JP6531449B2 - Voice processing apparatus, program and method, and exchange apparatus - Google Patents

Description

  The present invention relates to a voice processing device, a program and method, and an exchange device, and can be applied to, for example, a voice detection device that determines a voiced section and a silent section of a voice signal.

  Conventionally, in a voice processing apparatus that processes voice signals such as a telephone terminal and a switchboard, when processing voice signals, the voiced section where a person (speaker) is speaking is distinguished from the silent section where it is not speaking ) Processing is performed using the function of the presence detection. Conventionally, the voice detection function in voice processing is also called VAD (Voice Activity Detection). In the conventional speech processing apparatus, the function of speech detection is required in various speech signal processing, such as AGC (Automatic Gain Control), noise suppressor, speech recognition, and the like.

  In the conventional voice processing apparatus, the simplest method for realizing the function of the presence detection is a method of checking the power per unit time. In general, power is strong in the sound section and weak in the silent section. Therefore, in the conventional voice processing apparatus, the power per unit time can be determined, and the presence / absence of sound can be determined with a certain degree of accuracy that is to be compared with an appropriate threshold value. However, even a voice in a silent section includes background noise, and therefore, a silent section does not mean that a silent voice is completely captured. Therefore, in the conventional voice processing apparatus, when the presence / absence of sound is determined based on the power per unit time, it is necessary to set a threshold value larger than the power of the background noise. However, depending on the environment in which speech is captured, the power of background noise and the power of the signal (the target sound signal) may change dynamically.

  As a prior art which respond | corresponds to the above problems, there exists a description technique of patent document 1, 2. In Patent Documents 1 and 2, a change in background noise is estimated, and a threshold value for determining presence / absence of sound is dynamically changed based on the estimation result. Further, in the techniques described in Patent Documents 1 and 2, power per unit time is accumulated for a certain period of time, and is represented by a histogram to statistically estimate the power of background noise.

Japanese Examined Patent Publication 1-14599 Patent No. 3255584

  However, in the techniques described in Patent Documents 1 and 2, in an environment with a low S / N, the threshold for performing the on / off determination may be larger than the power of the signal, and accurate voice detection can not be performed. is there.

  In view of the above problems, a voice processing device, a program and method, and an exchange device capable of performing voice detection with higher accuracy even in an environment with large background noise power are desired.

  A voice processing apparatus according to a first aspect of the present invention comprises (1) level value calculating means for calculating a level value of an input voice signal for each frame in a predetermined time unit, and (2) a level value calculated by the level value calculating means. , Frequency counting means for counting the appearance frequency for each level value, (3) level value estimation means for estimating the background noise level value and the target sound signal level value from the appearance frequency for each level value, and (4) background And determining means for performing a determination process of determining a voiced section or a silent section of the input audio signal based on the estimated value of the noise level value and the estimated value of the target sound signal level value.

  The speech processing program according to a second aspect of the present invention comprises a computer, (1) level value calculating means for calculating a level value of an input sound signal for each frame of a predetermined time unit, and (2) the level value calculating means And (3) a level value estimation means for estimating a background noise level value and a target sound signal level value from the appearance frequency for each level value. 4) It is characterized in that it functions as a determination unit that performs a determination process of determining a voiced section or a silent section for the input audio signal based on the estimated value of the background noise level value and the estimated value of the target sound signal level value.

  According to a third aspect of the present invention, in the speech processing method performed by the speech processing apparatus, (1) level value calculation means, frequency counting means, level value estimation means, determination means, and (2) the level value calculation means The level value of the input audio signal is calculated for each frame in a predetermined time unit, and (3) the frequency counting means counts the appearance frequency for each level value for the level value calculated by the level value calculating means 4) The level value estimation means estimates the background noise level value and the target sound signal level value from the appearance frequency for each level value, and (5) the determination means estimates the background noise level value and the target sound signal A determination process is performed to determine a voiced section or a silent section of the input voice signal based on the estimated value of the level value.

  The exchange apparatus according to the fourth aspect of the present invention is (1) an exchange process of voice communication between a plurality of terminals, wherein the level of the voice signal to be transmitted to the terminal or the voice signal received from the terminal is desired. (2) The exchange processing means uses the audio processing apparatus of the first invention to adjust the level of the audio signal transmitted to the terminal or the level of the audio signal received from the terminal Are adjusted to a desired level.

  According to the present invention, it is possible to realize a voice processing device, a program and method, and a switching device capable of performing sound presence detection with higher accuracy even in an environment where power of background noise is large.

