JPH01307800A - Voice detecting method - Google Patents

Abstract

PURPOSE:To improve voice detecting performance by regarding the rising time point of an envelope waveform where a read level becomes a constant value larger than a level which is read one sampling cycle before as the start point of a voice section and regarding the sum value of the last read waveform level and constant value as a voice detection threshold value. CONSTITUTION:The point of time of the rising of the envelope waveform where the newly read level becomes the constant value larger the level which is read one sampling cycle before is regarded as the start point of the voice section and the value obtained by adding the constant value to the envelope waveform level which is read one sampling cycle before on a logarithmic scale is regarded as the voice detection threshold value. Therefore, the noise level of an input signal can be measured without providing any noise level measurement period specially, and the voice detection threshold value is set to a value adaptive to the noise level. Consequently, invariably accurate voice detection is performed.

Description

[Detailed Description of the Invention] [Industrial Field of Application] This invention is applicable to the audio level and
The present invention relates to a voice detection method necessary for determining whether an input signal is a voice section or a silent section under conditions where the noise level changes significantly.

In this method, natural dialogue is realized by detecting the end of a user's utterance and starting sending the next response message. In addition, the content of the user's utterance is estimated based on the length of the user's voice or the presence or absence of the user's reaction after sending the response message (the user's voice in response to the response message), and the content of the response message to be sent next is selected. An interactive answering machine has also been proposed that further improves the naturalness of dialogue.

FIG. 6 is a diagram showing the operation flow of the above-mentioned interactive answering machine, and FIGS. 7(a) and (b).

(C) is a diagram showing a processing flow when measuring the length of a user's voice, confirming the presence or absence of a user's reaction to a response message, and detecting the end of utterance, respectively.

Next, the operation of this interactive answering machine will be briefly explained.

First, when an incoming call is received, the line is automatically closed, a first response message is sent, and the user's message is recorded until the end of utterance is detected. Next, the voice length of the user's message is measured, and if the user's voice length is determined to be short, a second response message "Yes" is sent to encourage the other party to speak, and further recording and voice length are performed. Take measurements. If the length of the user's voice is determined to be long, a fourth response message is sent to make the user announce the business.

In addition, if the user's voice length is determined to be medium, after sending the second response message mentioned above, the system checks whether or not the user has responded, and depending on the user's response, sends a message indicating that the user is required to announce the caller's name. Third
The response message or the fourth response message is sent. If the third response message is sent, the fourth response message is sent after recording the user's message.

The measurement of the length of the user's voice, the confirmation of the user's response to the response message, and the detection of the generated number 9 are all performed based on voice detection technology. For example, the length of a user's voice is measured as shown in FIG. 7(a).

That is, when the voice detection operation state is reached, the voice timer and silent timer are reset, and both timers are operated based on the voice detection result. Next, when the voice detection state is reached, the voice timer starts counting, and when the voice becomes silent, the voice timer stops counting. Then, the silence timer starts counting. Here, when the voice detection state returns within a predetermined time (utterance end determination threshold), the count of the silence timer is reset and the count of the sound timer is restarted. When this operation is repeated and the silent time continues to exceed the utterance end determination threshold, it is assumed that the utterance has ended, and the count value of the utterance timer up to that point is taken as the voice length.

In addition, for answering machines that record business messages or response messages in IC memory, a recording method has been proposed that actually lengthens the recording time by recording only the voice section of the input signal and not recording the silent section. (For example, Japanese Patent Laid-Open No. 62-47329).

Generally, the voice detection method used in these answering machines, etc. extracts the envelope waveform of the input signal,
The extracted envelope waveform level is compared with a certain voice detection threshold, and when the envelope waveform level is above the voice detection threshold, there is a sound state, and when the envelope waveform level is below the voice detection threshold, there is no sound. It is determined that the condition is

To extract the envelope waveform, an input signal is input to the CPU, and the envelope waveform is extracted by internal processing of the CPU. Alternatively, there is a method such as providing an envelope extraction circuit consisting of a rectifier circuit and a low-pass filter before importing into the CPU. In addition, the power change of the input signal may be treated as equivalent to an envelope waveform and processed.

The voice detection performance of this voice detection method is mainly determined by the method of determining the voice detection threshold. In particular, when a telephone line is used as an input signal source, the voice level and noise level vary depending on the line condition and the installation environment of the opposite telephone. Regarding voice levels, telephone lines allow a loss of up to 25 dB between terminals as a design standard for the transmission path.

