JP2018180482A - Speech detection apparatus and speech detection program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce degradation of speech detection accuracy.SOLUTION: The watching terminal 10 includes a sound signal acquisition unit that acquires a sound signal; a sound candidate extraction unit that extracts a candidate for a section likely to sound as a sound section candidate based on a sound section of the time waveform of the sound signal, a feature amount calculation unit that calculates a feature amount indicating a degree of speech likeness of a frequency change in a speech section candidate, a speech speed calculation unit that calculates a speech speed in a speech section candidate, a correction unit for correcting the feature quantity according to the speech speed, and a speech detection unit for detecting the presence or absence of a speech in the speech section candidate based on the corrected feature quantity.SELECTED DRAWING: Figure 4

Description

  The present invention relates to a voice detection apparatus and a voice detection program.

  Solutions that use the Internet of Things (IoT) to watch residents, such as the elderly and the care recipient, are attracting attention. For example, voice detection for detecting a voice by speech from living sound collected by a microphone of a watching terminal installed in a resident's house is utilized for the safety confirmation of the resident.

  In such voice detection, as one aspect, the fluctuation of the frequency of the fundamental sound, which is one of the features indicating the voice likeness, that is, the so-called fundamental frequency is used. For example, Fourier transform such as FFT (Fast Fourier Transform) is applied to the input signal. As a result, as a result of converting the input signal from the time domain to the frequency domain, a power spectrum is calculated. Of the frequencies at which the power value has a peak on this power spectrum, the lowest frequency is extracted as the "fundamental frequency".

  A feature amount is calculated from the time series data of the fundamental frequency. For example, while shifting the time window having a predetermined analysis length, the slope of the approximate straight line is obtained by performing function approximation or the like by a linear function on the time series data of the fundamental frequency included in the time window. As described above, when the time window is shifted, the statistical value of the slope, for example, the average value, is calculated by dividing the slope of the approximate straight line obtained for each of the obtained slopes into a slope having a positive value and a slope having a negative value.

  The average value of these positive slopes and the average value of the negative slopes are used as feature quantities of fluctuation related to the fundamental frequency. For example, the mean value of the positive slope and the mean value of the negative slope are compared to the threshold. That is, the average value of the positive slope is included in the lower limit value of the fluctuation, for example, 0.37 Hz / ms and the upper limit value, for example 1.09 Hz / ms, and the average value of the negative slope is the lower limit value of the fluctuation, For example, when it is included in the range of −1.19 Hz / ms and an upper limit value, eg, −0.29 Hz / ms, it is detected that the input signal has voice.

JP 2012-133346 A JP, 2011-145326, A JP, 2014-013302, A

  However, in the above-described technique, the detection accuracy of speech may be reduced.

  That is, the threshold value to be compared with the feature value of the fluctuation of the fundamental frequency is merely set under the assumption that speech is performed at an average speech speed. For this reason, it is difficult to detect speech uttered at a rapid pace and speech uttered slowly.

  Even so, even if correction is performed to widen the range of the threshold to be identified as speech, false detections in which sounds other than speech are detected as speech increase. For example, to make it easy to detect speech uttered slowly, when lowering the lower limit value to be compared with the average value of positive slopes or raising the upper limit value to compare with the average value of negative slopes, As a result of difficulty in discrimination from low frequency sounds such as car engine sounds, false detection of non-voice low frequency sounds increases. In addition, when raising the upper limit value to be compared with the average value of positive slopes or lowering the lower limit value to be compared with the average value of negative slopes so as to make it easier to detect speech uttered at a rapid pace, As a result of making it difficult to distinguish music other than voice, false detection increases.

  In one aspect, the present invention aims to provide a voice detection device and a voice detection program that can suppress degradation in voice detection accuracy.

  In one aspect, the voice detection device includes a sound signal acquisition unit for obtaining a sound signal, and a voice candidate extraction unit for extracting a candidate of a section likely to be a voice as a voice section candidate based on a sound section of the time waveform of the sound signal. A feature amount calculation unit for calculating a feature amount indicating a degree of speech likeness of frequency change in the voice section candidate; a speech speed calculation unit for calculating a speech speed in the speech section candidate; and the feature quantity according to the speech speed And a voice detection unit that detects the presence or absence of voice in the voice section candidate based on the feature amount after correction.

  It is possible to suppress the decrease in the voice detection accuracy.

FIG. 1 is a view showing a configuration example of a remote care system according to a first embodiment. FIG. 2 is a diagram showing an example of correction of the threshold value. FIG. 3A is a diagram showing an example of a spectrogram. FIG. 3B is a diagram showing an example of a spectrogram. FIG. 3C is a diagram showing an example of a spectrogram. FIG. 4 is a block diagram of an example of a functional configuration of the watching terminal according to the first embodiment. FIG. 5 is a diagram showing an example of the time waveform of the fundamental frequency in the speech segment candidate. FIG. 6 is a diagram showing a graph of correlation coefficients in speech segment candidates. FIG. 7 is a diagram showing an example of a derived function of the correction coefficient α. FIG. 8 is a flowchart illustrating the procedure of the voice detection process according to the first embodiment. FIG. 9 is a diagram showing an example of a derived function of the correction coefficient β. FIG. 10 is a diagram illustrating an example of a derived function of the expansion coefficient γ. FIG. 11 is a diagram illustrating an example of a derived function of the expansion coefficient Δ. FIG. 12 is a diagram illustrating an example of a hardware configuration of a computer that executes a voice detection program according to the first embodiment and the second embodiment.

  A voice detection device and a voice detection program according to the present application will be described below with reference to the attached drawings. Note that this embodiment does not limit the disclosed technology. And each Example can be suitably combined in the range which does not make processing contents contradictory.

[Remote Care System]
FIG. 1 is a view showing a configuration example of a remote care system according to a first embodiment. The remote care system 1 shown in FIG. 1 is a remote care service for confirming the safety of a resident by voice detection that detects speech due to speech from the living sound collected by the microphone of the watching terminal 10 installed in the resident's house. To achieve

  As shown in FIG. 1, the remote care system 1 includes a watching terminal 10, a server device 30, and a related party terminal 50. Although the watching terminal 10 and the party terminal 50 are illustrated one by one in FIG. 1, any number of the watching terminal 10 and the party terminal 50 may be accommodated in the remote care system 1.

  The monitoring terminal 10, the server device 30, and the related party terminal 50 are connected via the network NW. This network NW can be constructed by any communication network, whether wired or wireless. Furthermore, in part of the network NW, a wired or wireless communication network may be mixed. For example, when the concerned party terminal 50 is implemented as a wireless communication device, the network NW includes relay devices such as the nearest base station and access point corresponding to the cell accommodating the wireless communication device.

