is less affected by noise such as road traffic noise than in the past and is suitable for detecting human screams and screams, particularly in crime prevention applications. To do.

  In order to achieve this object, the present invention provides a peak emphasizing means for emphasizing a peak of a frequency spectrum of an input signal in an audio detection device that detects the audio component from an input signal in which an audio component and a noise component are mixed. A noise estimation unit that estimates a noise spectrum corresponding to a noise component of the enhanced spectrum after the peak is enhanced by the peak enhancement unit, and a subtraction unit that subtracts the noise spectrum from the enhanced spectrum. is there.

  That is, in the present invention, when the frequency spectrum of the input signal in which the voice component and the noise component are mixed is observed, the frequency spectrum of the input signal includes the respective peaks of the voice component and the noise component, These peaks pay attention to the fact that the speech component and the noise component have different properties. Based on this point of interest, first, the peak of the frequency spectrum of the input signal is emphasized by the peak emphasizing means, that is, it becomes noticeable including the nature of the peak. Of the spectrum after peak enhancement, the noise spectrum corresponding to the noise component is estimated by the noise estimation means. Further, the noise spectrum is subtracted from the peak enhanced spectrum by the subtracting means. As a result, the peak of the noise component is removed from the peaks included in the spectrum after peak enhancement, and only the peak of the speech component, that is, the formant remains. By detecting this formant peak, the detection of the voice component is realized.

  In the present invention, the peak emphasizing means emphasizes the peak by processing the input signal by a predicting means for predicting the current input signal based on the past input signal and an inverse arithmetic expression of the arithmetic expression by the predicting means. And emphasizing execution means. Here, the prediction means predicts a periodic component included in the input signal, that is, a formant. Then, the emphasis execution means emphasizes the formant included in the input signal by processing the input signal by the inverse operation expression of the operation expression by the prediction means. At this time, not only the formant but also the peak of the noise component is emphasized, but the peak of the noise component is removed by the subtracting means as described above.

  The prediction means here can be constituted by, for example, a linear prediction error filter. The enhancement execution means can be configured by an inverse filter of the linear prediction error filter.

In this case, it is desirable that each of the linear prediction error filter as the prediction unit and the inverse filter as the enhancement execution unit is a lattice type digital filter. That is, since the linear prediction error filter and the inverse filter are conjugate with each other, one of them is, for example, FIR (Finite).
When designed with an Impulse Response) filter, the other is necessarily an IIR (Infinite Impulse Response) filter. Here, the IIR filter generally has a defect that it is unstable, in other words, it is difficult to determine stability, but it is known that this defect can be solved if it is a lattice type. For example, when the linear prediction error filter is designed by a lattice type FIR filter and the inverse filter is designed by a lattice type IIR filter, the transversal type is used for the lattice type FIR filter as the linear prediction error filter. Excellent linear prediction performance, such as a higher convergence speed than that of other filters of the beginning, is exhibited. The filter coefficient of the lattice type FIR filter as the linear prediction error filter is directly applied to the filter coefficient of the lattice type IIR filter as the inverse filter, so that the inverse filter is designed. That is, there is an advantage that the design of the inverse filter is easy.

Furthermore, the noise estimation means in the present invention may estimate the noise spectrum by time-averaging the enhanced spectrum. That is, when the noise component exists substantially constantly, the peak of the noise component included in the post-emphasis spectrum is almost unchanged. On the other hand, compared with the peak of the noise component, the peak of the voice component is single (intermittent), that is, changes with time . Therefore, when the emphasized spectrum is time-averaged, only the peak of the noise component included therein remains, and the peak of the speech component is reduced as a whole. Thereby, estimation of the noise spectrum is realized.