It is the block diagram shown about the functional composition of the speech processing unit (voice presence detection device) concerning a 1st embodiment. It is the graph shown about the histogram (frequency distribution) hold | maintained by the frequency counting part which concerns on 1st Embodiment. It is the graph shown about the histogram smooth | blunted by the level determination part which concerns on 1st Embodiment. It is the graph which showed the convexity of the histogram which was digitized by the level judgment part concerning a 1st embodiment. It is the block diagram shown about the functional composition of the speech processing unit (head detector) concerning a 2nd embodiment. It is the block diagram shown about the functional composition of the speech processing unit (background noise reduction device) concerning a 3rd embodiment. It is the block diagram shown about the functional composition of the speech processing unit (adaptive gain control device) concerning a 4th embodiment. It is the block diagram shown about the functional composition of the speech processing unit (speech processing unit provided with a jitter buffer) concerning a 5th embodiment. It is the block diagram shown about the functional composition of the exchange concerning a 6th embodiment.

(A) First Embodiment Hereinafter, a first embodiment of a speech processing apparatus, program and method according to the present invention will be described in detail with reference to the drawings. In this embodiment, an example will be described in which the voice processing device, program and method of the present invention are applied to a sound presence detection device.

(A-1) Configuration of First Embodiment FIG. 1 is a block diagram showing the entire configuration of a voice detection apparatus 1 of this embodiment.

  When the voice signal is input, the voice detection apparatus 1 detects a voice section of the voice signal, and outputs the result.

  Although there is no limitation on the format (data format) of the audio signal input to the presence detection device 1, for example, various data formats such as a PCM (Pulse Code Modulation) format can be applied. In this embodiment, it is assumed that a frame in which PCM format voice data of 10 msec worth is stored is input to the voice presence detection device 1 as a voice signal. That is, it is assumed that voice data in frame units is supplied to the voice presence detection device 1 as an input voice signal. The sampling frequency and the bit rate of the audio signal (audio data) input to the voice presence detection device 1 are not limited. In the example of this embodiment, it is assumed that audio data having a sampling frequency of 8 kHz, 16-bit PCM, and monaural is input to the voice presence detection device 1. When voice data (for example, voice data such as ITU-T G. 711) encoded by a predetermined codec is input to the voice presence detection device 1, a component for performing decoding processing is added. You may do so.

  Further, the signal format (data format) output by the presence detection device 1 is not limited. For example, the speech detection apparatus 1 outputs one of a signal (for example, “1” or “True”) indicating a speech zone and a signal “for example,“ 0 ”or“ False ”indicating a silence zone. May be

  Next, the internal configuration of the noise detection apparatus 1 will be described.

  The voice presence detection device 1 includes a high frequency transmission filter (hereinafter referred to as “HPF”) 10, a level calculation unit 11, a frequency counting unit 12, a level estimation unit 13, and a voice presence determination unit 14.

  The HPS 10 performs a filtering process to attenuate the power of a low frequency (low frequency band) component (a component of a frequency lower than a predetermined frequency) included in the input audio signal. Background noise often includes relatively large power in the low band. Therefore, it is possible to improve the S / N ratio of the audio signal applied to the sound detection by attenuating the low frequency component by the HPS 10 prior to the sound detection. The frequency band to be attenuated by the HPS 10 is not limited. In the HPS 10, for example, processing of attenuating components of 300 Hz or less may be performed. Hereinafter, the signal output from the HPS 10 (a signal whose low frequency component is attenuated) is also referred to as an input sound signal x. The presence detection apparatus 1 may not have the HPF 10 mounted thereon. When the HPF 10 is not mounted on the voice presence detection device 1, the voice signal (frame) itself input to the voice presence detection device 1 is processed as the input voice signal x.

  The level calculator 11 calculates an audio level (power level) of the input audio signal x. The level calculation unit 11 of this embodiment performs processing of calculating the audio level of the frame for each frame. A specific example of the sound level calculation process performed by the level calculation unit 11 will be described later.

  The frequency counting unit 12 counts the distribution (frequency distribution) of the level (power) in the input audio signal x. Specifically, the frequency counting unit 12 holds, as a histogram (frequency distribution), the number of occurrences of each level (the level calculated by the level calculation unit 11) in the input audio signal x. In this embodiment, the frequency counting unit 12 includes a counter unit 121 including counters corresponding to respective levels of the input audio signal x. In FIG. 1, the counter unit 121 is illustrated as having N + 1 counters CT (CT_0 to CT_N) (N is an arbitrary integer). The frequency counting unit 12 performs a process of incrementing (adding 1) the counter CT corresponding to the level in the counter unit 121 each time one level is calculated by the level calculation unit 11. The number of counters CT included in the counter unit 121, the interval of corresponding levels, and the like are not limited.

