JP2002091470A - Voice section detecting device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice section detecting device capable of surely detecting a voice section even for a word including a double consonant, or a word including successive sound of the 'Sa' column or 'Ha' column of the kana syllabary. SOLUTION: A voice signal detected by a microphone 21 is amplified by a line amplifier 22 and digitized by an analog-digital converter 23, and then stored in a storage part 24. The stored voice signal is fetched into a pitch detection part 25, and a voice pitch is extracted by time-base processing. Based on this voice pitch, a gate signal generation part 26 controls a gate signal, and based on this gate signal, a voice section signal generation part 28 controls a voice section signal. A word is extracted by segmenting the voice signal stored in the storage part by the voice section signal.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice section detection device, and more particularly to a voice section capable of reliably detecting a voice section even for a word including a prompt sound or a word having a continuous line of sound. The present invention relates to a section detection device.

[0002]

2. Description of the Related Art In speech recognition, it is necessary to extract a speech section to be subjected to speech recognition from a time-series signal taken from a microphone. A method has been proposed in which a period in which the short-time power of a voice is equal to or greater than a predetermined threshold is set as a voice section, but when the purpose is to recognize many types of words from the voice of an unspecified speaker, It was difficult to secure sufficient accuracy.

The present applicant has already proposed a pitch period extracting apparatus and method capable of detecting a pitch, which is the pitch of a voice, from a speech signal in the time domain with high accuracy (Japanese Patent Laid-Open No. 9-50297). Publication) can determine the voice section based on the pitch period.

[0004]

However, the word A (for example, "Nitrogen") containing a prompting sound in the word, the word B (for example, "Sushi shop") having a continuous line, or the word C (for a sushi bar) having a continuous line. For example, "dermatology")
It is not possible to avoid the possibility of erroneous detection in which all sounds constituting a word are not detected as one continuous voice section.

[0005] Fig. 1 shows a conventional speech section detection result based on a pitch period, wherein (a) targets "word A", (b) targets "word B", and (c) targets "word C". Shows a case where a voice section is detected. In each case, the upper row contains audio signals,
The lower part shows a voice section. As can be seen from this figure,
In the case of "word A", the front sound ("nitrogen"
Although "Chi" is detected in the voice section, the sound at the rear ("So" for "Nitrogen") is not detected.

[0006] In the case of "sushi restaurant", the voice section is interrupted between "sushi" and "ya", and in the case of "dermatology", between "hifu" and "ka", one voice is interrupted. Not detected as a section. The following can be considered as causes of this erroneous detection. A: The fricative sound "so" following the prompting sound of the word A "tsu" and the fricative sound "shi" following the sound of the word B "su" are not only low in level, but also difficult to distinguish from noise, and are therefore pitches. It is difficult to detect the period itself.

B: If there is no audible or noise portion preceding the word and the pitch is low, the pitch period cannot be detected. C: In the case of the word C, the cue sound ("hifu" for "skin") and the sound following it ("ka" for "skin")
The silence period between is long. D: Noise during silence period The present invention has been made in view of the above problem, and it is possible to reliably detect a voice section even for a word including a prompting sound or a word having a continuous line of sound. It is an object of the present invention to provide a simple voice section detection device.

[0008]

According to a first aspect of the present invention, there is provided a speech section detecting apparatus, comprising: preprocessing means for removing noise contained in a speech signal; and a speech signal from which noise has been removed by the preprocessing means. Voice pitch extracting means for extracting a voice pitch signal from a voice signal, a gate signal generating means for generating a gate signal based on the voice pitch extracted by the voice pitch extracting means, and a voice section signal based on the gate signal generating means. Voice section signal generation means.

In the present invention, the gate signal is controlled based on the voice pitch extracted from the voice signal, and the voice section signal is controlled based on the gate signal. In the voice section detection device according to the second invention, the voice sectioning the voice signal from which the noise has been removed by the preprocessing means on the basis of the voice section signal generated by the voice section signal generating means into a plurality of voice signals. The apparatus further includes signal dividing means.

In the present invention, the audio signal is divided into a plurality of sections by the audio section signal. In the voice segment detection device according to the third invention, the voice pitch extracting means removes a voice signal having a predetermined amplitude or less from the voice signal from which noise has been removed by the preprocessing means. , An equalizing means for adjusting the amplitude of the audio signal subjected to the subtraction processing to the substantially constant amplitude, and a positive peak and the positive peak from the audio signal adjusted to a substantially constant amplitude by the equalizing means. Negative peak emphasizing means for detecting a negative peak following the above and subtracting the positive peak from the negative peak to generate an audio signal in which the negative peak is emphasized, and detecting and processing the audio signal in which the negative peak is emphasized by the negative peak emphasizing means Differential processing means for differentiating the processed signal.

In the present invention, the voice pitch is extracted by processing in the time domain. In the speech section detection device according to the fourth invention, the subtraction processing means calculates the positive envelope and the negative envelope of the audio signal from which the noise has been removed by the preprocessing means, and calculates the positive envelope and the negative envelope. An envelope difference calculating means for calculating an envelope difference which is a difference between side envelopes; and a subtraction for calculating a subtraction processing threshold by multiplying the envelope difference calculated by the envelope difference calculating means by a predetermined multiplication factor. Processing threshold value calculating means, and subtraction processing for subtracting the subtraction processing threshold value from the audio signal amplitude when the amplitude of the audio signal from which noise has been removed by the preprocessing means is equal to or greater than the subtraction processing threshold value calculated by the subtraction processing threshold value calculating means. Threshold subtraction means.

In the present invention, a predetermined multiple of the envelope difference of the audio signal is set as the subtraction processing threshold. In the speech section detection device according to a fifth aspect, the subtraction processing means has an amplitude of the audio signal from which the noise has been removed by the preprocessing means is smaller than the subtraction processing threshold value calculated by the subtraction processing threshold value calculation means. In such a case, the apparatus further comprises zero setting means for setting the amplitude of the audio signal to zero.