  The watching terminal 10 is a terminal device that realizes a solution for watching a resident, for example, an elderly person or a care recipient.

  As one embodiment, watching terminal 10 is installed in a resident's house. For example, the watching terminal 10 has a microphone that converts sound into an electrical signal, a so-called microphone. Using this microphone, the watching terminal 10 executes voice detection processing for detecting voice from a sound signal input from the microphone. Then, the watching terminal 10 uploads the voice detection result to the server device 30. For example, uploading to the server device 30 can be performed at any timing such as when a voice detection process is executed, and a voice detection result is accumulated for a predetermined period or a predetermined number of times.

  The server device 30 is a computer that provides the above-described remote care service to the related party terminal 50.

  In one embodiment, the server device 30 accumulates voice detection results collected from the watching terminal 10. Then, the server device 30 executes various statistical processes on the voice detection result accumulated for a predetermined period, for example, every time the voice detection result is accumulated for a predetermined period, for example, one hour. For example, the server device 30 calculates the ratio of time with voice detection in one hour, or calculates the ratio of time without voice detection in one hour. Then, the server device 30 notifies the concerned person terminal 50 of the ratio of time with voice detection or the ratio of time without voice detection.

  Here, the server device 30 can also set a condition in the notification from the server device 30 to the terminal 50 of the person concerned. For example, the server device 30 outputs an alert to the terminal 50 when the percentage of time with voice detection is less than a predetermined threshold or when the percentage of time without voice detection is equal to or greater than a predetermined threshold. Can. In addition, the server device 30 can also output an alert to the concerned party terminal 50 when no voice is detected for a predetermined period. Here, although the case where an alert is output when safety is suspected is illustrated as one aspect to the last, notification can also be performed when safety is confirmed. For example, when the server device 30 includes the result with voice detection in a predetermined period, the percentage of time with voice detection is greater than or equal to a predetermined threshold, or the percentage of time without voice detection is less than a predetermined threshold If so, the concerned person's terminal 50 can be notified of the activity of the resident.

  The concerned party terminal 50 is a terminal device used by a resident related party. The "relevant person" mentioned here includes, by way of example only, the resident's family and relatives, the person in charge of the call center involved in the operation of the remote care service described above, the medical worker resident in the call center, etc. I do not mind.

  As one embodiment, not only mobile communication terminals such as smartphones, mobile phones and PHS (Personal Handyphone System) but also tablet terminals and slate terminals can be adopted as the related party terminal 50. As described above, the present invention is not limited to the handheld portable terminal device, and the related party terminal 50 may employ any IoT device other than a head mounted display, a wearable device such as a smart glass or a smart watch. In addition, although the example of the portable terminal device was given to the last, the concerned person terminal 50 may be a stationary information processing device, for example, a personal computer or the like.

[One aspect of improvement of voice detection processing]
Here, in the above-described speech detection process, fluctuation of the fundamental frequency, which is one of the features indicating speech likeness, is used. For example, feature quantities calculated from time series data of the fundamental frequency, for example, an average value of positive slopes and an average value of negative slopes are compared with a threshold. That is, the average value of the positive slope is included in the lower limit value of the fluctuation, for example, 0.37 Hz / ms and the upper limit value, for example 1.09 Hz / ms, and the average value of the negative slope is the lower limit value of the fluctuation, For example, when it is included in the range of −1.19 Hz / ms and an upper limit value, for example, −0.15 Hz / ms, it is detected as voice present in the input signal.

  However, as described in the background art section, the threshold value to be compared with the feature value of the fluctuation of the fundamental frequency is only set under the assumption that speech is performed at an average speech speed. For this reason, it is difficult to detect speech uttered at a rapid pace and speech uttered slowly.

  Even so, even if correction is performed to widen the range of the threshold value to be identified as speech, false detection in which sounds other than speech are detected as speech increases. FIG. 2 is a diagram showing an example of correction of the threshold value. In the upper part of the table of FIG. 2, the threshold for average speech speed, that is, the threshold without correction, the voice that can be detected with the threshold without correction, the voice that causes an omission of detection, and a sound other than voice that is erroneously detected It is shown. Further, in the middle part of the table in FIG. 2, there are shown thresholds corrected for slow speech speed, voices that can be detected with the thresholds after correction, voices that cause omission of detection, and sounds other than voices that are erroneously detected. It is done. Furthermore, the lower part of the table in FIG. 2 shows a threshold corrected for steep speech speed, a voice that can be detected with the threshold after correction, a voice that causes a detection failure, and a voice other than a voice that is erroneously detected. It is done.

  For example, when the fluctuation range of the fundamental frequency to be identified as voice is set in accordance with the average speech speed, the threshold including the lower limit value and the upper limit value is as shown in the upper part of the table of FIG. Among these, 0.37 Hz / ms is set to the positive gradient, that is, the lower limit value to be compared with the monotonically increasing average value, and 1.09 Hz / ms is set to the upper limit value. In addition, -1.19 Hz / ms is set as the lower limit compared with the negative slope, that is, the average value of monotonically decreasing, and -0.29 Hz / ms is set as the upper limit. When these threshold values are used, it is possible to detect speech of average speech speed, but as described above, it is difficult to detect speech uttered at a fast pace and speech uttered slowly. Thus, while the detection capability is limited, there is little risk of false detection in which sounds other than voice are detected as voice.

  In addition, when the range of fluctuation of the fundamental frequency to be identified as voice is set in accordance with the slow speech speed, the threshold value including the lower limit value and the upper limit value is as shown in the middle part of the table of FIG. Among them, the positive slope, that is, the lower limit compared to the average value of monotonically increasing, is corrected to 0.15 Hz / ms, while the upper limit is not corrected, and remains 1.09 Hz / ms. Ru. Also, the negative slope, ie the lower limit compared to the mean value of monotonically decreasing, is not corrected and remains -1.19 Hz / ms while the upper limit is corrected to -0.15 Hz / ms Be done. When these threshold values are used, it is possible to detect speech with moderate speech speed and speech with average speech speed, but it is still difficult to detect speech uttered at high speed. In addition, it becomes difficult to distinguish between voice and low frequency sound such as car engine sound, resulting in an increase in false detection of non-voice low frequency sound.

  In addition, when the range of fluctuation of the fundamental frequency to be identified as voice is set in accordance with the steep speech speed, the threshold including the lower limit value and the upper limit value is as shown in the lower part of the table of FIG. Among these, the positive slope, that is, the lower limit value to be compared with the average value of monotonically increasing, is kept at 0.37 Hz / ms while the upper limit value is corrected to 1.3 Hz / ms. Also, the negative slope, ie, the lower limit compared to the monotonically decreasing average, is corrected to -1.3 Hz / ms while the upper limit remains at -0.29 Hz / ms. When these threshold values are used, it is possible to detect speech of average speech speed or speech of steep speech speed, but it is still difficult to detect speech uttered at moderate speech speed. In addition, it becomes difficult to distinguish between voice and music other than voice, music or sound, etc., resulting in an increase in false detection.