  In the counter section 121 of this embodiment, it is assumed that the counter CT is set every 1 [dB]. For example, it is assumed that the counter CT_0 corresponds to M [dB] (M is an arbitrary integer). Specifically, counters CT_0, CT_1, CT_2,..., CT_N correspond to M [dB], M + 1 [dB], M + 2 [dB],..., M + N [dB] respectively (corresponding to 1 dB increments) ) Are described. That is, the counter unit 121 can hold a histogram (frequency distribution) for levels within the range of M [dB] to M + N [dB]. In this embodiment, the counter unit 121 is described as capable of holding a histogram (frequency distribution) in 1 dB increments in the range of 10 dB to 70 dB.

  As described above, in the frequency counting unit 12, the counter unit 121 holds a histogram obtained by counting the number of appearances of each level in the input audio signal x. Hereinafter, the histogram (frequency distribution) held by the counter unit 121 will be referred to as a histogram H. Furthermore, in the following, the counter value of the counter CT corresponding to an arbitrary level v is represented as H (v).

  Based on the histogram H held by the counter unit 121, the level estimation unit 13 determines the level of background noise (hereinafter referred to as "background noise level") included in the input speech signal and the level of speech (that is, target sound). A process of estimating (hereinafter referred to as "signal level") is performed.

  The sound determination unit 14 performs processing to determine whether a frame currently being processed (a frame acquired most recently) is a sound section or a silent section from the background noise level and the signal level estimated by the level estimation unit 13. Then, the sound presence determining unit 14 performs a process of outputting the content (either a signal indicating a sound interval or a signal indicating a sound interval) according to the determination result.

(A-2) Operation of the First Embodiment Next, an example of the specific operation (voice processing method according to the embodiment) of the voiced detection device 1 of the first embodiment having the configuration as described above will be described. Do.

  In the presence detection device 1, when voice data for one frame is input, the HPF 10 first performs high-pass filter processing (processing to attenuate the power of frequency components lower than a predetermined frequency). The HPF 10 outputs the processed audio signal (frame) as an input audio signal x.

  The level calculator 11 calculates the power for each frame of the input audio signal x. For example, the level calculation unit 11 may perform level calculation by logarithmically converting the power of the input audio signal x for one frame. When the level calculation unit 11 calculates the level, the power to be the reference point (0 dB) may be appropriately set. Further, the level calculation unit 11 may calculate the level of the current frame based on the moving average of the sound levels of the past frames. As a result, the level calculation unit 11 can suppress fine fluctuations in level between frames.

  The frequency counting unit 12 increments a counter CT corresponding to the level calculated by the level calculation unit 11. As a result, the frequency counting unit 12 updates the histogram H held by the counter unit 121.

  At this time, the frequency counting unit 12 performs processing to round the level calculated by the level calculation unit 11 according to a predetermined method. Then, the frequency counting unit 12 increments the counter CT corresponding to the rounded value (level). As described above, in the counter section 121 of this embodiment, the counter CT is set to have a width of 1 dB. Therefore, the frequency counting unit 12 may, for example, round to 10 dB for a level of 9.5 dB or more and less than 10.5 dB, and increment the counter CT corresponding to 10 dB.

  As described above, the counter unit 121 of the frequency counting unit 12 holds the histogram H by the counter CT set in 1 dB steps.

  FIG. 2 is a graph showing the histogram H held by the frequency counting unit 12 (counter unit 121). The graph shown in FIG. 2 shows the histogram H held by the frequency counting unit 12 (counter unit 121) when an audio signal is actually input to the voice presence detection device 1.

  In the graph of FIG. 2, the horizontal axis indicates the level of the input speech signal x, and the vertical axis indicates the number of occurrences of each level (the value of the counter CT of each level).

The level estimation unit 13 estimates the background noise level and the signal level included in the input audio signal x based on the histogram H held by the counter unit 121 at fixed time intervals. Furthermore, the presence determination unit 14 obtains a threshold used for the presence determination based on the estimated background noise level and the signal level.

  The voice determination unit 14 performs threshold calculation processing (threshold update processing) using, for example, the histogram H based on a frame (frame of the input sound signal x) for the latest predetermined time. The presence determination unit 14 may perform threshold calculation processing based on, for example, a frame (a frame of the input sound signal x) for the last 10 seconds. The timing at which the presence determination unit 14 performs the threshold value calculation process is not limited. The presence determination unit 14 may perform the presence determination, for example, every predetermined period (for example, every 10 seconds).

The timing at which the presence determination unit 14 performs threshold calculation processing, the number of samples of the histogram H held by the frequency counting unit 12, and the like are not limited. For example, sound determination unit 14, when performing the threshold value calculation processing for each predetermined period, initializing each counter CT of the counter unit 121 (the counter value reset to zero) may be performed a process of.

Next, an example of processing in which the level estimation unit 13 estimates the background noise level and the signal level will be described.