  If the input signal includes colored noise, the frequency spectrum of the input signal has a generally inclined characteristic such that the power is approximately inversely proportional to the frequency. If the peak of the frequency spectrum is directly enhanced by the peak emphasizing means, various inconveniences such as a steep slope of the frequency spectrum occur. For this reason, in the present invention, flattening means for flattening the frequency spectrum of the input signal may be further provided. However, the flattening means flattens the frequency spectrum to such an extent that the peak contained in the frequency spectrum of the input signal is maintained without being flattened. The peak emphasizing means emphasizes the peak of the flattened spectrum after being flattened by the flattening means.

  Such flattening means is constituted by a low resolution filter having a low frequency resolution that is insufficient to follow a peak included in the frequency spectrum of the input signal, for example, a digital filter having a relatively small number of taps (filter order). can do.

  As described above, according to the present invention, focusing on the fact that the peak of the speech component and the peak of the noise component included in the frequency spectrum of the input signal are different from each other, the peak of the frequency spectrum of the input signal is By emphasizing and removing the noise component peak from the emphasized peak, only the speech component peak is captured. That is, in an environment where noise such as road traffic noise exists, the influence of the noise can be eliminated. Therefore, accurate voice detection can be realized as compared with the above-described conventional technology that is easily affected by noise. This is particularly suitable for accurately detecting human screams and screams in crime prevention applications.

It is a block diagram which shows schematic structure of one Embodiment of this invention. It is an illustration figure which shows the frequency spectrum of the input signal in the same embodiment. It is a block diagram which shows the specific structure of the linear prediction error filter in the same embodiment. It is a block diagram which shows the structure which combined the reverse filter with the linear prediction error filter. It is a block diagram which shows the more specific structure of the linear prediction error filter. It is a block diagram which shows the more specific structure of the inverse filter in the embodiment. It is an illustration figure for demonstrating the necessity of the planarization circuit in the embodiment. It is an illustration figure which shows the frequency spectrum of the signal after a process by the same planarization circuit. It is an illustration figure which shows the frequency spectrum of the signal after a process by the inverse filter in the embodiment. It is an illustration figure which shows the specific structure of the spectrum subtraction in the same embodiment. It is an illustration figure for demonstrating the operation | movement of the spectrum subtraction. It is an illustration figure which shows one experimental result in the same embodiment. It is an illustration figure for demonstrating operation | movement of the peak determination circuit in the same embodiment. It is an illustration figure which shows another aspect of FIG.

  An embodiment of the present invention will be described with reference to FIGS.

  The voice detection device 10 according to the present embodiment is applied to a security device such as a super security light, for example. Specifically, when a human scream or screaming voice is picked up by a microphone provided in the security device. This is for detecting this. In order to realize this voice detection, the voice detection device 10 has a flattening circuit 20 as a flattening means, as shown in FIG. X (z) (z; variable in z conversion) is input.

  Here, the signal X (z) input to the flattening circuit 20 may include noise components such as road traffic noise in addition to the above-described audio components such as screams and screams. In this case, in the frequency spectrum of the input signal X (z), for example, peaks corresponding to the speech component and the noise component appear as shown in FIG. Among these, the peak α having the lowest frequency f is the peak of the noise component. The other peaks β and γ are the peaks of the voice component, and specifically the peaks of the second formant and the third formant in order from the lowest frequency f. In FIG. 2, the alternate long and short dash line is the average power of the input signal X (z) including the peaks α, β and γ. In this embodiment, the frequency band of discrete Fourier transform (DFT) for obtaining a frequency spectrum is limited to f = 1200 Hz to 3000 Hz. Therefore, in practice, a first formant peak also exists, but this first formant peak does not appear in FIG. 2 because it is outside the frequency band. Further, the input signal X (z) includes colored noise. Therefore, as can be seen from FIG. 2, the frequency spectrum of the input signal X (z) including the colored noise has a generally inclined characteristic such that the power P is approximately inversely proportional to the frequency f.

Returning to FIG. 1, the flattening circuit 20 performs a flattening process to be described later on the input signal X (z). Then, the signal X ′ (z) after the flattening processing includes a linear prediction error filter (LPEF) 30 as a prediction unit and an inverse filter (LPEF ⁻¹ ) 40 of the linear prediction error filter 30. And each of them.