Furthermore, the level of occurrence of calls to the other party's telephone differs from person to person and changes. Regarding the noise level, when the opposing phone is placed in a quiet room such as a bedroom, the ambient noise level is around 40 dBspt, whereas when the opposing phone is a public telephone installed on the street, the ambient noise level is 80 dB.
,, sometimes exceeds L.

In order to perform voice detection accurately, the voice detection threshold must be set to a value greater than the noise level. On the other hand, if the voice detection threshold is too large, there is a possibility that voice sections with low levels cannot be detected. For this reason, the voice detection threshold must be set to a value greater than the noise level, and moreover, it is desirable that the value be slightly greater than the noise level.

In the method of comparing the envelope waveform of the input signal and the voice detection threshold using CPU internal processing, it is possible to easily measure the noise level and determine the voice detection threshold based on that value. Therefore, there is a method of setting a period as a noise level measurement period during which there is a low possibility that the user will cause noise generation and in which only the noise level can be measured purely, and determining the voice detection threshold based on the noise level measured during that period. There is. In this case, the noise level will vary slightly over time, so
The noise level measurement period is often set to be long, and the average value within the period is treated as the noise level.

[Problems to be Solved by the Invention] However, depending on the service, it may not be possible to provide such a special noise level setting period. For example, an interactive answering machine provides natural responses without making the user feel uncomfortable talking to the answering machine, so voice messages are constantly exchanged between the user and the answering machine.

For this reason, it is difficult to specify a period during which a user is unlikely to cause an outbreak. Furthermore, even if a noise level measurement period is set, there is no guarantee that users will not generate noise during this period, so if a user generates noise during this period, the observed noise level may be higher than the actual one. Therefore, accurate noise level measurements cannot be performed. It is conceivable to suppress the user's vocalization by sending out a guidance voice that says, "Please do not speak for a while," but this would detract from the user's degree of freedom and degrade service performance.

Further, in a telephone line, the noise level may change over time depending on the installation condition of the opposite telephone and the line condition. In such a case, the voice detection threshold determined based on the noise level determined during a single noise level measurement period cannot avoid deterioration of the voice detection performance.

SUMMARY OF THE INVENTION An object of the present invention is to provide a voice detection method that can always perform highly accurate voice detection without being affected by noise levels.

[Means to solve the problem]

In the voice detection method according to claim (1) of the present invention, an envelope waveform level read one sampling period ago is stored, and the envelope waveform level newly read by the CPU is stored as an envelope waveform level read one sampling period ago. When compared with the waveform level, the rising point of the envelope waveform at which the newly read level becomes larger than the level read one sampling period ago by a certain value (Vd+r) is regarded as the starting point of the audio section, and this one sampling period Add a constant value (Voff-) to the previously read envelope waveform level on a logarithmic scale.
The value obtained by adding a) is used as the voice detection threshold (V th).

In the voice detection method according to claim (2), when the envelope waveform level exceeds the voice detection threshold (Vth) and then falls below the voice detection threshold (Vth), the envelope waveform rises again. detection is started, and the point in time when the rising edge of the envelope waveform is detected is regarded as the start point of a new voice section, and a constant value (Voff-a) on a logarithmic scale is set to the envelope waveform level read one sampling period ago. The added value is used as a new voice detection threshold (V th).

The voice detection method according to claim (3) is one in which the value obtained by adding a constant value (V art-b) on a logarithmic scale to the envelope waveform level read one sampling period before is set as a constant value (Vd+r). It is.

In the voice detection method according to claim (4), if the envelope waveform level does not exceed a voice detection threshold (V th) within a certain time (Tedt.) after the envelope waveform rises, The interval up to this point is regarded as a silent interval, and the detection of the rise of the envelope waveform is started again.The time when the rise of the envelope waveform is detected is regarded as the starting point of a new voice interval.
Add a constant value (Vote-a) to the envelope waveform level read before the sampling period on a logarithmic scale.
The new voice detection threshold (v th)
That is.

In the voice detection method according to claim (5), after the envelope waveform level rises and exceeds the voice detection threshold (V th) within a certain time (radg.), the minimum voice duration time (Tvc) is determined. If the voice detection threshold value (V th) falls below the voice detection threshold (V th) again within that period, this section is regarded as a silent section.