  It is also conceivable to perform frequency analysis on the sound signal using two time windows having different time lengths and to detect speech from the difference between the two analysis results, in addition to correcting the threshold value in this manner. That is, frequency analysis is performed on a partial waveform of a short interval in the time waveform of the sound signal, and frequency analysis is performed on a partial waveform of a long interval including the short interval. The analysis results of the short section and the long section differ in the case where the sound signal includes speech. For example, when the frequency analysis is performed in the long section, the fluctuation of the spectrum is large and the frequency spectrum appears smoothly, while when the frequency analysis is performed in the short section, the fluctuation of the spectrum is small, the peaks and valleys appear in the frequency spectrum. Thus, the presence or absence of speech is determined depending on whether or not the difference in frequency spectrum between the short interval and the long interval is large.

  However, since the utterance is not always performed at a constant speech speed, even if two time windows are used for frequency analysis, the fluctuation of the fundamental frequency is not limited to one of the short or long time windows. It does not always appear to fit.

  3A to 3C are diagrams showing an example of a spectrogram. FIG. 3A shows a spectrogram of a sound signal including a voice uttered by a woman at an average speech speed. Further, FIG. 3B shows a spectrogram of a sound signal including a voice uttered by a woman at a slow speech speed. Further, FIG. 3C shows a spectrogram of a sound signal including the engine sound of a vehicle.

  As shown in FIG. 3A, a plurality of syllables are included in the spectrogram when speech is performed at an average speech speed. As a result, the slope of the fundamental frequency appears to change in syllable units in the spectrogram. That is, in the spectrogram shown in FIG. 3A, three periods of a period in which the fundamental frequency does not fluctuate, a period in which the fundamental frequency fluctuates monotonously and a period in which the fundamental frequency monotonically decreases appear in time series. As described above, when fluctuation of the fundamental frequency appears in a time window of the sound signal in which frequency analysis is performed, it is possible to endure detection of voice.

  On the other hand, in the spectrograms shown in FIGS. 3B and 3C, the fundamental frequency does not fluctuate over the entire interval. For example, in the case of the spectrogram shown in FIG. 3B, the number of syllables included in the section of the sound signal in which the frequency analysis is performed is small and the monotonicity of the fundamental frequency is due to the speaker uttering slowly. Fluctuations such as increase and monotonous decrease do not appear in the spectrogram.

  Therefore, the spectrogram shown in FIG. 3B is difficult to distinguish from the spectrogram shown in FIG. 3C. When the section of the sound signal for generating the spectrogram shown in FIG. 3B is expanded, the number of syllables included in the section increases. As a result, it becomes easy to detect speech when speech is performed at a slow speech speed, but steep speech The response of voice detection is delayed when speech is made quickly.

  As described above, even if the threshold value for slow speech speed, average speech speed, or steep speech speed is fixedly used as the threshold value to be compared with the feature value of fluctuation of the fundamental frequency, the time length is Even if frequency analysis is performed using a plurality of different time windows, the speech detection accuracy is naturally limited.

  Therefore, the watching terminal 10 according to the present embodiment obtains the speech speed from the speech candidate section of the input sound signal as part of the above speech detection process, and uses the feature quantity indicating the fluctuation of the fundamental frequency according to the speech speed. It correct | amends and determines the presence or absence of the audio | voice of an audio | voice candidate area. As a result, even if the appearance of the fluctuation of the fundamental frequency changes due to the change of the speech speed, normalization of the speech speed can be performed on the feature quantity of the fluctuation of the fundamental frequency and speech detection can be performed. Detection can be suppressed. Therefore, according to the watching terminal 10 according to the present embodiment, it is possible to suppress the decrease in the voice detection accuracy.

[Configuration of watching terminal 10]
FIG. 4 is a block diagram illustrating an example of a functional configuration of the watching terminal 10 according to the first embodiment. As shown in FIG. 4, the watching terminal 10 includes a sound signal acquisition unit 11, a speech candidate extraction unit 12, a feature quantity calculation unit 13, a speech speed calculation unit 14, a correction unit 15, and a speech detection unit 16. Have.

  Functional units such as the sound signal acquisition unit 11, the speech candidate extraction unit 12, the feature amount calculation unit 13, the speech speed calculation unit 14, the correction unit 15, and the speech detection unit 16 shown in FIG. (Virtual processing) by a hardware processor such as (Central Processing Unit). That is, the above functional unit is virtually realized by the processor expanding a speech detection program for executing the speech detection processing on a memory such as a RAM (Random Access Memory) as a process. Here, the CPU and the MPU are illustrated as an example of the processor, but the above functional unit may be realized by any processor regardless of general purpose type or special purpose type. In addition to the above, the above-described functional units may be realized by hard wired logic such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA).

  Further, the functional units shown in FIG. 4 are merely examples, and the functional configuration of the watching terminal 10 does not prevent the functional configuration other than the example shown in FIG. 4 from being included. That is, the watching terminal 10 may have other functional units other than the above-described functional units. For example, the watching terminal 10 may have a communication interface or the like connected to the network NW, in addition to a microphone for converting sound waves into electric signals, although not shown in FIG. In addition to the above, since the remote care service described above can not only detect voice but also combine the movement of a person, temperature and humidity, etc. to confirm the safety, it may also have a human sensor, a temperature and humidity sensor, etc. it can. In addition, the above-mentioned remote care service can further include a call function for making a call with the call center.

  In addition, although a solid line representing the relationship between the input and output of signals and data to the functional unit shown in FIG. 4 is illustrated, the relationship between the input and output of data indicates that data is transmitted from at least one to the other. It is not always necessary to exchange data in both directions.

[Sound signal acquisition unit 11]
The sound signal acquisition unit 11 is a processing unit that acquires a sound signal.

  In one embodiment, the sound signal acquisition unit 11 acquires, as an input signal, a sound signal converted from a sound wave by a microphone (not shown). Here, the source from which the sound signal acquisition unit 11 acquires the sound signal may be arbitrary, and is not limited to the microphone. For example, the sound signal acquisition unit 11 can also acquire a sound signal by reading from an auxiliary storage device such as a hard disk or an optical disk that stores sound data, or a removable medium such as a memory card or a USB (Universal Serial Bus) memory. In addition to this, the sound signal acquisition unit 11 can also acquire a sound signal by receiving from an external device via the network NW.

[Speech candidate extraction unit 12]
The speech candidate extraction unit 12 is a processing unit that extracts speech segment candidates from the sound signal. The term "speech segment candidate" as used herein refers to a candidate of a segment that seems to be speech in the time waveform of the sound signal. As shown in FIG. 4, the voice candidate extraction unit 12 has a sounded section detection unit 12a, a basic sound calculation unit 12b, and a determination unit 12c.