  As will be described later, the linear prediction error filter 30 predicts the current flattened signal X ′ (z) based on the past flattened signal X ′ (z), and the prediction error E (z) is minimized. Adaptive operation to become. In accordance with the adaptive operation of the linear prediction error filter 30, an inverse filter 40 conjugate with the linear prediction error filter 30 is formed, and the flattened signal X '(z) is processed by the inverse filter 40. Thereby, the above-mentioned peaks α, β and γ included in the post-flattened signal X ′ (z) are emphasized. This peak enhancement will also be described in detail later.

  Further, the post-emphasis signal W (z) after the peak enhancement by the inverse filter 40 is input to a spectrum subtraction (SS) 50. The spectral subtraction 50 estimates the peak α of the noise component included in the input enhanced signal W (z), and subtracts the peak α from the enhanced signal W (z). As a result, the noise component peak α is removed, and a post-subtraction signal G (z) in which only the speech component peaks β and γ are left is generated. The post-subtraction signal G (z) is input to the peak determination circuit 60. The operation of the spectral subtraction 50 for generating the post-subtraction signal G (z) will be described later in detail.

  The peak determination circuit 60 determines whether or not the audio signal peaks β and γ are included in the post-subtraction signal G (z). When the peaks β and γ of the audio component are included, for example, an alarm device (not shown) provided in the crime prevention device is operated, or a notification signal is transmitted to a predetermined disaster prevention center. The point of the peak determination process by the peak determination circuit 60 will also be described in detail later.

  As described above, according to the speech detection device 10 of the present embodiment, speech detection is realized by the configuration including the flattening circuit 20, the linear prediction error filter 30, the inverse filter 40, the spectral subtraction 50, and the peak determination circuit 60. However, these will be described more specifically below.

  First, the necessity of the linear prediction error filter 30 will be described. That is, the linear prediction error filter 30 predicts the current flattened signal X ′ (z) based on the past flattened signal X ′ (z) as described above. The predictable component is canceled and only the unpredictable component is output as the prediction error E (z). Such a linear prediction error filter 30 includes, for example, a delay element 302 for one sample, an FIR type adaptive filter 304, and an adder 306 as shown in FIG.

  In the configuration shown in FIG. 3, for example, it is assumed that the input signal X (z) described above is directly input instead of the flattened signal X ′ (z). In this case, the input signal X (z) is delayed by the delay element 302 and then processed by the adaptive filter 304. The signal U (z) after processing by the adaptive filter 304 is input to the adder 306. The adder 306 also receives the input signal X (z), and the adder 306 subtracts the signal U (z) processed by the adaptive filter 304 from the input signal X (z). The signal E (z) after this subtraction is output as a prediction error (corresponding to the input signal X (z)), and the adaptive filter 304 performs an adaptive operation so that the prediction error E (z) is minimized.

  Here, assuming that the transfer function of the adaptive filter 304 is H (z), the signal U (z) processed by the adaptive filter 304 is expressed by the following equation (1).

  If the transfer function of the entire linear prediction error filter 30 including the transfer function H (z) of the adaptive filter 304 is L (z), the transfer function L (z) is expressed by the following equation (2).

Furthermore, when the number of taps of the adaptive filter 304 is N, the transfer function H (z) of the adaptive filter 304 is expressed by the following formula 3. In Equation 3, h _n is an n-th tap filter coefficient.

  Then, after the expression of Equation 3 is rewritten for convenience, when Equation 3 is substituted into Equation 2, the transfer function L (z) of the entire linear prediction error filter 30 is as shown in Equation 4 below. It is expressed in

  On the other hand, voice, particularly voiced sound P (z), is expressed as in the following formula 5. In Equation 5, A (z) is a transfer function (resonance characteristic) of the entire vocal tract of the speaker who emits the voiced sound, and B (z) is a characteristic of vocal cord vibration of the speaker. .