In the voice detection method according to claim (6), after the envelope waveform rises and the envelope waveform level exceeds the voice detection threshold (Vth) within a certain period of time (Tedga), the minimum voice continuous time (Tvc) is determined. After the elapsed time, the envelope waveform level again reaches the voice detection threshold (v th
), the voice detection threshold (v t
h), and then calculate the minimum duration of silence (
If the rising edge of the envelope waveform is detected again within T nvc), the stored value is set as the voice detection threshold CV th).

[Effect]

In the voice detection method according to claim (1) of the present invention, the rising point of the envelope waveform at which the newly read level is greater than the level read one sampling period ago by a certain value (V(11r)) is the point in the voice section. It is regarded as the starting point, and the value obtained by adding a constant value (■.ft-a) on a logarithmic scale to the envelope waveform level read one sampling period before is taken as the voice detection threshold (V th).

That is, it becomes possible to measure the noise level of an input signal without providing a special noise level measurement period, and the voice detection threshold (v th) is set at a value that is appropriate for the noise level.

In the voice detection method of claim (2), when the envelope waveform level exceeds the voice detection threshold (V th) and then falls below the voice detection threshold (V th), the envelope waveform rises again. detection has started,
The point in time when the rising edge of the envelope waveform is detected is regarded as the start point of a new voice section, and the new value is the sum of the envelope waveform level read one sampling period ago and a constant value (Voff-a) on a logarithmic scale. This is the voice detection threshold (V th).

In other words, if it turns out that noise has been mistakenly detected as speech, or if the end of a speech section is detected, the noise level is newly measured the next time the start of the speech section is detected, and speech detection is performed. A threshold (v th) is set.

In the voice detection method of claim (3), the value obtained by adding a constant value (Voff-b) on a logarithmic scale to the envelope waveform level read one sampling period before is (vdI,) total.

That is, the value to be compared with the envelope waveform level difference value when detecting the start end of the voice is also adapted to the noise level at that time.

In the voice detection method of claim (4), the envelope waveform level reaches the voice detection threshold (V th ) within a certain time (T-dt-) after the envelope waveform rises.
If it does not exceed , the section from the rise to that point is considered a silent section, detection of the rise of the envelope waveform is started again, and the point at which the rise of the envelope waveform is detected is regarded as the start point of a new voice section. At the same time, a constant value (V,
tv-) is the new voice detection threshold (V
th). Moreover, in the voice detection method of claim (5), the envelope waveform level rises and CT-at for a certain period of time. ) within the voice detection threshold (V t
h) and then falls below the voice detection threshold (V th) again within the minimum continuous time of voice 1a (Tvc), this section is regarded as a silent section.

In other words, even if noise is mistakenly detected as the start of speech, the envelope waveform change after the start is detected or after the speech detection threshold (Vth) has been exceeded is different from that of the speech section, so it is not a speech section. When this is determined, the section is determined to be a silent section.

In the voice detection method of claim (6), after the envelope waveform rises and the envelope waveform level exceeds the voice detection threshold (V th) within a certain time (Tedm-), the minimum voice duration time (Tvc) is determined. ), the envelope waveform level again reaches the voice detection threshold (V t
h), the voice detection threshold (V t
h) is stored, and then the minimum duration of silence (T
If the rising edge of the envelope waveform is detected again within the time interval (V th ), the stored value is set as the voice detection threshold (V th ).

In other words, if the envelope waveform level momentarily decreases during a voice section and falls below the voice detection threshold, the voice detection threshold (Vth) is not set anew, but the previous value is used. Used as is.

〔Example〕

If we focus on the envelope waveform of the human signal, we can see that S
/N is ensured, the sound section has a higher level than the preceding and succeeding sections. Therefore, the rising point of the envelope waveform is regarded as the starting point of the voice section. Specifically, an envelope waveform (here, the envelope waveform level input to the CPU is assumed to be Xn, where n represents the chronological order of input to the CPU) is input at a constant sampling period, and 1 is input to the register. Stores the envelope waveform level before the sampling period, Xo. -Xn-1≧V dlF...
... (1) (However, Vdl is constant) The time when n is satisfied is the rising point of the envelope waveform. The rise of this envelope waveform depends on the type of first syllable that constitutes the speech section. The first syllable is a voiceless consonant (S sound, P sound, ni sound).

Consonants), the envelope waveform rises slowly, and in other cases it rises steeply.

Therefore, in order to accurately detect the voice start point even when the first syllable includes a voiceless consonant (S sound, P sound, Ni sound, consonant), V dlf needs to be set to a sufficiently small value. .