  Among these, the sounded section detection unit 12a is a processing section that detects a sounded section from the time waveform of the sound signal.

As one embodiment, the voiced section detection unit 12 a calculates a power envelope for each frame every time a new frame of the sound signal is acquired by the sound signal acquisition unit 11. For example, assuming that the frame length is T, the voiced section detection unit 12a inputs the amplitude x _i of each sampling point included in the frame length T according to the following equation (1) to obtain the power envelope of the frame. calculate. Then, the voiced section detection unit 12a determines whether the power envelope of the frame is equal to or greater than a predetermined threshold. Then, when the power envelope of the frame is equal to or greater than the threshold, the sounded section detection unit 12a detects the frame as a "sent section". Thus, when the sounding section detected from the new frame and the sounding section detected from the previous frame are connected, by connecting the sounding sections connected to each other, two or more A sound interval can be integrated into one sound interval.

  The fundamental sound calculation unit 12 b is a processing unit that calculates a fundamental frequency.

  As one embodiment, the basic sound calculation unit 12 b calculates the fundamental frequency by autocorrelation of the time waveform of the sound signal, as an example only. For example, the fundamental sound calculation unit 12b shifts the replica waveform in which the waveform of the sounding section is duplicated with respect to the original waveform of the sounding section detected by the sounding section detection unit 12a according to the following equation (2). Meanwhile, the correlation coefficient R [x] between the original waveform of the sound interval and the duplicate waveform of the sound interval is calculated for each shift width τ. Thereafter, the basic sound calculation unit 12b identifies the shift width τ when the largest correlation coefficient is calculated among the correlation coefficients calculated for each shift width τ as the “basic cycle”. Then, the basic sound calculation unit 12b calculates the fundamental frequency by taking the reciprocal of the shift width τ when the maximum correlation coefficient is calculated.

  In addition, although the case where the autocorrelation of a time waveform is used was illustrated as an example of the calculation method of fundamental frequency to the last, the calculation method of fundamental frequency is not limited to this. For example, as described in the background section above, the fundamental frequency can also be determined from the frequency spectrum. Also, the fundamental frequency may be calculated using another method including a waveform envelope method, a zero crossing method, a cepstrum method, and the like.

  The determination unit 12c is a processing unit that determines whether or not the sound section corresponds to a speech section candidate.

  In one embodiment, the determination unit 12c determines, for each correlation coefficient calculated for each shift width by the basic sound calculation unit 12b, whether the correlation coefficient is equal to or more than a predetermined threshold. As a result, the ratio of the number of times the correlation coefficient equal to or greater than the threshold value is calculated is obtained among the total number of times the correlation coefficient is calculated by the basic calculation unit 12b. In addition, when the above ratio is equal to or higher than the threshold, the determination unit 12c determines whether the fundamental frequency calculated by the basic sound calculation unit 12b is within the allowable range for voice, for example, in the range of 100 Hz to 400 Hz. Do. If these two conditions are satisfied, that is, if the above ratio is equal to or higher than the threshold and the fundamental frequency is within the allowable range for voice, the determining unit 12c may Detect as a candidate.

  Here, although a case where a sounded section is detected as a speech section candidate conditionally is illustrated, a sounded section can be detected as a speech section candidate unconditionally. In this case, the determination of the above two conditions can be omitted.

[Feature amount calculation unit 13]
The feature amount calculation unit 13 is a processing unit that calculates a feature amount of fluctuation of the fundamental frequency.

  As one embodiment, the feature quantity calculation unit 13 is included in the time window while shifting the time window having a predetermined analysis length on the time waveform of the fundamental frequency in the speech section candidate extracted by the speech candidate extraction unit 12 The inclination of the approximate straight line is obtained by performing function approximation or the like by a linear function on the time series data of the fundamental frequency to be obtained.

  FIG. 5 is a diagram showing an example of the time waveform of the fundamental frequency in the speech segment candidate. The vertical axis of the graph in FIG. 5 indicates frequency, and the horizontal axis indicates time. FIG. 5 shows a time window in the case where the analysis length for performing the function approximation by the linear function is 60 msec. As shown in FIG. 5, the feature quantity calculation unit 13 shifts a time window of an analysis length of 60 msec by a predetermined shift width on the time waveform of the fundamental frequency in the speech segment candidate. That is, the time window (i), the time window (ii) and the time window (iii) are shifted in this order. As described above, every time the time window is shifted, the feature quantity calculation unit 13 performs function approximation using a linear function on time series data of the fundamental frequency included in the time window. For example, in the case of the time window (i), the slope and the intercept of the approximate straight line (i) are estimated by performing linear regression analysis that minimizes the sum of squares of residuals for the first-order model function. As described above, when the time window is shifted, the statistical value of the slope, for example, the average value, is calculated by dividing each of the slopes of the approximate straight line obtained with the positive value and the negative value.

  Here, although the case of calculating the average value of the positive slope and the average value of the negative slope has been illustrated as an example of the feature quantity, the present invention is not limited to this, and it is possible to Other feature quantities such as proportions can also be calculated. In this case, for example, by calculating the ratio between the first section in which the slope of any positive or negative value is detected in the voice section candidate and the second section in which the slope is zero in the voice section candidate, The ratio of time in which the fundamental sound fluctuates can be calculated as the feature value.

[Speech rate calculator 14]
The speech speed calculation unit 14 is a processing unit that calculates the speech speed of the speech segment candidate.

  As one embodiment, the speech speed calculation unit 14 estimates the number of syllables included in the speech segment candidates extracted by the speech candidate extraction unit 12. For the estimation of the number of syllables, the correlation coefficient calculated for each shift width τ by the basic sound calculation unit 12b is used as the voice section candidate detected by the determination unit 12c, as an example. That is, since there is periodicity in the part where vowels are uttered, the correlation coefficient tends to be high, while there is no periodicity in the part where consonants are uttered, so the correlation coefficient tends to be low. is there. As described above, in order to estimate the number of syllables using the characteristics of continuous speech sounds, the total number of continuous intervals with the value of the correlation coefficient at a predetermined threshold, for example, “0.7” or more, in the speech interval candidate The number of syllables in the speech segment candidate is estimated by calculating Here, as an example, the case of calculating the total number of consecutive sections with the value of the correlation coefficient being a predetermined threshold or more in the speech section candidate has been illustrated, but the waveform of the correlation coefficient is observed in the speech section candidate The number of syllables in the speech segment candidate can also be estimated by calculating the number of valleys.