  The characteristics of the voiced sound P (z) represented by this equation 5, particularly the characteristics of the vowel formant, depend on the transfer function A (z) of the vocal tract. Therefore, when the transfer function A (z) of the vocal tract is expressed by, for example, a finite-length all-pole model, the transfer function A (z) is expressed as the following Expression 6. In Equation 6, M is the number of taps of the all-pole model.

  Therefore, assuming that only the voiced sound P (z) is input as the input signal X (z), the prediction error E (z) is expressed by the following Expression 7.

  In addition, the adaptive filter 304 has a sufficient number of taps N to represent the vocal tract transfer function A (z), and the prediction error E (z) represented by Equation 7 is minimized. Assuming that the adaptive operation is performed, the following formula 8 is established in the formula 7.

  This means that the inverse of the vocal tract transfer function A (z) is predicted by the linear prediction error filter 30 including the adaptive filter 304.

Therefore, the input signal X (z) is processed by the inverse filter 40 of the linear prediction error filter 30, that is, the transfer function L ⁻¹ (z) of the inverse filter 40 is multiplied to the input signal X (z). Thus, the formants included in the input signal X (z) are emphasized. The transfer function L ⁻¹ (z) of the inverse filter 40 is expressed by the following formula 9.

  As shown in FIG. 4, the inverse filter 40 has a dependent filter 402 to which the transfer function H (z) of the adaptive filter 304 in the linear prediction error filter 30 is copied, and a signal processed by the dependent filter 402. An adder 404 that adds to the input signal X (z), and a delay element 406 for one sample that delays the signal W (z) added by the adder 404 and inputs the delayed signal to the dependent filter 402. . Then, the signal W (z) after the addition by the adder 404 is output as a signal after processing by the inverse filter 40, that is, a signal after enhancement. However, since the configuration of the inverse filter 40 is a so-called IIR type, there is a concern that its operation becomes unstable. Therefore, in order to eliminate this drawback, a lattice type filter is used as the inverse filter 40. In accordance with this, the linear prediction error filter 30 is also of a lattice type.

Specifically, first, as shown in FIG. 5, the linear prediction error filter 30 includes a delay side (backward prediction side) adder 310 to which an output of the delay element 302 is input, and an input signal X (z). And another adder 312 on the non-delay side (forward prediction side) that is directly input. The output of the delay element 302 is also input to the multiplier 314, and the output of the multiplier 314 is input to the non-delay side adder 312. The non-delay side adder 312 subtracts the output of the multiplier 314 from the input signal X (z) and inputs the signal after the subtraction to the adder 312a at the next stage. In addition, the input signal X (z) is also input to another multiplier 316, and the output of this multiplier 316 is input to the adder 310 on the delay side. The adder 310 on the delay side subtracts the output of the multiplier 316 from the output of the delay element 302 and inputs the signal after the subtraction to the delay element 302a in the next stage. The delay element 302a in the next stage forms a configuration similar to that in the previous stage together with the two adders 310a and 312a and the two multipliers 314a and 316a. This configuration is cascaded over M stages, and the final M-th non-delay side adder 312b serves as the adder 306 shown in FIGS. That is, the output of the M-th non-delay side adder 312b is the prediction error E (z). It should be noted that the same filter coefficient (reflection coefficient) δ ₁ is set in the two multipliers 314 and 316 constituting the first stage. The same applies to the other stages. Since the calculation methods of these filter coefficients δ ₁ , δ ₂ ,..., Δ _M are well known, detailed description thereof is omitted here.