  As described above, the graph shown in FIG. 2 shows the histogram H held by the frequency counting unit 12 (counter unit 121) when an audio signal is actually input to the voice presence detection device 1. Then, in the histogram H shown in FIG. 2, the distribution of the actual signal level (distribution of the level of the sound interval) and the distribution of the background noise level (distribution of the level of the silent interval) were confirmed. Then, in the histogram H of FIG. 2, the first peak formed mainly by the distribution of the background noise level could be confirmed in the range of the levels B1 and B2. Further, in the histogram H of FIG. 2, the second peak formed mainly by the distribution of the signal level (the level of the sound interval) was confirmed in the range of levels B3 to B4 higher than the range of levels B1 to B2. .

  As described above, in the histogram H shown in FIG. 2, the first peak (peak in the range of levels B1 to B2) mainly formed by the distribution of the background noise level and the second peak mainly formed by the distribution of the signal level Peak (a peak in the range of levels B3 to B4 higher than the first peak) occurs. That is, the histogram H shown in FIG. 2 is a histogram including two peaks (bimodal).

  As a result of a plurality of experiments by the applicant, it has become clear that, in the histogram H held by the frequency counting unit 12, the occurrence of the above-mentioned two peaks is generally valid (reproducible) .

Therefore, the level estimation unit 13 of this embodiment detects the first peak mainly formed by the distribution of the background noise level and the second peak mainly formed by the distribution of the signal level, and the sound presence determination unit 14 will be made a sound judgment based on two peaks detected.

Next, an example of a specific procedure of the presence determination by the level estimation unit 13 and the presence determination unit 14 (a specific example of the presence determination based on the above-described two peaks) will be described.

Since the curve indicated by H (v) includes fine asperities, the level estimation unit 13 performs a smoothing process for the purpose of removing the asperities. The method of smoothing H (v) is not limited, but, for example, a method such as weighted averaging may be used.

Specifically, the level estimation unit 13 may perform smoothing of H (v) using the following equation (1). In the following equation (1), Hs (v) represents a value after H (v) is smoothed. The level estimation unit 13 performs smoothing processing by obtaining Hs (v) for each of all the levels constituting the histogram H.
Hs (v) = {H (v-2) + 2H (v-1) + 3H (v)
+ 2H (v + 1) + H (v + 2)} / 9 (1)

FIG. 3 is a graph showing the smoothing process of H (v) by the level estimation unit 13 .

  In FIG. 3, a graph (curve) before H (v) is smoothed is shown by a dotted line, and a graph (curve) after H (v) is smoothed is shown by a solid line.

Next, the level estimation unit 13 digitizes the convexity of Hs (v) in order to detect the above-mentioned two peaks from the smoothed Hs (v). The specific method of quantifying the convexity of Hs (v) in the level estimation unit 13 is not limited. In this embodiment, the level estimation unit 13 is described as quantifying the convexity of Hs (v) using the differentiated second-order differential value. Specifically, the level estimation unit 13 performs a process of digitizing the convexity of Hs (v) using the following equation (2). In the following equation (2), C (v) represents the convexity of Hs (v). The level estimation unit 13 digitizes the convexity by obtaining C (v) for each of the levels constituting the histogram H. A section in which C (v) is a positive value indicates that the shape is convex in the upward direction (positive direction).

Then, the level estimation unit 13 regards a section in which C (v) is positive as one peak, and searches for the peak over the entire section.

As described above, the histogram H usually has a first peak mainly formed by the background noise level distribution and a second peak mainly formed by the signal level distribution (a peak higher than the first peak). Distribution is bimodal). Therefore, the level estimation unit 13 can usually detect two peaks based on the convexity of the histogram H (Hs (v)). Then, the level estimation unit 13 regards the lower one of the two detected peaks as the first peak associated with the background noise level, and the second peak associated with the signal level with the higher peak. It shall be regarded as When only one peak can be found, the level estimation unit 13 may regard the peak as a peak related to the background noise level. In addition, when three or more peaks are detected, the level estimation unit 13 selects two in order from the one with the widest section (the one with the positive C (v) section), and the level is lower. The peak may be regarded as the first peak related to the background noise level, and the higher peak may be regarded as the second peak related to the signal level.
C (v) = Hs (v)-
{Hs (v-10) + Hs (v + 10)} / 2 (2)

  FIG. 4 is a graph in the case where the convexity at each level of Hs (v) shown in FIG. 3 is digitized (digitized based on the above equation (2)). In FIG. 4, a graph (curve) obtained by digitizing the convexity of Hs (v) is illustrated by a solid line, and a graph indicating Hs (v) is illustrated by a dotted line.

In the graph shown in FIG. 4, two sections (peaks) in which C (v) is a positive value are formed. Therefore, the level estimation unit 13 regards the lower one of the two peaks as the first peak related to the background noise level, and selects the higher peak as the second peak related to the signal level. It will be considered. Hereinafter, the section of the first peak (section including the first peak) related to the background noise level is referred to as a peak section PN. Moreover, below, the area (area containing a 2nd peak) of the 2nd peak which concerns on a signal level shall be called peak area PS.