  FIG. 6 is a diagram showing a graph of correlation coefficients in speech segment candidates. FIG. 6 shows, as an example, a graph of the correlation coefficient in the speech segment candidate in which the utterance "man" is performed. The vertical axis of the graph shown in FIG. 6 indicates the value of the correlation coefficient, and the horizontal axis indicates the shift width τ. In the speech section candidate 60 shown in FIG. 6, in the section where the correlation coefficient is equal to or more than the threshold value "0.7", the correlation coefficient becomes high at the place where the periodicity of vowels is observed, and the relationship is the place where the syllable breaks off. According to the characteristics of the speech sound whose number is low, there are four sections (d), (d), (d) and (t). Therefore, the number of syllables of the speech segment candidate 60 is estimated to be “4”. Here, although the case where the number of syllables is calculated | required by counting the number of area where a correlation coefficient becomes more than threshold value "0.7" was illustrated as an example to the last, the calculation method of the number of syllables is not limited to this. For example, valleys appearing in the voice section candidate 60, that is, valleys before and after the section (d), valleys behind the section (e), valleys behind the section (f) and valleys after the section (t) The number of syllables can also be estimated to be "4" by obtaining the number "4" of peak portions sandwiched between five valleys from the number "5" of portions.

  As described above, after the number of syllables in the speech segment candidate is estimated, the speech speed calculation unit 14 converts the number of syllables in the speech segment candidate into unit time, for example, the number of syllables per second. The speed (syllables / second) can be calculated.

[Correction unit 15]
The correction unit 15 is a processing unit that corrects the feature amount calculated by the feature amount calculation unit 13 according to the speech speed calculated by the speech speed calculation unit 14.

  In one embodiment, the correction unit 15 calculates a correction coefficient α by which the feature amount calculated by the feature amount calculation unit 13 is multiplied from the speech speed calculated by the speech speed calculation unit 14. The correction coefficient α is calculated according to a derived function of the correction coefficient α shown in FIG. 7 as an example. FIG. 7 is a diagram showing an example of a derived function of the correction coefficient α. FIG. 7 shows a graph showing the relationship between the speech speed and the correction coefficient α. The vertical axis of the graph shown in FIG. 7 indicates the correction coefficient α, and the horizontal axis indicates the speech speed [syllable / second]. As shown in FIG. 7, for the derivation of the correction coefficient α, a larger value is derived as the speech speed is smaller, while a function is used which derives a smaller value as the speech speed is larger. In other words, as the speech in the speech segment candidate is slower, the larger correction coefficient α is derived, while the earlier the speech in the speech segment candidate is smaller, the smaller correction coefficient α is derived. Thus, the magnitude of the correction coefficient α can be changed according to the magnitude of the speech speed, but the derivation function of the correction coefficient may not necessarily be a monotonically decreasing graph, as shown in FIG. That is, when the speech speed falls within the range of the average, the correction coefficient α is derived as “1” regardless of the value of the speech speed within the range of the average. As described above, when the correction coefficient α is “1”, the feature amount calculated by the feature amount calculating unit 13 is not substantially corrected. Also, when the speech speed is 2.9 [syllables / second] or less, the speech speed is the same as 2.9 [syllables / second] even if the speech speed is 2.9 [syllables / second] or less The correction coefficient α of is derived. If the speech speed is 11 [syllables / second] or more, the same correction coefficient α is derived as in the case where the speech speed is 11 [syllables / second] even if the speech speed is 11 [syllables / second] or more. Be done. The correction unit 15 can calculate the correction coefficient α from the speech speed in the speech segment candidate through such a derivation function.

  After the correction coefficient α is calculated as described above, the correction unit 15 corrects the feature amount by multiplying the feature amount calculated by the feature amount calculation unit 13 by the correction coefficient α. That is, while the correction unit 15 multiplies the correction coefficient α by the positive slope calculated by the feature amount calculation unit 13, that is, the average value [frequency / ms] of monotonous increase, the correction unit 15 multiplies the negative coefficient, that is, the average of monotonous decrease. The value [frequency / ms] is also multiplied by the correction factor α. Thereby, the average value of monotonous increase after correction and the average value of monotonous decrease after correction are calculated.

[Voice detection unit 16]
The voice detection unit 16 is a processing unit that detects a voice based on the feature amount of the voice section candidate corrected by the correction unit 15.

  In one embodiment, the voice detection unit 16 compares the average value of monotonous increase and the average value of monotonous decrease corrected by the correction unit 15 with a threshold. For example, the voice detection unit 16 determines whether or not the average value of monotonically increasing after correction is included in the lower limit value of fluctuation, for example, 0.37 Hz / ms and the upper limit value, for example, 1.09 Hz / ms. . At this time, when the average value of monotonically increasing after correction is included in the range between the lower limit value and the upper limit value of fluctuation, the voice detection unit 16 determines that the average value of monotonically decreasing after correction is the lower limit value of fluctuation, for example, −1. It is further determined whether it falls within the range of 19 Hz / ms and an upper limit value, for example, −0.29 Hz / ms. Then, when the average value of monotonically decreasing after the correction is included in the range of the lower limit value and the upper limit value of the fluctuation, the voice detection unit 16 detects the voice section candidate as “with voice”. On the other hand, when the average value of monotonically increasing after correction is not included in the lower limit value and the upper limit of fluctuation, or the average value of monotonous decrease after correction is included in the lower limit value and upper limit of fluctuation. If not, the voice segment candidate is detected as "no voice". The voice detection result is output from the watching terminal 10 to the server device 30 each time the voice detection result is obtained or the voice detection result is accumulated for a predetermined period, for example, one hour.

[Flow of processing]
FIG. 8 is a flowchart illustrating the procedure of the voice detection process according to the first embodiment. This process is started, for example, when a new frame of the sound signal is acquired by the sound signal acquisition unit 11.

  As shown in FIG. 8, when a sound signal frame is newly acquired by the sound signal acquisition unit 11 (step S101), the sound section detection unit 12a determines that the power envelope of the frame acquired in step S101 is predetermined. If it is equal to or greater than the threshold value, the frame is detected as a "sound interval" (step S102). As described above, when the sounding section detected from the new frame is connected to the sounding section detected from the previous frame, two or more sounding sections can be connected by connecting the sounding sections to be connected. A section can be integrated into one sound section.

  Subsequently, as one example, the basic sound calculation unit 12b calculates a fundamental frequency by autocorrelation of the time waveform in the sounding section detected by the sounding section detection unit 12a (step S103). Then, the determination unit 12c determines whether or not the sound section corresponds to a speech section candidate (step S104). For example, in step S104, the ratio of the number of times the correlation coefficient equal to or greater than the threshold is calculated among the total number of times the correlation coefficient is calculated in the sounded section and whether the fundamental frequency in the sounded section satisfies a predetermined condition Thus, it is determined whether the sound section corresponds to the voice section candidate.