On the other hand, as shown in FIG. 6, the inverse filter 40 includes an adder 410 on the feedback side (side corresponding to backward prediction) to which the output of the delay element 406 is input, and the order in which the post-emphasis signal W (z) is output. And another adder 412 on the direction side (side corresponding to the forward prediction). The output of the delay element 406 is also input to the multiplier 414, and the output of the multiplier 414 is input to the adder 412 on the forward direction side. The forward side adder 412 adds the output of the multiplier 414 to the signal input via the previous stage adder 412a, and outputs the signal after the addition as an enhanced signal W (z). In addition, the emphasized signal W (z) is also input to another multiplier 416, and the output of this multiplier 416 is input to the adder 410 on the feedback side. The adder 410 on the feedback side subtracts the output of the multiplier 416 from the output of the delay element 406, and inputs the signal after the subtraction to the delay element 406a of the next stage. The delay element 406a in the next stage has the same configuration as that of the delay element 406 in the previous stage together with the two adders 410 and 412 and the two multipliers 414 and 416, and the two adders 410a and 412a and the two multipliers. Build with 414a and 416a. This configuration is cascaded over M stages, and the M-th forward-direction adder 412b serves as the adder 404 shown in FIG. That is, the input signal X (z) is input to the M-th forward direction adder 412b. The first stage multipliers 414 and 416 are set with the filter coefficient δ ₁ of the first stage multipliers 314 and 316 of the linear prediction error filter 30 shown in FIG. The same applies to the other stages. Thereby, the inverse filter 40 of the linear prediction error filter 30 is configured.

Such a lattice-type inverse filter 40 is known to exhibit a stable operation although it is an IIR type. In other words, stability determination is easy, and specifically, it is known that if each of the filter coefficients δ ₁ , δ ₂ ,..., Δ _M is less than ± 1, the operation of the inverse filter 40 is stabilized. It has been. Further, the linear prediction error filter 30 is also of a lattice type, so that excellent linear prediction performance such as a higher convergence speed than other filters such as a transversal type can be obtained. Moreover, the inverse filter 40 is realized by applying the filter coefficients δ ₁ , δ ₂ ,..., Δ _{M of the} linear prediction error filter 30 to the inverse filter 40 as they are.

  In this way, the input signal X (z) is processed by the inverse filter 40 of the linear prediction error filter 30 to emphasize the formant included in the input signal X (z). In this case, that is, the input signal When peak enhancement is directly applied to X (z), the following inconvenience occurs.

  That is, the input signal X (z) exhibits a frequency spectrum that is totally inclined as shown in FIG. 2 described above, and if the input signal X (z) is directly subjected to peak enhancement, As shown by the solid curve in FIG. 7, the slope of the frequency spectrum after peak emphasis becomes steep. Note that the dashed curve in the figure is the frequency spectrum of the input signal X (z) before peak enhancement, that is, the same as the solid curve shown in FIG. In addition, since the slope of the frequency spectrum becomes steep in this way, there is a possibility that the power of portions other than the peaks α, β, and γ is particularly larger than the formant peaks β and γ. It becomes difficult for the circuit 60 to determine the peaks β and γ of the formant. Further, in this peak emphasis, not only the formant peaks β and γ but also the noise component peak α is emphasized, so that the noise component peak α is particularly excessive, that is, the range overflows.

  In order to avoid this inconvenience, a planarization circuit 20 is provided. That is, as the input signal X (z) is processed by the flattening circuit 20, the sharpness of each peak α, β, and γ is maintained as shown by the solid curve in FIG. To the extent, the entire frequency spectrum of the input signal X (z) is flattened and the tilt is corrected. Such a flattening circuit 20 is realized by a filter having a low frequency resolution that cannot follow the peaks α, β, and γ. For example, the flattening circuit 20 has the same configuration as that of the linear prediction error filter 30 and is linear. This is realized by a filter having fewer taps than the prediction error filter 30. Of course, the planarization circuit 20 may be realized by other configurations.

  Then, the flattened signal X ′ (z) that has been lightly flattened by the flattening circuit 20 is input to each of the linear prediction error filter 30 and the inverse filter 40. Accordingly, unlike the solid line curve shown in FIG. 7, as shown by the solid line curve in FIG. 9, a post-emphasis signal W (z) in which the peaks α, β and γ are moderately emphasized is obtained. It is done. 9 is a frequency spectrum of the flattened signal X ′ (z) before peak enhancement, that is, the same as that indicated by the solid curve in FIG.