In the present invention, when the amplitude of the audio signal is equal to or smaller than the subtraction threshold, the amplitude of the audio signal is set to zero. In the voice section detection device according to the sixth invention, the uniform amplitude means calculates the positive envelope and the negative envelope of the voice signal from which the noise has been removed by the preprocessing means, and calculates the positive envelope and the negative envelope. An envelope difference calculating means for calculating an envelope difference which is a difference between side envelopes; and a maximum envelope difference holding means for holding a maximum envelope difference among the envelope differences calculated before and now by the envelope difference calculating means. Means, and a matching amplitude gain calculating means for calculating a matching amplitude gain by dividing the maximum envelope difference held by the maximum envelope difference holding means by the current envelope difference.

In the present invention, the uniform amplitude gain is determined based on the envelope difference of the audio signal. In the speech section detection device according to the seventh invention, the matching amplitude means is configured to use the matching amplitude gain as a unit when the matching amplitude gain calculated by the matching amplitude gain calculating means is equal to or greater than a predetermined threshold value. The apparatus further includes unit gain setting means for setting a gain. In the present invention, when the matching amplitude gain is equal to or more than a predetermined threshold, the matching amplitude gain is set to a unit gain.

In the voice section detection device according to an eighth aspect of the present invention, the gate signal generation means determines an average value of a predetermined number of continuous voice pitches extracted by the voice pitch extraction means. A gate signal opening means for opening a gate signal when a gate opening threshold is exceeded is provided. According to the present invention, the gate signal is opened when the average value of the predetermined number of voice pitches is equal to or greater than the gate opening threshold.

In the speech section detection device according to the ninth invention, the gate signal generation means is configured such that, when the gate signal is once opened by the gate signal opening means, the continuous pitch extracted by the voice pitch extraction means. If the average value of the predetermined number of voice pitches is equal to or greater than a predetermined gate closing threshold which is smaller than the gate opening threshold, the apparatus further comprises a gate signal open maintaining means for maintaining the gate signal in an open state.

According to the present invention, the gate signal is maintained in the open state if the average value of a predetermined number of continuous voice pitches is equal to or greater than the gate closing threshold. Tenth
In the voice section detection device according to the invention, the gate signal generation means, when the average value of a predetermined number of continuous voice pitches extracted by the voice pitch extraction means is less than the gate open threshold The apparatus further comprises a gate signal closing means for closing the gate signal.

In the present invention, the gate signal is closed when the average value of the voice pitch becomes smaller than the gate closing threshold. In the voice segment detection device according to the eleventh invention,
A first predetermined time period measuring means for measuring a first predetermined time period from a point in time when the gate signal generated by the gate signal generation means is opened; a first predetermined time period; There is provided a voice section signal opening means for opening the voice section signal retroactively for a predetermined second predetermined period from the time when the timing of the first predetermined period by the timing means is completed.

In the present invention, when the gate signal is continuously open for the first predetermined period, the voice section signal is opened retroactively for a second predetermined period from the time when the first predetermined period has elapsed. You. In the voice segment detection device according to the twelfth invention,
A third predetermined time period measuring means for measuring a predetermined third predetermined period from the time when the gate signal generated by the second gate signal generating means is closed; The apparatus further includes voice section signal closing means for closing the voice section signal when the timing of the third predetermined time period by the predetermined time counting means is completed.

In the present invention, the voice section signal is closed after a lapse of a third predetermined period from the time when the gate signal is closed. In the voice section detection device according to the thirteenth aspect, the voice section signal generation means may switch the voice section signal opening means from the voice section signal opening means before the time measurement of the third predetermined time period by the third predetermined time counting means does not end. And a voice section signal open state maintaining means for maintaining the voice section signal in an open state when the voice section signal is opened retroactively for a predetermined period of time.

In the present invention, the voice section signal is the third
When the predetermined period and the second predetermined period overlap, the open state is maintained.

[0022]

FIG. 2 is a functional block diagram of a voice section detection apparatus according to the present invention. A voice signal converted into an electric signal by a microphone 21 is amplified by a line amplifier 22 and then analog / digital. The audio signal is sampled by the conversion unit 23 at every predetermined sampling time Δt, converted into a digital signal, and stored in the storage unit 24.

The gate signal generator 26 is provided with a pitch detector 25
A gate signal is generated on the basis of the pitch detected in step (1), and the voice section signal generation unit 27 generates a voice section signal based on the gate signal generated by the gate signal generation unit 26. The word extracting section 28 processes the digital signal stored in the storage section 24 based on the voice section signal generated by the voice section signal generating section 27 to extract and output words included in the voice section.

In this embodiment, the analog / digital conversion unit 23, the storage unit 24, the pitch detection unit 25,
The gate signal generation unit 26, the voice section signal generation unit 27, and the word extraction unit 28 are configured by, for example, a personal computer, and the pitch detection unit 25, the first gate generation unit 26, the second gate generation unit 27, and the word extraction unit 28 It is configured as hardware.

FIG. 3 is a flowchart of a voice sampling routine executed by the analog / digital conversion unit 23 and the storage unit 24, and is executed as an interrupt process every sampling time Δt. First, in step 30, the audio signal V sampled by the analog / digital converter 23 is fetched, and in step 31, preprocessing is performed on the audio signal V. The details of the preprocessing will be described later.

At step 32, an index i indicating the storage order of the storage unit 23 is set to "1". At steps 33 to 35, the audio signal X already stored in the storage unit 23 is set.
(I) is forwarded by the following processing. X (i + 1) ← X (i) When the forward feeding is completed, the currently read audio signal V is stored in the earliest address X (1) of the storage unit 23, and this routine ends.