  As shown in FIG. 4, the peak section PN related to the background noise level tends to be narrower than the peak section PS related to the signal level. Therefore, it can be said that the distribution of the second peak of the signal level tends to be larger than the distribution of the first peak of the background noise level.

Next, the level estimation unit 13 determines, for each of the peak sections PN and PS, a representative value in the section (a value to be applied to the presence determination). The method of determining the representative value of each peak section by the level estimation unit 13 is not limited. In this embodiment, the level estimation unit 13 determines the representative value of each peak section using the centroid method. For example, the level estimation unit 13 may determine the representative value of the peak interval PN based on the following equation (3). In the following equation (3), LvN is a representative value of the peak interval PN (estimated value of background noise level). Further, for example, the level estimation unit 13 may determine the representative value of the peak section PS based on the following equation (4). In the following equation (4), LvS is a representative value (estimated value of signal level) of the peak interval PS.
LvN = v v H (v) / H H (v) (v ∈ PN) (3)
LvS = v v H (v) / H H (v) (v ∈ PS) (4)

Next, a process of determining a threshold based on the background noise level LvN and the signal level LvS estimated by the presence determination unit 14 will be described.

  The sound determination unit 14 determines whether the current processing frame is a sound section or a silent section using the estimated value LvN of the background noise level and the estimated value LvS of the signal level. Here, the noise determination unit 14 uses the estimated value LvN of the background noise level and the estimated value LvS of the signal level to determine a threshold LvTh to be compared with the frame level Lv of the current processing frame. Here, it is assumed that the noise determination unit 14 obtains the threshold LvTh using the following equation (5). In the following equation (5), α is a coefficient set to any value between 0 and 1 (0 ≦ α ≦ 1). For example, α may be a fixed (static) value (for example, a value on the order of 0.5), but may be varied dynamically.

If the noise determination unit 14 finds only one peak from the histogram H, it updates only the estimated value LvN of the background noise level to a value based on the latest histogram H, and the estimated value LvS of the signal level. The threshold LvTh may be obtained by continuously using the previously calculated one.
LvTh = αLvN + (1-α) LvS (5)

  In the example of this embodiment, the voice determination unit 14 compares the frame level Lv of the currently processed voice frame (for example, the voice frame input most recently) with the threshold LvTh, and generates the voice of the voice frame. A process of determining (determining whether it is an audio frame belonging to a sound interval or an audio frame belonging to a silent interval) is performed. Specifically, the voice determination unit 14 determines that the voice frame belongs to a voiced section if Lv> = LvTh, and determines that the voice frame belongs to a silent segment if Lv <LvTh.

(A-3) Effects of the First Embodiment According to the first embodiment, the following effects can be achieved.

  In the speech detection apparatus 1 according to the first embodiment, from the histogram H, a section of a first peak mainly formed by the distribution of background noise level and a section of a second peak mainly formed by the distribution of signal level Is performed to estimate the background noise level and the signal level. Then, in the voice presence detection device 1 of the first embodiment, the threshold used for the voice presence determination is estimated using both the background noise level and the signal level. In the prior art, since only the background noise level is estimated, there is a case where an appropriate threshold can not be set in a state where the S / N ratio is bad. However, in the speech detection apparatus 1 according to the first embodiment, since both the background noise level and the signal level are estimated from the histogram H to set an appropriate threshold, the S / N ratio is bad. Also, it becomes possible to perform threshold setting more appropriate than before. That is, in the speech detection apparatus 1 according to the first embodiment, the speech detection can be performed more stably than in the related art.

(B) Second Embodiment Hereinafter, a second embodiment of the speech processing device, program and method according to the present invention will be described in detail with reference to the drawings. In this embodiment, an example in which the speech processing device, program and method of the present invention are applied to a speech detection device will be described.

  FIG. 5 is an explanatory view showing a functional configuration of the head detection apparatus 100 of the present invention.

  The head detection apparatus 100 is an apparatus (an apparatus having a function of a talk head detection unit) that detects a talk head (a start time of sound) from an input voice signal. For example, the talkhead detection apparatus 100 may be implemented as software by installing a program (including the speech processing program according to the embodiment) on a computer including a processor and a memory. The head detection apparatus 100 may be incorporated, for example, in an apparatus that performs voice processing, such as a telephone terminal.

  The speech head detection apparatus 100 includes a speech detection unit 101 and a speech head detection unit 102.

  The voice presence detection unit 101 performs voice presence detection on the input audio signal. In this embodiment, it is assumed that the voice detection apparatus 1 of the first embodiment is applied as the voice detection unit 101. The presence detection unit 101 outputs the presence determination or the silence determination at predetermined intervals based on the input audio signal.