  Furthermore, the post-emphasis signal W (z) is input to the spectral subtraction 50. The spectral subtraction 50 is configured as shown in FIG. That is, the spectral subtraction 50 has a moving average circuit 502, and the enhanced signal W (z) is input to the moving average circuit 502. The moving average circuit 502 performs a moving average (time average) over the input emphasized signal W (z) over a predetermined period of Ta, for example, Ta = 5 seconds. Then, the averaged signal Wa (z) after the moving average by the moving average circuit 502 is input to the multiplier 504, where it is multiplied by a certain coefficient ε. The value of the coefficient ε is appropriately determined according to the situation, and for example, ε = 1.5. The averaged signal Wa ′ (z) after multiplication by the multiplier 504 is input to the adder 506.

  Further, the spectrum subtraction 50 has a delay circuit 508, and the enhanced signal W (z) is also input to this delay circuit 508. The delay circuit 508 delays the input post-emphasis signal W (z) by a certain period of Td. The delay time Td by the delay circuit 508 is ½ of the moving average time Ta by the moving average circuit 502, that is, Td = 2.5 seconds. The signal Wd delayed by the delay circuit 508 is also input to the adder 506.

  The adder 506 generates the above-described subtracted signal G (z) by subtracting the averaged signal Wa ′ (z) after the multiplication by the multiplier 504 from the delayed signal Wd (z) by the delay circuit 508. Here, the post-delay signal Wd (z) is the post-emphasis signal W (z) at a time before the current time by the delay time Td, and has a frequency spectrum as shown in FIG. 11A, for example. On the other hand, the averaged signal Wa ′ (z) is apparently the signal W after the enhancement over the average time Ta of a total of 5 seconds, 2.5 seconds before and after it, centering around the time that is delayed by the delay time Td by the delay circuit 508. (Z) is a moving average and the level is multiplied by ε. In particular, in the averaging time of Ta = 5 seconds, the peak α of the standing noise component is almost unchanged. On the other hand, since speech components (long vowel components) such as screams and screams are sporadic, their peaks β and γ vary. As a result, the averaged signal Wa ′ (z) has a frequency spectrum in which only the noise component peak α remains and the speech component peaks β and γ are greatly reduced, as shown in FIG. 11B. Therefore, the subtracted signal G (z) becomes a frequency spectrum in which the noise component peak α is removed and only the speech component peaks β and γ are left, as shown in FIG. 11C. That is, only the peaks β and γ of the speech component are extracted. Therefore, voice detection is realized by capturing the peaks β and γ of the voice component.

  Even if the noise component peak α is substantially unchanged over the averaging time Ta by the moving average circuit 502 as described above, the noise component peak α is also somewhat reduced by the moving average processing over the averaging time Ta. The In order to compensate for this reduction, the above-described multiplier 504 is provided. That is, the coefficient ε of the multiplier 504 is set so that the noise component peak α is appropriately removed from the subtracted signal G (z).

  Here we report the actual experimental results.

  That is, using an evaluation sound source (not shown), pink noise and a 1400 Hz sine wave are generated as noise. Then, a voice of “Help me” is generated as a voice. The SNR (Signal-Noise Ratio) between these voices and noise is about -15 dB. The sampling frequency for obtaining the input signal X (z) including these sounds and noise is 32 kHz, and the above-mentioned discrete Fourier transform score is 800. Further, the tap length of the linear prediction error filter constituting the flattening circuit 20 is 40, and the step size is 0.00625. The tap length of the linear prediction error filter 30 is 800, and the step size is 0.25. The inverse filter 40 also has the same tap number and step size as the linear prediction error filter 30. The above-described averaging time Ta of the spectral subtraction 50 is Ta = 5 seconds, and the coefficient ε is ε = 1.5.