FIG. 4 is a detailed flowchart of the pre-processing routine executed in step 31.
Executes high frequency noise removal processing on the digital signal. For this processing, for example, a low-pass filter having a cutoff frequency of 4 KHz and a cutoff characteristic of 18 db / oct is used. In step 311, low frequency noise removal processing is performed on the digital signal from which high frequency noise has been removed. For this processing, for example, a high-pass filter having a cutoff frequency of 300 Hz and a cutoff characteristic of 18 db / oct is used.

In the above embodiment, the high-frequency noise removal processing and the low-frequency noise removal processing are performed by software, but a hardware filter may be incorporated in the line amplifier 22. FIG. 5 is a detailed flowchart of the pitch detection routine executed by the pitch detection section 25. In step 50, the audio signal X (i) stored in the storage section 23 is read.

Then, a subtraction process is executed in step 51, an AGC process is executed in step 52, and a peak detection process is executed in step 53. Further, an extreme value detection clamp process is executed in step 54, and a pitch cycle detection process is executed in step 55, and this routine is ended. Steps 51-55
Will be described in detail below.

FIG. 6 shows a step 51 of the pitch detection routine.
Is a flowchart of a subtraction processing routine executed in step A. In the AGC processing for equalizing the amplitude of the audio signal, even a small noise component is removed from the predetermined amplitude or less in order to prevent the noise component from being AGC processed and amplified. The purpose is to do. First, the envelope value difference ΔE is calculated in step 51a, which will be described in detail with reference to FIG.

In step 51b, it is determined whether or not the envelope value difference ΔE is less than a predetermined threshold value r for amplitude removal. When the determination is affirmative, that is, when the envelope value difference ΔE is less than the threshold value r, the process proceeds to step 51c. The audio signal X (i) is set to "0", and the routine proceeds to step 51d. When a negative determination is made in step 51b, that is, when the envelope value difference ΔE is equal to or larger than the threshold value r, the process directly proceeds to step 51d.

In step 51d, the current positive-side envelope value Ep is calculated.
Is larger than the previous positive side envelope value Ebp. When an affirmative determination is made in step 51d, that is, when the current positive envelope value Ep is larger than the previous positive envelope value Ebp and the positive envelope value is increasing, the index s is set to "1" in step 51e. And the process proceeds to step 51g.

[0033] Conversely, when a negative determination is made in step 51d, i.e. positive envelope value of this positive envelope value E _p the previous E _pb
If it is smaller and the positive side envelope value is decreasing, the index s is set to "0" in step 51f, and the flow advances to step 51g. Step previous value s _b of the index s in 51g is "1" and this index s is "0", i.e., positive-side peak is detected whether the detected.

When an affirmative determination is made in step 51g, that is, when a positive peak is detected, a threshold value bc for the subtraction processing is calculated in step 51h using the following equation.
Go to 1i. bc ← α * ΔE Here, α is a predetermined value, and can be set to a fixed value “0.05” when the voice section detection device according to the present invention is used in a vehicle cabin.

Conversely, when a negative determination is made in step 51g, that is, when the positive peak is not detected, the flow directly proceeds to step 51i. In step 51i, the audio signal X
It is determined whether (i) is equal to or larger than the threshold value bc of the subtraction process, that is, whether the amplitude of the audio signal X (i) is large. Step 5
1i, that is, when the amplitude of the audio signal X (i) is equal to or larger than the threshold bc, the value obtained by subtracting the threshold bc of the subtraction processing from the audio signal X (i) in step 51j is obtained. Step 51i by setting the signal X _S (i)
Proceed to.

X _S (i) ← X (i) -bc On the other hand, when a negative determination is made in step 51i, that is, when the amplitude of the audio signal X (i) is smaller than the threshold bc, X _S (i) is determined in step 51k. ) Is set to zero and the routine proceeds to step 51i. Note that the process of step 51k may be omitted, and if a negative determination is made in step 51i, the process may directly proceed to step 51i.

[0037] Finally, update the positive envelope value E _pb, negative envelope value E _mb and previous index s _b of the previous last at step 51i and ends this routine. _{E pb ← E p E mb ←} E m s b ← s 7 is a flowchart of the envelope value difference calculation routine executed in step 51a of the subtraction process routine, following the current positive envelope value E _p in step a1 It is calculated by the formula.

[0038] The _{E p = E pb · exp {} -1 / (τ · f s)} where tau is a time constant f _s is negative envelope value E _m current at a sampling frequency step a2 is calculated by the following equation. The maximum value of _{E m = E mb · exp {} -1 / (τ · f s)} audio signal after subtraction processing in step a3 X _S (i) and the current positive envelope value E _p calculated in step a1 again this time replace the positive side envelope value E _p.

The audio signal X _S after the subtraction processing in step a4
(I) a negative side envelope value currently calculated at step 71 E _m
Replacing the minimum value of anew to this negative side envelope value E _m. In step a5, the envelope value difference ΔE is calculated by the following equation, and this routine ends. ΔE = E _p - E _m 8 is an explanatory view to the effects of the subtraction process, (b) shows an audio signal before subtraction, the (b) the audio signal after subtraction. From this figure, it is understood that minute noise has been removed by the subtraction processing.

FIG. 9 shows step 52 of the pitch detection routine.
Is a flowchart of an AGC processing routine executed in step (a), which aims to make the amplitude of the audio signal X _S (i) after the subtraction processing uniform. First, in step 52a, the initial value of the maximum envelope value difference ΔE _max is set to “0”, and in step 52b, the envelope value difference calculation routine shown in FIG. 7 is executed to calculate the envelope value difference ΔE. However, in this case, X (i) of steps a3 and a4 of the envelope value difference calculation routine is X
Needless to say, _S (i).