In this way, when the rising time is detected, since the section immediately before the rising is a silent section, the envelope waveform level value immediately before the rising can be said to be a representative value of the noise level. Therefore, based on this value, the voice detection threshold (
V th).

Normally, unless noise is applied in the transmission system, the input signal S
/N is the S/N at the opposite telephone end even if the line loss changes.
The value is the same as N. Additionally, users tend to increase the generation level as the ambient noise increases. From the above, when determining the voice detection threshold based on the noise level, it is desirable to set the ratio between the noise level and the voice detection threshold to be constant regardless of the noise level. Therefore, specifically, a constant value (Vofr-a) on a logarithmic scale is added to the noise level to determine the voice detection threshold. This can be expressed numerically as follows.

log [v th]” log [X n-1] +V
ate-a = ... (2) However, Xn-xn-,
≧Vdl, the way the envelope waveform rises at the beginning of the voice is also considered to depend on the noise level, so (1)
The voice detection sensitivity can be improved by improving the formula as follows.

X n −x n, ≧ Vd1f
(Xn-1) ... ... (3) However, log[V aIt (X 11-+)] = lo
g [x n-11+ V ore-b... (
4) Note that when the values of Vd1f (Xn-1) and Voff are set small, the voice detection sensitivity improves for input signals with low S/N, but on the other hand, the probability of erroneously detecting noise increases. Become. In order to reduce the probability of erroneously detecting noise while maintaining voice detection sensitivity for input signals with low S/N, the following two algorithms are required.

■ If the input signal consists only of noise, the envelope waveform changes randomly even after rising.

On the other hand, in the case of voice, the voice detection threshold (
Vth). Therefore, the rise time threshold T
mdt@ is provided, and the envelope waveform of the input signal rises T. The envelope waveform level is within Vt
If it does not exceed h, the section is regarded as a silent section.

■ The duration of human utterance when expressing a meaningful message is 500 m5ec or more. Therefore, the minimum sound duration time Tvc is set to about 500 m5ec, and even when the envelope waveform of the input signal rises and the envelope waveform level exceeds Vth within T*dg*, the T V
If it falls below Vth within e, that section is regarded as a silent section.

That is, when the envelope waveform level exceeds Vth and then falls below Vth, and the end of the voice section is detected, or from the rising edge of T, d. If it does not exceed the envelope waveform level (Vth) even after the elapse of time, the previously set voice detection threshold (Vth) is invalidated, and when the envelope waveform rises again, a new voice detection threshold (Vth) is set. Set. Through this process, even if the noise level changes somewhat during the voice detection process, it is possible to set the noise level in an appropriate manner.

Note that the envelope waveform also fluctuates during the voice section, and the envelope waveform momentarily changes to the voice detection threshold (v th).
It may fall below. This voice detection threshold (V th
), the envelope waveform level is
It has a form in which low-level audio is superimposed on the noise level. Therefore, if the voice detection threshold (Vth) is set based on this value at the time of the next rise of the envelope waveform, the voice detection threshold (V th) will be a larger value than the desired value. In this case, the rear part of the voice section will be determined to be a silent section.

Therefore, a voice detection threshold register and a minimum duration of silence T nVC are provided, and when the envelope waveform exceeds Vth and falls below Vth after the minimum duration of voice duration "rye" has elapsed, the voice detection threshold (v th) in the voice detection threshold register, and the minimum duration of silence T after the envelope waveform falls below Vth.
At the rise of the envelope waveform within nVc, the voice detection threshold (Vth) stored in the voice detection threshold register is used instead of calculating a new voice detection threshold (Vth).
th) is used. FIG. 1 is a block diagram for explaining an embodiment of the voice detection method of the present invention. In this figure, 1 is an envelope waveform extraction circuit composed of a rectifier circuit and a low-pass filter, 2 is an A/D converter for reading the envelope waveform level into the CPU, and 3
4 is a waveform register that stores the envelope waveform level; 5 is a waveform register that stores the envelope waveform level read one sampling period before; 6 is a threshold that stores the voice detection threshold. A value register, 7 is a timer that sends a signal to the A/D converter 2 and the control unit 3 every sampling period T□□1, 8 is a sound counter for measuring the length of a sound period, and 9 is a length of a silent period. 10 is an output status register that stores the voice detection result, and 11 is an invalid voiced section that stores the length of a section that was originally considered to be a voiced section but later turned out to be a silent section. It is a long register.