  At this time, when the sound section corresponds to the voice section candidate (Yes at step S104), the feature quantity calculation unit 13 shifts the time window having a predetermined analysis length on the time waveform of the fundamental frequency in the voice section candidate. The gradient of the approximate straight line is obtained by performing function approximation or the like by a linear function on the time series data of the fundamental frequency included in the time window (step S105).

  Then, the feature quantity calculation unit 13 divides the inclination of the approximate straight line obtained each time the time window is shifted into inclinations having positive values and inclinations having negative values, and the inclination statistical values, For example, the average value is calculated as a feature of fluctuation of the fundamental frequency (step S106).

  In addition, the speech speed calculation unit 14 estimates the number of syllables included in the speech segment candidate, and converts the number of syllables in the speech segment candidate into unit time, for example, the number of syllables per second, to calculate the speech velocity in the speech segment candidate (Syllables / second) is calculated (step S107).

  Thereafter, the correction unit 15 corrects the feature amount calculated in step S106 by multiplying the feature amount calculated in step S106 by the correction coefficient α determined by the speech speed calculated in step S107 (step S108). .

  Then, the voice detection unit 16 determines whether or not the feature value corrected in step S108 is within the fluctuation range of the fundamental frequency that is likely to be voice (step S109). For example, the voice detection unit 16 determines whether the average value of monotonically increasing is included in the range between the lower limit value and the upper limit value of fluctuation, and the average value of monotonously decreasing is included in the range of the lower limit value and upper limit value of fluctuation. Determine if

  After that, the voice detection unit 16 outputs the determination result in step S109 as to whether or not the corrected feature amount is within the fluctuation range of the fundamental frequency that seems to be voice to the server device 30 as the voice detection result in the voice section candidate. (Step S110) The process returns to the process of step S101.

[One side of effect]
As described above, the watching terminal 10 according to the present embodiment obtains the speech speed from the speech candidate section of the input sound signal, corrects the feature amount indicating the fluctuation of the fundamental frequency according to the speech speed, and calculates the speech candidate. Determine the presence or absence of voice in the section. As a result, even if the appearance of the fluctuation of the fundamental frequency changes due to the change of the speech speed, normalization of the speech speed can be performed on the feature quantity of the fluctuation of the fundamental frequency and speech detection can be performed. Detection can be suppressed. Therefore, according to the watching terminal 10 according to the present embodiment, it is possible to suppress the decrease in the voice detection accuracy.

  Although the embodiments of the disclosed apparatus have been described above, the present invention may be implemented in various different forms other than the above-described embodiments. Therefore, another embodiment included in the present invention will be described below.

[Threshold correction]
Although the case of correcting the feature quantity of the fluctuation of the fundamental frequency is illustrated in the above-described first embodiment, the present invention is not limited to this. For example, the correction unit 15 can correct the threshold value to be compared with the feature amount. At this time, the correction unit 15 can correct only the threshold value to be compared with the feature value, or can correct both the feature value and the threshold value.

  Specifically, the correction unit 15 calculates a correction coefficient β to be multiplied by a threshold value to be compared with the feature amount calculated by the feature amount calculation unit 13 from the speech speed calculated by the speech speed calculation unit 14. The correction coefficient β is calculated, for example, in accordance with a derived function of the correction coefficient β shown in FIG. FIG. 9 is a diagram showing an example of a derived function of the correction coefficient β. A graph showing the relationship between the speech speed and the correction coefficient β is shown in FIG. The vertical axis of the graph shown in FIG. 9 indicates the correction coefficient β, and the horizontal axis indicates speech speed [syllables / second]. As shown in FIG. 9, for the derivation of the correction coefficient β, a smaller value is derived as the speech speed is smaller, while a function is used in which a larger value is derived as the speech speed is larger. In other words, the slower the speech in the speech segment candidate is, the smaller the correction coefficient β is derived, while the earlier the speech in the speech segment candidate is, the larger the correction coefficient β is derived. As described above, the magnitude of the correction coefficient β can be changed according to the magnitude of the speech speed, but the derivation function of the correction coefficient β may not necessarily be a monotonically increasing graph, as shown in FIG. That is, when the speech speed falls within the range of the average, the correction coefficient β is derived as “1” regardless of the value of the speech speed within the range of the average. Thus, when the correction coefficient β is “1”, the threshold value to be compared with the feature amount is not substantially corrected. Also, when the speech speed is 2.9 [syllables / second] or less, the speech speed is the same as 2.9 [syllables / second] even if the speech speed is 2.9 [syllables / second] or less The value of “0.5” is derived as the correction coefficient β. Also, when the speech speed is 11 [syllables / second] or more, the same value “1.5” as in the case where the speech speed is 11 [syllables / second] even if the speech speed is 11 [syllables / second] or more “Is the correction coefficient β is derived. The correction unit 15 can calculate the correction coefficient β from the speech speed in the speech segment candidate through such a derivation function.

  After the correction coefficient β is calculated, the correction unit 15 corrects the threshold value by multiplying the threshold value to be compared with the feature amount calculated by the feature amount calculation unit 13 by the correction coefficient β. That is, the correction unit 15 multiplies the correction coefficient β by the threshold value to be compared with the average value of monotonically increasing, that is, the lower limit value and the upper limit value of the sound-like fluctuation. With such correction, the lower limit value of the threshold value to be compared with the average value of monotonically increasing becomes “0.37 × β (Hz / ms)” and the upper limit value is “1.09 × β (Hz / ms)”. Become. Furthermore, the correction unit 15 multiplies the correction coefficient β by the threshold value to be compared with the average value of monotonically decreasing, that is, the lower limit value and the upper limit value of the sound-like fluctuation. With such correction, the lower limit value of the threshold value to be compared with the average value of monotonically decreasing becomes “−1.19 × β (Hz / ms)” and the upper limit value is “−0.29 × β (Hz / ms) It becomes ".

  After the threshold value is corrected as described above, the voice detection unit 16 compares the average value of monotonous increase and the average value of monotonous decrease corrected by the correction unit 15 with the threshold value corrected by the correction unit 15. For example, the voice detection unit 16 sets the threshold value range of the fluctuation that seems to be voice after correction, ie, the lower limit value “0.37 × β (Hz / ms)” and the upper limit value “1.09. It is determined whether or not it falls within the range of x β (Hz / ms). At this time, when the average value of monotonically increasing after correction is included in the range of the threshold value of fluctuation that seems to be the voice after correction, the voice detection unit 16 detects the average value of monotonous decrease after correction is the fluctuation that seems to be voice after correction. It is further determined whether it falls within the range of the threshold value, that is, within the range of the lower limit value “−1.19 × β (Hz / ms)” and the upper limit value “−0.29 × β (Hz / ms)”. Then, when the average value of monotonically decreasing after correction is included in the threshold range of fluctuation that seems to be the voice after correction, the voice detection unit 16 detects the voice segment candidate as “voice present”. On the other hand, when the average value of monotonically increasing after correction is not included in the threshold range of fluctuation like speech after correction, or the average value of monotonous decrease after correction is within the threshold range of fluctuation like speech after correction If it is not included, the speech segment candidate is detected as "no speech". The voice detection result is output from the watching terminal 10 to the server device 30 each time the voice detection result is obtained or the voice detection result is accumulated for a predetermined period, for example, one hour. Here, as an example, although the case where both the feature amount and the threshold value are corrected is illustrated, it is also possible to correct only the threshold value.