Next, at step 52c, X _S (i-2) <X _S (i-1) and X _S (i) <X _S (i-1) and X (i-1) _S > 0. That is, it is determined whether the audio signal X _S (i−1) after the subtraction processing sampled before Δt has a positive peak.

When an affirmative determination is made in step 52c, that is, when the audio signal X _S (i-1) after the subtraction processing has a positive peak, in step 52d the envelope value difference ΔE and the maximum value determined before that are determined. the maximum value of the envelope value difference Delta] E _max again set to the maximum envelope value difference Delta] E _max, the process proceeds to step 52e to update the maximum envelope value difference Delta] E _max. Step 52c
In If a negative determination is made, i.e. the process proceeds to directly step 52e when audio signals X _S (i-1) is not a positive peak.

In step 52e, it is determined whether or not the envelope value difference ΔE calculated in step 52b is "0". If a negative determination is made, that is, if ΔE is not “0”, the gain G is set to ΔE _max / ΔE in step 52f. Next, at step 52g, it is determined whether the gain G is equal to or larger than a predetermined threshold value β (for example, 10). When the determination is affirmative, the gain G is set to "1" at step 52h, and the routine proceeds to step 52i. The determination in step 52g may be omitted, and the process may proceed directly from step 52f to step 52i.

Conversely, when a negative determination is made in step 52g, that is, when the gain G is less than the predetermined threshold value β, the flow directly proceeds to step 52i. Step 52e
When the affirmative determination is made in step 5, that is, when ΔE is “0”, the gain G is set to “1” in step 52h, and step 5
Proceed to 2i. Finally, in step 52i, the audio signal X _S (i-1) after the subtraction processing is multiplied by the gain G to calculate the audio signal X _G (i-1) after the AGC processing, and this routine ends.

X _G (i-1) ← G * X _S (i-1) FIGS. 10A and 10B are explanatory diagrams of the effect of the AGC processing. FIG. 10A shows an audio signal before the AGC processing, and FIG. 5 shows an audio signal after AGC processing. That is, when the amplitude of the voice waveform changes rapidly as in (a), occurrence of erroneous detection cannot be avoided in pitch period detection described later. Therefore, it is possible to prevent the occurrence of erroneous detection by making the audio waveform to have a substantially constant amplitude by the AGC process.

FIG. 11 shows step 5 of the pitch detection routine.
Detailed flowchart of the peak detection processing routine executed in step 3
And the voice after the AGC processing in step 53a.
It is determined whether a positive peak has been detected in the signal. That is,
When the condition of is satisfied, X _G(I-2) is a positive peak
It is determined that there is. X_G(I-3) <X_G(I-2) AND X_G(I-
1) <X_G(I-2) and 0 <X_G(I-2) When a positive determination is made in step 53a, that is, AGC processing
Step 5 when a positive peak is detected in the later audio signal
Peak value X at 3b_G(I-2) is stored as P, and
End the routine.

When a negative determination is made in step 53a, that is, when no positive peak is detected in the audio signal after the AGC processing, this routine is directly terminated. Figure 12 determines whether a detailed flowchart of extreme detection clamp processing routine executed at step 54 of pitch detection routine, a negative peak is detected in the speech signal X _G after AGC processing at step 324a. That is, it is determined that X _G (i-2) is a negative peak when the following condition is satisfied.

X _G (i-3)> X _G (i-2) AND
X _G (i-1)> X _G (i-2) and 0> X _G (i-
2) When a positive determination is made in step 54a, that is, when a negative peak is detected in the audio signal after the AGC processing, the peak value P is calculated from the audio signal X _G (i-2) after the AGC processing in step 54b. The subtraction is performed to calculate the clamped audio signal X _C (i-2) emphasizing the negative peak, and the routine ends.

X _C (i−2) ← X _G (i−2) −P When a negative determination is made in step 54a, that is, when no negative peak is detected in the audio signal after the AGC processing, the AGC is performed in step 54c. The audio signal X _G (i-2) after the processing is used as the audio signal X _C (i-2) after the clamp processing, and this routine ends.

X_C(I-2) ← X_G(I-2) FIG. 13 is executed in step 55 of the pitch detection routine.
The detailed flowchart of the pitch cycle detection processing routine
In step 55a, the detected output X _D(I-
3) is calculated by the following equation. X_D(I-3) ← E · exp (−Δt / τ) where Δt is a sampling time and τ is a predetermined time
Is a constant.

E will be described later. Step 5
In 5b, it is determined whether the absolute value of the clamped audio signal X _C (i-3) is greater than the absolute value of the detected output X _D (i-3). If a negative determination is made in step 55b, that is, when the absolute value of X _C (i-3) is less than the absolute value of the X _D (i-3) after detection in step 55c outputs X _D (i
-3) is set to E, and the routine proceeds to step 55f.

When an affirmative determination is made in step 55b, that is, when the absolute value of X _C (i-3) is larger than the absolute value of X _D (i-3), the audio signal after the clamp processing is performed in step 55d. Is determined whether or not there is a negative peak. That is,
When the following conditions are satisfied, it is determined that X _C (i-3) is a negative peak. X _C (i-4)> X _C (i-3) and X _C (i-
2)> X _C (i-3) and 0> X _C (i-3) When an affirmative determination is made in step 55d, that is, when a negative peak is detected in the clamped audio signal, a negative determination is made in step 55e. The peak value X _C (i-3) is set to E, and the routine proceeds to step 55f. When a negative determination is made in step 55d, that is, when no negative peak is detected in the audio signal after the clamp processing, the process proceeds to step 55c.

In step 55f, the value stored as E is set as the post-detection signal X _D (i-3).
At 5 g, a post-detection signal change ΔX _D is calculated by the following equation. ΔX _D ← X _D (i−3) −X _D (i−4) In step 55h, it is determined whether or not the absolute value of the detected signal change ΔX _D is equal to or larger than a predetermined threshold γ.