  The input signal X (z) under such conditions has a frequency spectrum as shown in FIG. Then, the flattened signal X ′ (z) after the input signal X (z) is flattened has a frequency spectrum as shown in FIG. As can be seen from these comparisons, in the frequency spectrum of the input signal X (z) in FIG. 12A, the peaks β and γ of the speech component are not particularly noticeable compared to the peak α of the noise component. In the frequency spectrum of the post-flattened signal X ′ (z) of 12 (b), the speech component peaks β and γ become conspicuous as much as the noise component peak α.

  Further, the post-emphasis signal W (z) after the post-flattening signal X ′ (z) is subjected to peak emphasis has a frequency spectrum as shown in FIG. Then, the averaged signal Wa (z) after the post-emphasis signal W (z) is moving averaged over the above-described averaging time Ta has a frequency spectrum as shown in FIG. As can be seen from these comparisons, in the frequency spectrum of the post-emphasis signal W (z) in FIG. 12 (c), although the peaks β and γ of the speech component are sufficiently large, the averaged signal Wa in FIG. 12 (d). In the frequency spectrum of (z), the peaks β and γ of the sound component are extremely reduced. On the other hand, although the peak α of the noise component is also reduced, the degree of reduction is small.

  Then, as described above, the post-emphasis signal W (z) is delayed by a delay time of Td to generate the post-delay signal Wd (z), and the average signal Wa (z) is multiplied by the coefficient ε. In this way, an averaged signal Wa ′ (z) after multiplication is generated, and a subtracted signal G (z) is generated by subtracting these signals. The post-subtraction signal G (z) has a frequency spectrum as shown in FIG. As is apparent from the frequency spectrum of the subtracted signal G (z) in FIG. 12 (e), the noise component peak α is removed, and only the speech component peaks β and γ remain.

  As described above, the effectiveness of the present embodiment was also confirmed by actual experiments.

  The post-subtraction signal G (z) is input to the peak determination circuit 60, where it is determined whether or not there are peaks β and γ of the speech component. The frequency spectrum of the post-subtraction signal G (z) is spectral subtraction. As a result of the processing by 50, since it also has a negative number, the presence or absence of the peaks β and γ of the speech component cannot be determined by absolute evaluation. That is, a certain threshold value is set, and whether or not there are peaks β and γ of the sound component cannot be determined based on whether or not there are sound component peaks β and γ larger than the threshold value. Therefore, the peak determination circuit 60 performs peak determination by the following relative evaluation.

  That is, it is assumed that a part of the frequency spectrum of the subtracted signal G (z) has characteristics as shown in FIG. Here, when the component value g [i] in an arbitrary frequency bin i satisfies the conditional expression expressed by the following equation 10, the component value g [i] is the peak β or γ of the audio component. Is determined.

  Based on the conditional expression expressed by this equation 10, that is, based on the component values g [i−2] to g [i + 2] in five consecutive frequency bins [i−2] to [i + 2], By determining whether or not the component value g [i] in the frequency bin i is the peak β or γ of the sound component, even if it is the frequency spectrum of the subtraction signal G (z) having a negative number, the sound The presence or absence of signal peaks β and γ can be accurately determined. Also, even if the peaks β and γ of the audio signal become unclear due to fluctuations, for example, as shown in FIG. 14, the presence or absence of them can be accurately determined. The numerical value “4” included in Q5 to Q8 in Equation 10 is a so-called experience value, and a different value may be set depending on the situation.

  As described above, according to the present embodiment, the frequency band of the discrete Fourier transform for obtaining the frequency spectrum of the input signal X (z) is limited to f = 1200 Hz to 3000 Hz, that is, the first formant is a voice. It is intentionally removed from the detection target. Therefore, even in an environment where there is a noise such as road traffic noise having a large power in the frequency band near 1 kHz that overlaps the frequency band of the first formant, voice detection is realized while eliminating the influence of the noise. Can do. In addition, according to the present embodiment, paying attention to the difference in properties between the noise component peak α and the speech component peaks β and γ included in the frequency spectrum of the input signal X (z), these peaks α , Β and γ are emphasized, and furthermore, only the noise component peak α is removed from the frequency spectrum after the peak enhancement, and the speech component peaks β and γ are captured, thereby realizing speech detection. Therefore, not only road traffic noise but also the influence of other noises can be eliminated. As a result, more accurate voice detection can be realized than the above-described conventional technique that is easily affected by the noise. This is particularly suitable for accurately detecting human screams and screams in crime prevention applications.