[Smoothing of speech speed]
In the first embodiment described above, the speech speed in the speech segment candidate is calculated. However, the speech speed can be calculated in another span. For example, the speech speed calculation unit 14 can also calculate the speech speed in units of sub-frames, for example, one second. In this case, the speech speed calculation unit 14 can smooth the speech speed between subframes according to the following equation (3). In addition, when each process of the speech speed calculation unit 14 and the processing unit subsequent thereto is performed in units of subframes, subframes on which each process is performed are narrowed down to subframes overlapping with part or all of the speech segment candidates. .

  Av_Speech_speed_tx = 0.2 × Speech_speed_tx + 0.8 × Av_Speech_speed_tx-1 (3)

  “Av_Speech_speed_tx” in the above equation (3) indicates the speech speed [syllabus / second] after smoothing in the current subframe. Also, "Speech_speed_tx" in the above equation (3) indicates the speech speed [syllables / second] of the current subframe. This "Speech_speed_tx" can similarly calculate the speech speed in a subframe by replacing the speech segment candidate portion described in the first embodiment above with the subframe. Further, “Av_Speech_speed_tx-1” in the above equation (3) indicates the smoothed speech speed [syllables / second] calculated in the previous subframe.

[Setting of feature amount calculation section 1]
Here, the watching terminal 10 can set the total length of the section effective for the calculation of the feature amount according to the speech speed. Hereinafter, the section in which the calculation of the feature amount is performed may be referred to as a “feature amount calculation section”. That is, in the above-described first embodiment, the case where the inclination of the primary approximate straight line is repeatedly calculated while sliding the time window having the analysis length of about 60 msec in the speech section candidates has been exemplified. Instead of setting the total length of the section to be performed as the speech section candidate, the feature amount calculation section is set by expanding and reducing the sub-frame according to the speech speed after smoothing. In addition, although the case where a sub-frame is expanded and contracted using the speech speed after smoothing is illustrated as an example to the last only, as the sub-frame is expanded and contracted using the speech speed calculated in the above-mentioned first embodiment. I don't care.

  For example, the watching terminal 10 calculates a scaling factor γ to be multiplied by the time length of the subframe from the smoothed speech speed calculated by the speech speed calculation unit 14. The scaling factor γ is calculated according to the derivation function of the scaling factor γ shown in FIG. 10 as an example. FIG. 10 is a diagram illustrating an example of a derived function of the expansion coefficient γ. FIG. 10 shows a graph showing the relationship between the speech speed and the expansion coefficient γ. The vertical axis of the graph shown in FIG. 10 indicates the expansion coefficient γ, and the horizontal axis indicates the speech speed [syllable / second]. As shown in FIG. 10, for the derivation of the expansion / contraction coefficient γ, a larger value is derived as the speech speed is smaller, while a function is used which derives a smaller value as the speech speed is larger. In other words, as the utterance in the sub-frame is slower, a larger expansion / contraction factor γ is derived, while as the utterance in the sub-frame is faster, the smaller expansion / contraction factor γ is derived. Thus, the magnitude of the expansion coefficient γ can be changed according to the size of the speech speed, but the derivation function of the expansion coefficient may not necessarily be a monotonically decreasing graph, as shown in FIG. That is, when the speech speed falls within the range of the average, the scaling factor γ is derived as “1” regardless of the value of the speech speed within the range of the average. As described above, when the expansion coefficient γ is “1”, the total length of the section in which the feature amount is calculated is substantially the same as the sub-frame time length. Also, when the speech speed is 2.9 [syllables / second] or less, the speech speed is the same as 2.9 [syllables / second] even if the speech speed is 2.9 [syllables / second] or less The scaling factor γ of is derived. When the speech speed is 11 [syllables / second] or more, the same expansion coefficient γ is derived as in the case where the speech speed is 11 [syllables / second] even if the speech speed is 11 [syllables / second] or more Be done. The watching terminal 10 can calculate the scaling factor γ from the speech speed through such a derivation function.

  After the expansion coefficient γ is calculated, the watching terminal 10 sets a feature amount calculation section by multiplying the time length of the sub-frame, for example, 1 second by the expansion coefficient γ. Taking the feature amount calculation section set in this way as the total length, the feature amount calculation unit 13 calculates the statistical value of the inclination, for example, the average value, by dividing the inclination having a positive value and the inclination having a negative value. By this, the feature quantity of the fluctuation of the fundamental frequency is calculated. After that, the correction unit 15 corrects the feature amount using the correction coefficient α determined from the speech speed after smoothing, and the voice detection unit 16 detects that the feature amount after correction falls within the fluctuation range of the fundamental frequency like speech. Whether or not there is voice in the subframe is determined.

[Setting of feature amount calculation section 2]
Here, although the case where speech speed is used is illustrated as an example of the setting method of a feature-value calculation area, a feature-value calculation area can also be set using this other factor. For example, the watching terminal 10 can set the feature value calculation section by expanding or reducing the subframe in accordance with the speech speed after smoothing and the reliability of the subframe.

  In this case, the watching terminal 10 calculates, for each subframe, a reliability that indicates the degree to which the data of the subframe can be trusted. As an example of this reliability, the denominator is the total number of frames included in a subframe, where the numerator is the number of frames in which the determination unit 12c can calculate two basic frequencies that use the determination in the subframe. It can be calculated according to the calculation formula

  For example, the watching terminal 10 calculates a scaling factor γ and a scaling factor Δ by which the time length of the subframe is multiplied. Among these, the expansion / contraction coefficient γ is a coefficient calculated by the speech speed after smoothing, and is calculated from the speech speed according to FIG. 10 as described above. On the other hand, the expansion coefficient Δ is a coefficient calculated from the reliability of the subframe. The scaling factor Δ is calculated according to a derived function of the scaling factor Δ shown in FIG. 11 as an example. FIG. 11 is a diagram illustrating an example of a derived function of the expansion coefficient Δ. FIG. 11 shows a graph showing the relationship between the reliability and the expansion coefficient Δ. The vertical axis of the graph shown in FIG. 11 indicates the expansion coefficient Δ, and the horizontal axis indicates the degree of reliability. As shown in FIG. 11, for the derivation of the expansion coefficient Δ, a larger value is derived as the reliability is smaller, while a function is used which derives a smaller value as the reliability is larger. As described above, the magnitude of the expansion coefficient Δ can be changed according to the degree of reliability, but the derivation function of the expansion coefficient may not necessarily be a monotonically decreasing graph, as shown in FIG. That is, when the reliability is 0.65 or more, the same value “1” as in the case of the reliability of 0.65 is derived as the expansion coefficient Δ even if the reliability is 0.65 or more. The watching terminal 10 can calculate the scaling factor Δ from the speech speed through such a derivation function. In addition, although the expansion / contraction coefficient Δ in the case where the reliability is less than 0.3 is not set in FIG. 11, since the possibility that voice is detected is low when the reliability is less than 0.3, It is possible to stop the calculation of the feature amount.