[0054] When an affirmative determination is made in step 55h, that is, when the post-detection output fell sharply, set to "-1" to speech pitch signal X _P (i-3) in step 55i and ends this routine. Conversely, when a negative determination is made in step 55h, that is, when the post-detection output is not sharply, and set to "0" to speech pitch signal X _P (i-3) in step 55j terminates this routine.

FIGS. 14 and 15 are explanatory diagrams (1/2) and (2/2) of the pitch period detecting method applied in the present invention. 14A shows the audio signal after the clamp processing, and FIGS. 14B and 14C show the enlarged audio signal of the corresponding portion.
The horizontal axis represents time, and the vertical axis represents amplitude. That is, when the sound signal after the clamp processing is inside the envelope starting from the negative peak (b), the envelope is maintained, and when the sound signal is outside the envelope (c).
Represents the audio signal after the clamp processing as an output after detection.

FIG. 15D shows a signal after detection, and FIG. 15E shows a waveform diagram of a voice pitch signal, showing that a pitch pulse is detected at times t ₂ , t ₄ and t ₆ . FIG.
6 is a flowchart of a first gate signal generation routine executed by the first gate generation unit 26, and is a flowchart of step 1;
At 60, it is determined whether the voice pitch signal _XP (i-3) is "-1" and the index j indicating the time when the voice pitch signal was "-1" immediately before is not equal to (i-3). to decide.

If a negative determination is made in step 160, that is, if the voice pitch signal X _P (i-3) is not "-1",
Alternatively, when j is equal to (i-3), this routine is directly terminated. When an affirmative determination is made in step 160, that is, when the voice pitch signal _XP (i-3) is "-1" and the index j is not equal to (i-3), the process proceeds to step 161 and the following expression is used. Calculate the pitch frequency f.

[0058] f (i-3) = f s / {(i-3) -j} Here, f _s equals 1 / Delta] t at the sampling frequency. At step 162, it is determined whether the pitch frequency f is equal to or higher than a predetermined maximum frequency of 500 Hz.
And the process proceeds to step 164. Step 1
If a negative determination is made in 62, the flow directly proceeds to step 164.

At step 164, immediately before the voice pitch signal
The index j indicating the time at which the value was "-1" is set to (i-
Update in 3). Then, after updating the pitch frequency by the following expression in step 165, calculates an average pitch frequency f _m. In this embodiment, the average pitch frequency is calculated by the arithmetic average of three pitch frequencies.
The number of pitch frequencies used is not limited to three. Also, the method of calculating the average pitch frequency is not limited to the arithmetic average, but may be calculated by another method such as a weighted average or a moving average.

[0060] _{f 3 ← f 2 f 2 ←} f 1 f 1 ← f (i-3) f m = (f 3 + f 2 + f 1) / 3 Then, the average pitch frequency f _m is predetermined in step 166 It is determined whether it is equal to or more than a _first threshold Th ₁ (for example, 200 Hz).

[0061] When an affirmative determination is made in step 166, that is, when the average pitch frequency f _m is the first threshold value Th ₁ or more, the step 167 as the speech segment began
To set the first gate signal g ₁ to "1" and terminate this routine. Conversely, when a negative determination is made in step 166, that is, when the average pitch frequency f _m is lower than the first threshold value Th _1, the average pitch frequency f _m at step 168
Is a predetermined second threshold value Th ₂ (for example, 80 Hz)
It is determined whether this is the case.

[0062] When an affirmative determination is made in step 168, that is, when the average pitch frequency f _m is the second threshold value Th ₂ or more, Step 1 as the speech section continues
The gate signal g ₁ proceeds to 67 and maintained at "1" and ends this routine. Conversely, when a negative determination is made in step 168, i.e., the average pitch frequency f _m is the second threshold value Th ₂
If it is less than the threshold, it is determined that the voice section has ended, and the flow advances to step 169 to reset the gate signal g ₁ to “0” and terminate this routine.

[0063] Figure 17 is a diagram for explaining the method of generating the gate signals, (b) the pitch frequency, (b) shows the gate signal g _1. The black circles in (a) represent the pitch frequency f at each time. That is, when three consecutive average pitch frequencies become equal to or higher than the first threshold value Th ₁ (200 Hz), the gate signal g ₁ becomes “1”, that is, opens.

The gate signal g ₁ remains open while the three consecutive average pitch frequencies are equal to or higher than the second threshold Th ₂ (80 Hz), and the three consecutive average pitch frequencies are maintained at the second average pitch frequency. the gate signal g ₁ when it becomes less than the threshold Th ₂ (80 Hz) is "0", that is, closed. FIG.
Is an example of audio signal processing, and (a) is an audio signal V
Shows an audio signal X in which a low frequency noise is removed by filtering with a high frequency pass filter having a cutoff frequency of 300 Hz in a preprocessing routine.

(B) AGC by AGC processing routine
A waveform of the audio signal X _G after treatment, the amplitude of the predetermined amplitude or more components are formed into a substantially constant. (C) shows the signal X _D after the detection processing by the pitch cycle detection processing routine.
(D) shows the pitch frequency f calculated in step 341 of the first gate signal generation routine.

(E) shows the gate signal g ₁ generated in the first gate signal generation routine. As can be understood from this figure, the existence period of the audio signal and the gate signal g ₁
Although but open at a time is consistent, resulting in delayed closing timing of the gate signal g ₁ pitch frequency due to noise ((d) of ○ mark) is generated when the noise after the voice is interrupted occurs .