FIG. 2 is a diagram showing a configuration example of the envelope waveform extraction circuit 1 shown in FIG. 1, and FIG. 3 is a diagram showing the processing flow in FIG. 1. Hereinafter, the explanation will be given according to FIGS. 1 and 3. Note that ST1 to ST38 in FIG. 3 indicate each step.

After the initial settings have been made (STO), the human input signal has a constant sampling period T sa determined by timer 7,
, pl, envelope waveform extraction circuit 1. The signal is read into the control unit 3 via the A/D converter 2 (STl). According to the sampling theorem, the sampling period T ia □1 at this time needs to be 1/2 or less of the change period of the envelope waveform. Since the change period of the envelope waveform is approximately equal to the human speech rate (10 to 20 [syllables/sec]), there is no problem if the sampling period T gaapl is about 20 m5ec.

Next, the control unit 3 moves the data stored in the waveform register 4 to the waveform register 5 (STl), and then stores the read envelope waveform value in the waveform register 4 (Sr3). Furthermore, the (4)
After finding V dlf (X n-1) according to the formula, calculate the difference between the data (xn) of waveform register 4 and the data (xn-I) of waveform register 5, and these values are
3) It is determined whether the formula is satisfied (Sr4). As a result of the judgment, if the equation (3) is checked, it is determined that there is no sound, and the output status register 1o is set to [silence] (Sr1), and the silence kakunta 9 is incremented (57B).

On the other hand, as a result of the determination, if the formula (3) is satisfied, the silent length up to that point is compared with the minimum duration of silent duration T nvc (Sr7). The value of the minimum duration of silence T nvc is the length of time that is considered to be the shortest among the time required for a human to pause vocalization due to breathing or the like. The length of silence up to that point is T nvc or more (silence counter value ≧ (Tnvc
/T1. , pl)), the voice detection threshold (Vth) is determined by equation (2).
is newly calculated (5T8), and this value is stored in the threshold register 6 (Sr9). The length of silence up to that point is T n
Less than Ve (silence counter value (Tnvc/T□p1))
At this time, it is determined that this is not the beginning of a voice section but an extension of the previous voice section, and the voice detection threshold value Vth stored in the threshold register 6 is read out (STIO);
After this, the envelope waveform level (Xn) and the voice detection threshold (V th) are compared (STII), and when the envelope waveform is equal to or higher than the voice detection threshold (x n≧v
th), it is determined that there is a sound, and the sound counter 8 is incremented (ST12), and the output status register 10 is set to "sound" (ST13).

On the other hand, when the envelope waveform is below the voice detection threshold (
Since the rising edge of the envelope waveform (that is, the start of the voice section) is also detected when
In addition to ST14), it is determined whether the voice detection threshold is exceeded or not before T and dge have elapsed since the rise of the envelope waveform (5T15), and if the voice detection threshold is exceeded, the output status register 10 is set to "sound present". (ST16).

However, T, d, from the rise of the envelope waveform. If the envelope waveform level does not rise above the voice detection threshold even after the passage of time (speech counter value - CT-dt./To□1)), it is assumed that the envelope waveform has risen due to noise, and The section that was determined to be a voiced section is re-judged as a silent section, and the length of the section counted in the voiced section 8 is set in the invalid voice section length register 11 (ST20), and at the same time, it is set in the output star register 10. Set “Sound section disabled” (
ST21).

Next, the sum of the silence counter value and the voice counter value is stored in the silence counter 9 (ST22), and after setting the voice counter 8 to 0 (ST23), the process proceeds to step S.
Return to TI.

In addition, Tedt. The value of must be set to a shorter time than the duration of a fricative sound such as a shivering sound at the beginning of the voice.

On the other hand, when the envelope waveform level exceeds the voice detection threshold, the envelope waveform level and the voice detection threshold are continuously compared (ST24.25, 26,
27, 28, 29.30), the sound section continues for longer than the minimum sound duration time T ve (speech counter value ≧ (T vc/T
sa□I)), the section is judged to be a true sound section, the silent counter 9 is set to 0 (5T31), and steps ST24 to ST30 are executed. After that, when the envelope waveform level becomes less than the voice detection threshold, "silence" is set in the output status register 10.
T33), after setting silence counter 9 to 1 (ST34)
, set the sound counter 8 to O (ST38), and step S
Return to T.I.