  The talker detection unit 102 detects a talker of the input audio signal based on the detection result of the voice detection unit 101. The speech head detection unit 102 also outputs a detection signal indicating that the speech head is detected at the timing when the judgment result of the speech detection unit 101 transits from silence judgment to speech judgment, and detects the speech head at other timings. Output a non-detection signal indicating that there is no.

(C) Third Embodiment Hereinafter, a third embodiment of the speech processing device, program and method according to the present invention will be described in detail with reference to the drawings. In this embodiment, an example in which the speech processing apparatus, program and method of the present invention are applied to a background noise reduction apparatus (noise suppressor) will be described.

  FIG. 6 is an explanatory view showing a functional configuration of the background noise reduction device 200 of the present invention.

  The background noise reduction device 200 is a device that reduces background noise by lowering the voice level in a silent section and outputting the input voice signal (a device that takes on the function of background noise reduction means). For example, the background noise reduction apparatus 200 may be implemented as software by installing a program (including the voice processing program according to the embodiment) on a computer including a processor and a memory. The background noise reduction device 200 may be incorporated into a device that performs voice processing, such as a telephone terminal, for example.

  The background noise reduction device 200 includes a voice presence detection unit 201, an audio frame buffer 202, a voice presence / non-speech determination buffer 203, a determination rewriting unit 204, and a gain superposition unit 205.

  The sound detection unit 201 detects sound of the input audio signal. In this embodiment, it is assumed that the voice detection apparatus 1 of the first embodiment is applied as the voice detection unit 201. The presence detection unit 201 outputs the presence determination or the silence determination at predetermined intervals based on the input audio signal.

  The voice frame buffer 202 buffers frames of the input voice signal for a fixed time.

  The speech / non-speech determination buffer 203 buffers the determination result of the speech presence detector 201 for a fixed period.

  The judgment rewriting unit 204 refers to the judgment result of the speech / non-speech judgment buffered in the speech / non-speech judgment buffer 203 and detects the speech / silence when a change from the speech judgment to the speech judgment is detected. With regard to the determination result of the sound / non-speech determination accumulated in the determination buffer 203, a process is performed in which the fixed time in the past is traced back and the non-speech determination is rewritten as the presence determination. The determination rewrite unit 204 is disposed for the purpose of preventing a dead end in the audio signal output from the background noise reduction device 200.

  The gain superimposing unit 205 acquires and outputs an arbitrary voice frame (for example, the oldest voice frame) from the speech / non-speech determination buffer 203. When outputting a frame, the gain superimposing unit 205 refers to the determination result of the sound / non-voice determination corresponding to the frame (refer to the determination result stored in the voice / non-voice determination buffer 203), and When the frame is a silent section, processing is performed to lower the sound level of the frame (processing to adjust the gain) and then output. The gain superimposing unit 205 performs processing for outputting a frame in a sound interval as it is.

(D) Fourth Embodiment Hereinafter, a fourth embodiment of the speech processing device, program and method according to the present invention will be described in detail with reference to the drawings. In this embodiment, an example in which the voice processing device, program and method of the present invention are applied to an adaptive gain controller (AGC) will be described.

  FIG. 7 is an explanatory view showing a functional configuration of the adaptive gain control apparatus 300 of the fourth embodiment.

  The adaptive gain control device 300 is a device (a device that takes on the function of a gain control unit) that adjusts and outputs the input voice signal to a desired voice level (a predetermined constant level). For example, the adaptive gain control apparatus 300 may be implemented as software by installing a program (including the voice processing program according to the embodiment) on a computer including a processor and a memory. The adaptive gain control device 300 may be incorporated, for example, in a device that performs voice processing, such as a telephone terminal.

  The adaptive gain control apparatus 300 includes a sound presence detection unit 301, a level calculation unit 302, a gain determination unit 303, and a gain superposition unit 304.

  The voice presence detection unit 301 performs voice presence detection on the input audio signal. In this embodiment, it is assumed that the voice detection apparatus 1 of the first embodiment is applied as the voice detection unit 301. The presence detection unit 301 outputs the presence determination or the silence determination at predetermined intervals based on the input audio signal.

  The level calculator 302 calculates the level of the input signal.

  The gain determination unit 303 also determines the gain to be superimposed based on the level of the input signal calculated by the level calculation unit 302. Further, the gain determination unit 303 takes into consideration the detection result (the result of the presence / absence determination) of the presence detection unit 301, and the gain to be superimposed on the input signal (gain for setting the input signal to a desired level) Decide. For example, the gain determination unit 303 performs a process of determining a gain that does not amplify background noise in a silent section (a section in which a silence determination is detected by the noise detection unit 301).

  The gain superposition unit 304 superimposes the gain determined by the gain determination unit 303 on the input signal and outputs the resultant signal. The level of the audio signal output by the gain superimposing unit 304 is a desired level set in advance.