FIG. 19 is a flowchart of the second gate signal generation routine. The object of the present invention is to solve the above problem by adding steps 190 to 193 to the first gate signal generation routine. That is, step 190
Then, immediately before the voice pitch signal X _p (i-3) is obtained by the following equation,
-1 "from the index j indicating (i-
The elapsed time Dt up to 3) is calculated.

[0068] Dt ← {(i-3) -j} / f s next threshold time Dt _th the elapsed time Dt is predetermined in step 191 (e.g. 0.025 seconds), and the gate signal g ₁ is " 1 "(i.e., the gate is open). If an affirmative determination is made in step 191, that is, if the gate is open and more than 25 milliseconds have elapsed since the last detection of the voice pitch signal of “−1”, in step 193 the modified gate signal g _{Set 1} to "0" to close the gate, update the index j,
Reset the f ₂ and f ₃ ends the routine.

Conversely, when a negative determination is made in step 192, that is, the gate is closed, or finally "-1"
If 25 milliseconds have not elapsed since the detection of the voice pitch signal, the first step shown in FIG.
Is executed and the routine is terminated. In the above embodiment, the threshold time Δt _{th is set} to 25 milliseconds because the frequency 40H is used for 25 milliseconds or more.
z or less, but the pitch frequency of the human voice is 40H
This is because it is hard to imagine that it is less than z.

FIG. 18F shows the corrected gate signal generated by the second gate signal generation routine. The corrected gate signal is not affected by the pitch frequency (marked by ニ in (d)) caused by noise. It turns out that is closed. Although the voice section can be accurately detected by using the correction gate, the voice section can be detected more accurately by further solving the following problems.

[0071] Since the average value of 1.3 one pitch frequency is the gate opens when a first threshold value Th ₁ or more, the opening timing is delayed is bee. 2. Cannot distinguish between sporadic large-amplitude noise and speech signal. 3. Inability to distinguish between noise and noise. 4. Since the prompting sound has a small amplitude, the prompting sound cannot be detected.

In the present invention, the above-mentioned problem is solved by introducing a speech section signal controlled as follows by a gate signal (including a modified gate signal). That is, in order to solve the above 1, 2 and 3, when the gate signal is kept open for a first predetermined period (for example, 50 milliseconds) or more, a second predetermined period (for example, 100 milliseconds) from the current time point Second) Open the voice section signal retroactively.

To solve the above problem 4, the voice section signal is kept open for a third predetermined period (for example, 150 milliseconds) from the time when the gate signal is closed. FIG. 20 is a flowchart of a voice section signal generation routine executed by the voice section signal generation unit 27, and it is determined whether the gate signal g _1b previously calculated in step 200 is “0”, that is, whether the gate is closed. Is determined.

[0074] When an affirmative determination is made in step 200, i.e., either when the gate was closed, a gate signal g ₁ which is calculated this time in step 201 is "0", i.e. whether the gate is maintained closed judge. When an affirmative determination is made in step 201, that is, when the gate is kept closed, the closed maintenance process is executed in step 202, and step 207 is performed.
Proceed to.

When a negative determination is made in step 201, that is, when the gate is opened, the opening process is executed in step 203 and the process proceeds to step 207. Step 200
Is negative, that is, when the gate is open, the gate signal g ₁ calculated this time in step 204
1 ", that is, whether the gate remains open.

When an affirmative determination is made in step 204, that is, when the gate is kept open, an open maintaining process is executed in step 205, and the routine proceeds to step 207. Step 2
When a negative determination is made in step 04, that is, when the gate shifts to closing, the closing process is executed in step 206 to execute step 20.
Go to 7. In step 207, a voice section signal is output, and in step 208, the gate signal g _1b calculated last time is updated with the gate signal g ₁ calculated this time, and this routine ends.

[0077] Figure 21 is a flowchart of open maintenance process routine executed in step 202 of the speech section signal generation routine, the closing duration t is the time the gate signal g ₁ continues the closed state in step 2a _The sampling time Δt is added to _ce . In step 2b, the closing duration t
It is determined whether _ce has become equal to or longer than the third predetermined period of 150 milliseconds.

[0078] When an affirmative determination is made in step 2b, that is, when the gate signal g ₁ has elapsed 150 milliseconds in the closed can, the speech section signal g ₂ in step 2c in the processing time (i-3) " Set to 1 "and end this routine. Conversely, when a negative determination is made in step 2b time, that is, when the gate signal g ₁ has not exceeded the 150 milliseconds in the closed is an index showing the processing time in step 2d is (i-3) The second gate signal g ₂ (i−3) is set to “0”, and this routine ends.

FIG. 22 is a flowchart of the opening processing routine executed in step 203 of the voice section signal generation routine. In step 3a, the gate signal g _1b calculated last time is set to “1”. In step 3b, the closing duration t
return to the _ce "0", an index indicating the processing time in step 3c is the routine is set to g ₂ a (i-3) "1" is a speech section signal when a (i-3) finish.

[0080] Figure 23 is a flowchart of open maintenance process routine executed in step 205 of the speech section signal generation routine, the opening duration is a time gate signal g ₁ continues the opened state in step 5a t _The sampling time Δt is added to _oe . In step 5b, the open duration time t
It is determined whether or not _oe is _equal to or longer than a first predetermined period of 50 milliseconds.

[0081] If a negative determination is made in step 5b, that is, when when the gate signal g ₁ has not reached the 50 milliseconds in the open is an index showing the processing time in step 5c is (i-3) G is the voice section signal of
₂ Set (i-3) to "0" and end this routine. If a negative determination is made in step 5b, that is, when the gate signal g ₁ has elapsed 50 milliseconds in the open, the
In step 5d, the second predetermined time from the processing time, ie, 10
An index i _B indicating a time that has been traced back by 0 milliseconds is calculated by the following equation.