In addition, before the sound interval length continues longer than the minimum sound duration time Tvc (sound count value < (Tvc/T□□I)),
When the envelope waveform level becomes less than the voice detection threshold (ST27.32), it is assumed that the envelope waveform level has increased due to the influence of noise, and the section that was determined to be a sound section from the rise to that point is It is re-judged as a silent section, and the section length counted by the voice counter 8 is set in the invalid voice section length register 11 (
ST35), the output status register 10 is set to “
"Speech interval invalid" is set (ST36). Next, the sum of the silence counter value and the voice counter value +1 is stored in the silence counter 9 (ST37), and after setting the voice counter 8 to 0 (ST38), the process returns to step STI.

Note that the value of Tvc needs to be set to the shortest value of the voice interval when a human generates a meaningful message.

FIG. 4 shows a configuration example when the embodiment of the present invention explained in FIG. FIG. In this figure, 12 is a voice analyzer that analyzes voice signals into a form that can be stored in an IC memory, and 13 is a CP.
14 is an audio storage section consisting of an IC memory; 15 is a timer that determines the read timing of audio data from this audio storage section 14 to the control section 13; and 16 is the audio storage section. 14 is an address pointer specifying an access address; 17 is a voice detection circuit according to an embodiment of the present invention explained in FIG. 3; 18 is an output status register which is a component of this voice detection circuit 17; This is an invalid voice section length register which is a component of the detection circuit 17.

FIG. 5 shows a processing flow in the embodiment of FIG. 4.

The control unit 3 analyzes the voice data analyzed by the voice analyzer 12 at a period T determined by a timer 15. Audio storage section 14 at p2
Load from. The loading speed at this time is Speech Analyzer 1
2 is normal PCM and the input signal band is telephone audio band 3.4
If it is kHz, it is 8kHz2. Next, the control unit 3 stores the read audio data at the address indicated by the address pointer 16 in the audio storage unit 14, and advances the value of the address pointer 16 by one.

Normally, the above process is repeated, but the above-mentioned T sa□8
The output status register 18 of the audio detection circuit 17 periodically
When a new voice detection result is written to , the following operations are performed.

(2) When "silence" is written in the output status register 18, it is determined that there is no sound during T ga □1 immediately before, so this portion is overwritten with the next audio data. Therefore, the address pointer 16 is returned to the address where the previous data was stored by T□□1.

(2) When "voiceable section invalid" is written in the output status register 18 Since the section from the immediately preceding audio rise to this point is determined to be silent, this section is overwritten with the next audio data. Therefore, the address pointer 16 is returned to the address where the audio data at the beginning of the section is stored. Since the section length is written in the invalid voice section length register 19, that value is read out and used.

■ T1. immediately before "audio" is written to the output status register 18. □Since it was determined that there was sound during the 3rd period, recording continued.

The above process is repeated until an operation end command is input or until there is no more free space in the audio storage section 14 and an operation stop command is issued.

〔Effect of the invention〕

Since this invention is configured as described above, it produces the effects described below.

That is, in claim (1), it is possible to measure the noise level of a human signal without providing a special noise level measurement period, and the voice detection threshold (Vt
Since h) is set to a value that is appropriate to the noise level, highly accurate voice detection can always be performed without being affected by the noise level.

In addition, in claim (2), if it is found that noise has been detected incorrectly or if the end of a speech section is detected, a new noise is detected when the next start of the speech section is detected. The level is measured and the voice detection threshold (Vtr+
), it is possible to maintain voice detection accuracy even if the noise level changes somewhat during voice detection processing.

Furthermore, in claim (3), the value to be compared with the envelope waveform level difference value when detecting the start of speech is also adapted to the noise level at that time, so that the possibility of erroneously detecting noise as the start of speech is kept low. At the same time, it is possible to improve the configuration sensitivity of the voice start end.

Furthermore, in claims (4) and (5), even if noise is mistakenly detected as a voice start, the envelope waveform change after the start edge is detected or after the voice detection threshold (V th) has been exceeded is the same as that of the voice. If it is determined that the section is not a voice section because it differs from the envelope waveform change of the section, the section is determined to be a silent section, so the value to be compared with the envelope waveform level difference value and the voice detection threshold are set when inspecting the start of the voice. Even if the value is set to a small value and the voice detection sensitivity is increased, the probability that noise will be mistakenly detected as voice can be lowered.