(E) Fifth Embodiment Hereinafter, a fifth embodiment of the speech processing device, program and method according to the present invention will be described in detail with reference to the drawings. In this embodiment, an example in which the speech processing apparatus, program and method of the present invention are applied to a speech processing apparatus provided with a jitter buffer will be described.

  For example, the voice processing apparatus 400 may be implemented as software by installing a program (including the voice processing program according to the embodiment) on a computer including a processor and a memory. The voice processing device 400 may be incorporated into, for example, a device that performs voice processing, such as a telephone terminal.

  FIG. 8 is an explanatory view showing a functional configuration of the speech processing device 400 according to the fifth embodiment.

  The voice processing device 400 includes a voice detection unit 401, a jitter buffer 402, and a PCM decoding unit 403.

  The jitter buffer 402 suspends (buffers) a voice packet (packet in which a voice frame is inserted) arriving via the IP network N by the jitter buffer 402 and absorbs jitter (fluctuation) to suspend voice at a predetermined interval. A packet is output. The jitter buffer 402 includes a storage buffer 402 a that stores (holds) voice packets, and a jitter buffer control unit 402 b that controls processing (for example, discarding) of the voice packets in the storage buffer 402 a.

  In this embodiment, it is assumed that voice packets in the form of RTP (Real-Time Transport Protocol) are input to the jitter buffer 402.

  The jitter buffer 402 (storage buffer 402 a) reads the sequence number in the voice packet (packet in RTP format) arriving from the IP network N, and stores the voice packet in ascending order of the sequence number. Also, the jitter buffer 402 (storage buffer 402 a) outputs voice packets stored in ascending order of sequence numbers.

  When the amount (number) of voice packets stored in the storage buffer 402a is equal to or more than a threshold, the jitter buffer control unit 402b performs processing to discard some voice packets and reduce the storage amount.

  The PCM decoding unit 403 decodes the payload (encoded audio data) of the audio packet supplied from the storage buffer 402 a. The PCM decoding means 403 is, for example, an ITU-T G.I. The payload of the audio packet is decoded according to a predetermined codec such as 711 to obtain decoded audio data (for example, a frame of audio data in PCM format).

  The voice presence detection unit 401 performs voice presence detection on the input audio signal (frame of voice data output from the PCM decoding unit 403). In this embodiment, it is assumed that the voice detection device 1 of the first embodiment is applied as the voice detection unit 401. The sound presence detection unit 401 outputs sound presence determination or silence determination at predetermined intervals based on the input audio signal.

  The jitter buffer control unit 402b determines whether to discard the voice packet of the storage buffer 402a (determination of processing timing for discarding the voice packet) in consideration of the determination result of the voice presence detection unit 401. For example, the jitter buffer control unit 402b may determine that the voice packet in the storage buffer 402a is to be discarded only while the noise detection unit 401 outputs the silence determination. As a result, in the jitter buffer 402, it is possible to suppress the influence on voice due to voice packet discarding (for example, deterioration of decoded voice on the decoding side).

(F) Sixth Embodiment The sixth embodiment of the speech processing device, program and method according to the present invention, and the exchange device will be described in detail with reference to the drawings. In this embodiment, an example in which the voice processing device, program and method of the present invention are applied to a switching device will be described.

  FIG. 9 is an explanatory view showing a functional configuration of the exchange device 500 of the sixth embodiment.

  The exchange apparatus 500 is connected to a plurality of IP telephone terminals 600 via an IP network N, and performs call control processing and media communication processing (processing of media data such as voice data) between the IP telephone terminals 600 (so-called so-called It is a device that takes on the function of IP-PBX.

  For example, the switching device 500 may be implemented as software by installing a program (including the voice processing program according to the embodiment) on a computer including a processor and a memory.

  In the example of this embodiment, the switching device 500 includes a call control unit 501 and a media processing unit 502. In other words, the switching device 500 includes the call control unit 501 and the media processing unit 502 as components for realizing the switching processing means. The switching method (call control processing method, media data processing method, etc.) in the switching device 500 is not limited.

  The call control unit 501 performs call control processing between the IP telephone terminals 600. The call control unit 501 performs call control processing in accordance with a call control protocol such as, for example, SIP (Session Initiation Protocol).

  The media processing unit 502 performs media communication processing (processing of media data such as voice data) with the IP telephone terminal 600. Media processing unit 502 receives and processes voice data (voice packet) from IP telephone terminal 600, and performs processing such as transmitting to another IP telephone terminal 600. The media processing unit 502 includes an adaptive gain control unit 503. The adaptive gain control unit 503 sets a desired voice level for a voice signal (voice data) based on a voice packet received from the IP telephone terminal 600 or a voice signal (voice data) to be inserted into a voice packet to be transmitted to the IP telephone terminal 600. A process of adjusting to (a predetermined constant level) is performed. In this embodiment, it is assumed that the adaptive gain control device 300 of the third embodiment is applied as the adaptive gain control unit 503. That is, the switching device 500 processes an audio signal to be transmitted or received using the adaptive gain control device 300 of the third embodiment.