I _B ← (i−3) −0.1 / Δt The second term on the right side is the number of samplings existing in 100 milliseconds. In step 5e, the index i _B is set to be equal to or greater than zero in order to prevent the sound signal from going back to an area where no audio signal exists. In step 5f, g ₂ (i) is the audio section signal when the index indicating the time is i _{B. B} )
Set to "1".

Whether the index i _B has become the index (i−3) indicating the processing time in step 5g,
It is determined whether the retroactive processing has been completed for the predetermined period.
And when a negative determination is made, i.e. when the retroactive process has not been completed it is returned to step 5f and decrementing the index i _B at step 5h. Conversely, when an affirmative determination is made in step 5g, that is, when the retroactive processing is completed, this routine ends.

FIG. 24 is a flowchart of an opening process routine executed in step 206 of the voice section signal generation routine. In step 6a, the gate signal g _1b calculated last time is set to “0”. In step 6b, the open duration time t
_oe returned to "0", an index indicating the processing time in step 6c is the routine is set to (i-3) is a speech section signal when a g ₂ a (i-3) "0" finish.

FIG. 25 is a flowchart of a second gate signal output routine executed in step 207 of the voice section signal generation routine. In step 7a, a time which is 100 ms earlier as the second predetermined time, which is a second predetermined time. the index i _b showing a calculating by the following equation. i _b ←
(I-3) setting the index i _b zero or more to -0.1 / Delta] t audio signal in step 7b is prevented from going back to the region that does not exist,
In step 7c, g ₂ (i _b ) is output, and this routine ends.

FIG. 26 is a flow chart of a word extraction routine executed by the word extraction unit 28.
0 index indicates the time in i _b is a word signal W when the (i _b) is calculated by the following equation. W (i _b ) ← X (i _b ) * g ₂ (i _b ) where X (i _b ) is an audio signal stored in the storage unit 24.

In step 261, W (i _b ) is output, and this routine ends.

[0088]

According to the voice section detection apparatus of the first invention, the gate signal is controlled based on the voice pitch extracted by processing the voice signal in the time domain, and the voice section is detected by the gate signal. Therefore, a voice section can be detected with a simple configuration. According to the voice section detection device according to the second invention, it is possible to divide the voice signal into a plurality of voice sections based on the voice section.

According to the voice section detection apparatus of the third invention, the voice section is detected based on the voice pitch extracted by processing the voice signal in the time domain, so that the voice section is detected almost in real time. It is possible to do. According to the voice section detection device according to the fourth aspect, it is possible to suppress the fluctuation of the amplitude of the voice signal.

According to the speech section detection device according to the fifth aspect, it is possible to reliably remove noise present in the speech signal. According to the voice section detection device according to the sixth aspect, the voice pitch can be reliably extracted by making the amplitude of the voice signal substantially uniform. According to the speech section detection device according to the seventh aspect, when the uniform amplitude gain is the predetermined threshold value, the mixing of the noise can be prevented by resetting the uniform amplitude gain to the unity gain.

According to the speech section detection apparatus according to the eighth aspect, it is possible to prevent the gate signal from being erroneously opened due to the influence of noise. According to the speech section detection device according to the ninth aspect, it is possible to prevent the gate signal from being accidentally closed due to the influence of noise. Tenth
According to the voice section detection device according to the invention, it is possible to reliably close the first gate signal when the voice pitch is no longer extracted.

According to the speech section detection apparatus of the eleventh aspect, it is possible to compensate for the delay of opening of the gate signal and to reliably eliminate noise by distinguishing it from aero sounds. First
According to the voice section detection device of the second aspect, it is possible to reliably detect a prompt sound having a small amplitude. According to the voice section detection device according to the thirteenth aspect, it is possible to prevent erroneous detection even when voice sections overlap.

[Brief description of the drawings]

FIG. 1 shows a result of detecting a voice section based on a conventional pitch cycle.

FIG. 2 is a functional configuration diagram of a voice section detection device according to the present invention.

FIG. 3 is a flowchart of an audio sampling routine.

FIG. 4 is a flowchart of a preprocessing routine.

FIG. 5 is a flowchart of a pitch detection routine.

FIG. 6 is a flowchart of a subtraction processing routine.

FIG. 7 is a flowchart of an envelope difference calculation routine.

FIG. 8 is an explanatory diagram of an effect of a subtraction process.

FIG. 9 is a flowchart of an AGC processing routine.

FIG. 10 is an explanatory diagram of an effect of AGC.

FIG. 11 is a flowchart of a peak detection processing routine.

FIG. 12 is a flowchart of an extreme value detection / clamping processing routine.

FIG. 13 is a flowchart of a pitch cycle detection processing routine.

FIG. 14 is an explanatory diagram (1/2) of a pitch period detection method.

FIG. 15 is an explanatory diagram (2/2) of a pitch period detection method.

FIG. 16 is a flowchart of a first gate signal generation routine.

FIG. 17 is an explanatory diagram of a generation method of a gate signal.

FIG. 18 is an example of audio signal processing.

FIG. 19 is a flowchart of a second gate signal generation routine.

FIG. 20 is a flowchart of a voice section signal generation routine.

FIG. 21 is a flowchart of a closing maintenance processing routine.

FIG. 22 is a flowchart of an opening process routine.

FIG. 23 is a flowchart of an open maintaining process routine.

FIG. 24 is a flowchart of a closing process routine.

FIG. 25 is a flowchart of a voice section signal output routine.