Furthermore, in claim (6), when the envelope waveform level momentarily decreases during a voice section and falls below the voice detection threshold, the voice detection threshold is not set anew, but is set as before. Since the value of V th is used as is, it is possible to avoid a situation where the voice detection threshold (V th ) gradually increases during the voice section and the voice at the end of the sentence is not detected as a result.

[Brief explanation of the drawing]

FIG. 1 is a block configuration diagram for explaining an embodiment of the voice detection method of the present invention, FIG. 2 is a diagram showing an example of the configuration of an envelope waveform extraction circuit, and FIG. 3 is a diagram illustrating the processing flow in FIG. 1. FIG. 4 is a block configuration diagram showing an application example of the present invention, FIG. 5 is a diagram showing the processing flow in FIG. 4, FIG. 6 is a diagram showing the operation flow of an interactive answering machine, and FIG. Figures (a), (b), and (c) show measurement of user voice length and confirmation of user reaction to response message 1
FIG. 6 is a diagram showing a processing flow when detecting the end of generation. In the figure, 1 is an envelope waveform extraction circuit, 2 is an A/D converter, 3 is a control unit, 4.5 is a waveform register, 6 is a threshold register, 7.15 is a timer, 8 is a sound counter, 9 is the silence counter, 10°18 is the output status register,
11. 19 is an invalid voice section length register, 12 is a voice analyzer, 13 is a control section, 14 is a voice storage section, 16 is an address pointer, and 17 is a voice detection circuit. Figure 4 Figure 7 (b) Figure 7 <c>

Claims (6)

[Claims] (1) Extract the envelope waveform of the input signal through a rectifier circuit and low-pass filter, compare the magnitude relationship between this envelope waveform level and the voice detection threshold using internal processing of the CPU, and determine whether the envelope waveform level is the voice detection threshold. In the voice detection circuit, which determines the portion where the envelope waveform level exceeds the threshold as a voice section and the portion where the envelope waveform level is below the voice detection threshold as a silent section, the envelope waveform level read one sampling period ago is stored. Wait, C
The envelope waveform level newly read by PU is set to 1.
Compared with the envelope waveform level read before the sampling period, the newly read level is a constant value (V_d_i
The rising point of the envelope waveform that increases by more than _f) is regarded as the starting point of the voice section, and the envelope waveform level read one sampling period before is set to a constant value (V_o_f_f_−_a) on a logarithmic scale.
) is set as a voice detection threshold value (V_t_h). (2) The envelope waveform level is the voice detection threshold (V
_t_h) is exceeded, the voice detection threshold (V_t
_h), the detection of the rise of the envelope waveform is started again, and the point at which the rise of the envelope waveform is detected is regarded as the start point of a new audio section, and the envelope waveform level read one sampling period before is detected. is a constant value (V_o_
Claim (1) characterized in that a value obtained by adding f_f_−_a) is set as a new voice detection threshold value (V_t_h).
) Voice detection method described. (3) Add a constant value (V_o
_f_f_-_b) is added to a constant value (V_d_
The voice detection method according to claim 1, wherein the voice detection method is: i_f). (4) A certain period of time (T
If the envelope waveform level does not exceed the voice detection threshold (V_t_h) within ____d_s_■),
The section from the rising edge to that point is considered a silent section,
The detection of the rising edge of the envelope waveform is started again, and the point at which the rising edge of the envelope waveform is detected is regarded as the starting point of a new audio section, and a constant value (( 2. The voice detection method according to claim 1, wherein a value obtained by adding V_o_f_f_-_a) is set as a new voice detection threshold value (V_t_h). (5) The envelope waveform level rises and within a certain period of time (T_■_d_s_■) the voice detection threshold (V_
t_h), the minimum voice duration time (T_v_
The voice detection according to any one of claims (1) to (4), characterized in that if the voice detection threshold (V_t_h) falls below the voice detection threshold (V_t_h) again within c), this section is regarded as a silent section. Method. (6) The envelope waveform rises for a certain period of time (T_
After the envelope waveform level exceeded the voice detection threshold (V_t_h) within ■_d_s_■), the envelope waveform level fell below the voice detection threshold (V_t_h) again after the minimum voice duration time (T_v_c) had elapsed. In this case, the value of the voice detection threshold (V_t_h) is stored, and then the minimum duration of silence (T_n_v_
If the rising edge of the envelope waveform is detected again within c), the stored value is used as the voice detection threshold (
The voice detection method according to any one of claims (1) to (5), characterized in that: V_t_h).