Such a voice detection device 10 is, for example, a CPU (Central Processing).
Unit) or a combination of the CPU and a DSP (Digital Signal Processor). For these CPUs and DSPs, it is possible to detect voices in the manner described above with a relatively small amount of processing, so it is possible to use relatively inexpensive CPUs and DSPs, especially as DSPs. Can be a fixed-point type. Furthermore, the sampling frequency for obtaining the input signal X (z) can be reduced from the above-described 32 kHz, and can be down-sampled to 12 kHz, for example. This also greatly contributes to cost reduction of the CPU and DSP.

  In this embodiment, although the case where this invention is applied to a crime prevention use was demonstrated, it is not restricted to this. That is, the present invention can be applied to any application where it is necessary to detect only the audio component from the input signal X (z) in which the audio component and the noise component are mixed.

  Further, by limiting the frequency band of the discrete Fourier transform, as a result, the first formant is excluded from the target of speech detection. However, the present invention is not limited to this. That is, another means for removing the first formant from the target of voice detection, for example, a frequency limiting means such as a low-pass filter may be employed.

  Furthermore, although the lattice-type FIR filter shown in FIG. 5 is adopted as the linear prediction error filter 30, a filter having another configuration such as a transversal type may be adopted. In this case, it is important to use a filter having a configuration conjugate with the linear prediction error filter 30 for the inverse filter. However, as described above, a tremendous advantage can be obtained by adopting a lattice-type filter.

  And although the thing of the structure shown in FIG. 10 was employ | adopted as the spectrum subtraction 50, it is not restricted to this. In particular, devices other than the moving average circuit 502 may be employed as means for estimating the noise component peak α.

DESCRIPTION OF SYMBOLS 10 Speech detector 20 Flattening circuit 30 Linear prediction error filter 40 Inverse filter 50 Spectral subtraction

Claims (5)

In a voice detection device that detects a voice component from an input signal in which a voice component and a noise component are mixed,
Peak emphasizing means for emphasizing the peak of the frequency spectrum of the input signal;
Noise estimation means for estimating a noise spectrum corresponding to the noise component in the enhanced spectrum after the peak is enhanced by the peak enhancement means;
Subtracting means for subtracting the noise spectrum from the enhanced spectrum;
Peak determination means for determining whether or not the peak is included in the signal after subtraction after the noise spectrum is subtracted from the spectrum after enhancement by the subtraction means;
Equipped with,
The peak enhancement means is configured to set a linear prediction error filter that predicts the current input signal based on the past input signal, and an inverse transfer function that is the reciprocal of the transfer function of the linear prediction error filter. Including an inverse filter that enhances the peak by processing,
A voice detection device characterized by the above. Each of the linear prediction error filter and the inverse filter is a lattice type.
The voice detection device according to claim 1 . The noise estimation means estimates the noise spectrum by time averaging the emphasized spectrum.
Speech detection apparatus according to claim 1 or 2. In a voice detection device that detects a voice component from an input signal in which a voice component and a noise component are mixed,
Peak emphasizing means for emphasizing the peak of the frequency spectrum of the input signal;
Noise estimation means for estimating a noise spectrum corresponding to the noise component in the enhanced spectrum after the peak is enhanced by the peak enhancement means;
Subtracting means for subtracting the noise spectrum from the enhanced spectrum;
Comprising
Flattening means for flattening the frequency spectrum of the input signal to such an extent that the peak is maintained;
The peak enhancing means emphasizes the peak of the flattened spectrum after being flattened by the flattening means ;
A voice detection device characterized by the above . The flattening means includes a low resolution filter having a low frequency resolution that is insufficient to follow the peak;
The voice detection device according to claim 4 .

Images