FIG. 26 is a flowchart of a word extraction routine.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 21 ... Microphone 22 ... Line amplifier 23 ... Analog / digital conversion part 24 ... Storage part 25 ... Pitch detection part 26 ... Gate signal generation part 27 ... Voice section signal generation part 8 ... Word extraction part

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Hideki Kitao 1-2-28 Gosho-dori, Hyogo-ku, Kobe City, Hyogo Prefecture Inside Fujitsu Ten Limited (72) Inventor Shinichi Iwamoto 1-chome, Goshodori, Hyogo-ku, Kobe City, Hyogo Prefecture 2-28 Inside Fujitsu Ten Limited (72) Inventor Osamu Iwata 1-2-28 Goshodori, Hyogo-ku, Kobe City, Hyogo Prefecture Inside Fujitsu Ten Limited (72) Inventor Masataka Nakamura Miyake Saiki-ku, Hiroshima City, Hiroshima Prefecture Chome 1-1 Inside Tsuru Gakuen (72) Inventor Yoshinao Omoto 1-5-25 Kugayama, Suginami-ku, Tokyo F-term (reference) 5D015 AA05 DD03 EE05

Claims (13)

[Claims] 1. A pre-processing means for removing noise contained in an audio signal, a voice pitch extracting means for extracting a voice pitch signal from the audio signal from which noise has been removed by the pre-processing means, and a voice pitch extracting means A voice section detection device, comprising: a gate signal generation unit that generates a gate signal based on the voice pitch extracted in step (a); and a voice section signal generation unit that generates a voice section signal based on the gate signal generation unit. 2. The audio further comprising audio signal classification means for dividing the audio signal, from which noise has been removed by the preprocessing means, into a plurality of audio signals based on the audio section signal generated by the audio section signal generation means. Section detection device. 3. The subtraction processing means for performing a subtraction process for removing an audio signal having a predetermined amplitude or less from the audio signal from which the noise has been removed by the preprocessing means, An equalizing means for adjusting the amplitude of the audio signal subjected to the subtraction processing to a substantially constant amplitude; and detecting a positive peak and a negative peak following the positive peak from the audio signal adjusted to a substantially constant amplitude by the equalizing means. A negative peak emphasizing means for subtracting the positive peak from the negative peak to generate an audio signal in which the negative peak is emphasized; and detecting and processing the audio signal in which the negative peak is emphasized by the negative peak emphasizing means. The voice section detection device according to claim 1, further comprising: a differential processing unit that performs differential processing on a subsequent signal. 4. The subtraction processing means calculates a positive envelope and a negative envelope of the audio signal from which noise has been removed by the preprocessing means, and calculates a difference between the positive envelope and the negative envelope. An envelope difference calculating means for calculating a certain envelope difference, and a subtraction processing threshold value calculating means for calculating a subtraction processing threshold value by multiplying the envelope difference calculated by the envelope difference calculating means by a predetermined predetermined coefficient multiple. When the amplitude of the audio signal from which the noise has been removed by the preprocessing unit is equal to or greater than the subtraction threshold calculated by the subtraction threshold calculation unit, a subtraction threshold subtraction unit that subtracts the subtraction threshold from the amplitude of the audio signal. The voice section detection device according to claim 3, comprising: 5. The method according to claim 1, wherein the subtraction processing means sets the amplitude of the audio signal to zero if the amplitude of the audio signal from which the noise has been removed by the preprocessing means is smaller than the subtraction processing threshold value calculated by the subtraction processing threshold value calculation means. 5. The voice section detection device according to claim 4, further comprising a zero setting means for setting the value to. 6. The uniform amplitude means calculates a positive envelope and a negative envelope of the audio signal from which noise has been removed by the preprocessing means, and calculates a difference between the positive envelope and the negative envelope. An envelope difference calculating means for calculating a certain envelope difference; a maximum envelope difference holding means for holding a maximum envelope difference among the envelope differences calculated before and now by the envelope difference calculating means; and 4. The voice section detection device according to claim 3, further comprising: a matched amplitude gain calculating means for calculating a matched amplitude gain by dividing the maximum envelope difference held by the envelope difference holding means by the current envelope difference. 7. A unit gain setting means for setting the matching amplitude gain to a unit gain when the matching amplitude gain calculated by the matching amplitude gain calculating means is equal to or greater than a predetermined threshold value. The voice section detection device according to claim 6, further comprising: 8. The gate signal generating means, wherein the gate signal is output when an average value of a predetermined number of continuous voice pitches extracted by the voice pitch extracting means is equal to or greater than a predetermined gate opening threshold value. The voice section detection device according to claim 1, further comprising a gate signal opening unit that opens the signal. 9. The gate signal generating means, when the gate signal is once opened by the gate signal opening means, is an average value of a predetermined number of continuous voice pitches extracted by the voice pitch extracting means. 9. The voice section detection device according to claim 8, further comprising a gate signal open maintaining means for maintaining said gate signal in an open state when is greater than or equal to a predetermined gate close threshold value smaller than said gate open threshold value. 10. The gate signal generating means, when an average value of a predetermined number of continuous voice pitches extracted by the voice pitch extracting means becomes less than the gate open threshold, outputs the gate signal. The voice section detection device according to claim 9, further comprising a gate signal closing unit that closes the gate signal. 11. A first predetermined period timer means for measuring a predetermined first predetermined period from a time point when the gate signal generated by the gate signal generation means is opened, and And a voice section signal opening means for opening the voice section signal retroactively for a predetermined second predetermined period from the time when the first predetermined period clocking by the first predetermined period clocking means ends. Item 3. The voice section detection device according to item 1 or 2. 12. A third predetermined period timer means for measuring a predetermined third predetermined period from a point in time when the gate signal generated by the gate signal generation means is closed. 12. The voice section detection device according to claim 11, further comprising voice section signal closing means for closing a voice section signal when the timing of the third predetermined period by the third predetermined time counting section is completed. 13. The voice section signal generation means, before the time measurement of the third predetermined time period by the third predetermined time counting means does not end, retroactively from the voice section signal opening means for the second predetermined time period. 13. The voice section detection device according to claim 12, further comprising voice section signal open state maintaining means for maintaining the voice section signal in an open state when the voice section signal is opened.