(G) Other Embodiments The present invention is not limited to the above-described embodiments, and may include modified embodiments as exemplified below.

  (G-1) The voice processing device of the present invention (the voiced detection device of the first embodiment) is a voice processing device other than the devices exemplified in the above embodiments (for example, a telephone terminal, a conference terminal , Voice recording apparatus etc.).

  (G-2) In the voice determination unit 14 of the first embodiment, when the frame level Lv is close to the threshold LvTh, the voice determination and the silence determination alternate in a short time (a so-called flapping state). there is a possibility. Therefore, in order to prevent the above-mentioned state, the sound determination unit 14 always determines that a frame for a predetermined time period after that is a sound interval (a so-called "hangover") when a frame is determined to be a sound interval. "Function" may be performed. For example, a time of about 500 msec may be applied as the above-mentioned predetermined time.

  DESCRIPTION OF SYMBOLS 1 ... existence detection apparatus (voice processing apparatus), 10 ... HPF, 11 ... level calculation part, 12 ... frequency counting part, 12 ... 1 counter part, 13 ... level estimation part, 14 ... existence determination part.

Claims (10)

Level value calculation means for calculating the level value of the input speech signal for each frame of a predetermined time unit;
Frequency counting means for counting the frequency of appearance of each level value for the level value calculated by the above-mentioned level value calculating means;
Level value estimation means for estimating a background noise level value and a target sound signal level value from an appearance frequency for each level value;
Speech processing apparatus characterized by further comprising: determination means for performing determination processing for determining a voiced section or a silent section of an input speech signal based on an estimated value of a background noise level value and an estimated value of a target sound signal level value. .   The determination method according to claim 1, wherein the determination means calculates a threshold based on the estimated value of the background noise level value and the estimated value of the target sound signal level value, and performs the determination process using the calculated threshold. Voice processing device.   The above-mentioned level value estimating means detects the section of the peak of two appearance frequency in the appearance frequency for each level value, and estimates the background noise level value based on the appearance frequency of the low level first section of the two sections. 3. The speech processing apparatus according to claim 1, wherein the target sound signal level value is estimated based on the appearance frequency of the high level second section of the two sections.   The speech processing apparatus according to any one of claims 1 to 3, further comprising a speech head detection means for detecting a speech head in the input speech signal by using the judgment result of the judgment means.   The speech processing apparatus according to any one of claims 1 to 3, further comprising background noise reduction means for reducing background noise from the input speech signal using the determination result of the determination means.   The voice according to any one of claims 1 to 3, further comprising gain control means for adjusting the level of the input voice signal to a desired level in consideration of the determination result of the determination means. Processing unit. The input speech signal comes from the network in frame units,
A jitter buffer for holding frames coming from the network and outputting the frames at predetermined intervals;
The jitter buffer control means is for controlling the jitter buffer, and further includes a jitter buffer control means for discarding a frame reserved in the jitter buffer at a timing taking into consideration the determination result of the determination means. The voice processing apparatus according to any one of claims 1 to 3. Exchange processing of voice communication between a plurality of terminals, and having exchange processing means for adjusting a level of a voice signal to be transmitted to the terminal or a voice signal received from the terminal to a desired level;
The exchange processing means adjusts the level of an audio signal to be transmitted to the terminal or an audio signal received from the terminal to a desired level using the speech processing apparatus according to claim 6. apparatus. Computer,
Level value calculation means for calculating the level value of the input speech signal for each frame of a predetermined time unit;
Frequency counting means for counting the frequency of appearance of each level value for the level value calculated by the above-mentioned level value calculating means;
Level value estimation means for estimating a background noise level value and a target sound signal level value from an appearance frequency for each level value;
A voice characterized in that it functions as determination means for performing a determination process of determining a voiced section or a silent section for an input voice signal based on an estimated value of a background noise level value and an estimated value of a target sound signal level value. Processing program. In the speech processing method performed by the speech processing device,
Level value calculation means, frequency counting means, level value estimation means, determination means
The level value calculating means calculates the level value of the input speech signal for each frame of a predetermined time unit,
The frequency counting means counts the appearance frequency for each level value for the level value calculated by the level value calculating means,
The level value estimation means estimates a background noise level value and a target sound signal level value from the appearance frequency for each level value,
The voice processing method characterized in that the determination means performs a determination process of determining a voiced section or a silent section for the input voice signal based on the estimated value of the background noise level value and the estimated value of the target sound signal level value. .

Images