  After the expansion coefficient γ and the expansion coefficient Δ are calculated, the watching terminal 10 sets a feature amount calculation section by multiplying the time length of the subframe, for example, 1 second by the expansion coefficient γ and the expansion coefficient Δ. Can. Taking the feature amount calculation section set in this way as the total length, the feature amount calculation unit 13 calculates the statistical value of the inclination, for example, the average value, by dividing the inclination having a positive value and the inclination having a negative value. By this, the feature quantity of the fluctuation of the fundamental frequency is calculated. At this time, if the degree of reliability is less than 0.3, the possibility of the voice being detected is low, so that the calculation of the feature amount can be stopped. After that, the correction unit 15 corrects the feature amount using the correction coefficient α determined from the speech speed after smoothing, and the voice detection unit 16 detects that the feature amount after correction falls within the fluctuation range of the fundamental frequency like speech. Whether or not there is voice in the subframe is determined.

[Feature value]
In the first embodiment described above, the feature amount related to the fluctuation of the fundamental frequency is calculated as an example of the feature amount, but the feature amount calculated by the watching terminal 10 is not limited to the one related to the fluctuation of the fundamental frequency. For example, the watching terminal 10 may calculate a feature amount related to the fluctuation of the frequency called harmonic of the fundamental frequency. Also in this case, apart from arranging the derivation of the correction coefficient α to the derivation corresponding to the harmonic or arranging the threshold to be compared with the corrected feature amount to the threshold corresponding to the harmonic, calculation of speech speed, feature amount There is no change in the contents of the processing such as the correction and the voice detection, and the presence or absence of voice in the voice segment candidate can be detected as in the first embodiment.

Distributed and integrated
Further, each component of each illustrated device may not necessarily be physically configured as illustrated. That is, the specific form of the distribution and integration of each device is not limited to the illustrated one, and all or a part thereof may be functionally or physically dispersed in any unit depending on various loads, usage conditions, etc. It can be integrated and configured. For example, the sound signal acquisition unit 11, the speech candidate extraction unit 12, the feature amount calculation unit 13, the speech speed calculation unit 14, the correction unit 15 or the speech detection unit 16 are connected via a network as an external device of the terminal 10 It is also good. In addition, another device has the sound signal acquisition unit 11, the speech candidate extraction unit 12, the feature quantity calculation unit 13, the speech speed calculation unit 14, the correction unit 15, or the speech detection unit 16, and they are connected by a network and cooperate Thus, the function of the watching terminal 10 described above may be realized.

  In addition, although the example in which the voice detection process shown in FIG. 8 is executed by the watching terminal 10 has been described in the first embodiment, the execution subject of the voice detection process shown in FIG. 8 is not limited to the watching terminal 10. For example, the server device 30 may receive an upload of the sound signal collected from the microphone of the watching terminal 10, and may execute voice detection processing on the uploaded sound signal.

[Voice detection program]
The various processes described in the above embodiments can be realized by executing a prepared program on a computer such as a personal computer or a workstation. Therefore, an example of a computer that executes a voice detection program having the same function as that of the above embodiment will be described below with reference to FIG.

  FIG. 12 is a diagram illustrating an example of a hardware configuration of a computer that executes a voice detection program according to the first embodiment and the second embodiment. As illustrated in FIG. 12, the computer 100 includes an operation unit 110a, a speaker 110b, a camera 110c, a display 120, and a communication unit 130. The computer 100 further includes a CPU 150, a ROM 160, an HDD 170, and a RAM 180. The components 110 to 180 are connected via a bus 140.

  In the HDD 170, as shown in FIG. 12, the sound signal acquisition unit 11, the speech candidate extraction unit 12, the feature quantity calculation unit 13, the speech speed calculation unit 14, the correction unit 15, and the speech detection unit shown in the first embodiment A voice detection program 170a that exerts the same function as that of S.16 is stored. The speech detection program 170a is similar to the respective components of the sound signal acquisition unit 11, the speech candidate extraction unit 12, the feature amount calculation unit 13, the speech speed calculation unit 14, the correction unit 15, and the speech detection unit 16 shown in FIG. , May be integrated or separated. That is, the HDD 170 may not necessarily store all the data described in the first embodiment, and data used for processing may be stored in the HDD 170.

  Under such an environment, the CPU 150 reads the voice detection program 170 a from the HDD 170 and develops the voice detection program 170 a in the RAM 180. As a result, the speech detection program 170a functions as a speech detection process 180a as shown in FIG. The voice detection process 180a expands various data read from the HDD 170 in the area allocated to the speech detection process 180a in the storage area of the RAM 180, and executes various processes using the expanded various data. For example, the process shown in FIG. 8 or the like is included as an example of the process executed by the speech detection process 180a. In the CPU 150, all the processing units described in the first embodiment may not necessarily operate, and processing units corresponding to processing to be executed may be virtually realized.

  The voice detection program 170 a may not necessarily be stored in the HDD 170 or the ROM 160 from the beginning. For example, the voice detection program 170 a is stored in a “portable physical medium” such as a flexible disc, a so-called FD, a CD-ROM, a DVD disc, a magneto-optical disc, an IC card or the like inserted into the computer 100. Then, the computer 100 may acquire the voice detection program 170 a from these portable physical media and execute it. Further, the voice detection program 170a is stored in another computer or server device connected to the computer 100 via a public line, the Internet, LAN, WAN, etc., and the computer 100 acquires the voice detection program 170a from these. It may be carried out.

DESCRIPTION OF SYMBOLS 1 remote care system 10 watching terminal 11 sound signal acquisition part 12 audio | voice candidate extraction part 12a sounding area detection part 12b fundamental sound calculation part 12c determination part 13 feature-value calculation part 14 speech speed calculation part 15 correction part 16 audio | voice detection part 30 server apparatus 50 concerned party terminal

Claims (6)

A sound signal acquisition unit for acquiring a sound signal;
A voice candidate extraction unit which extracts a candidate of a section likely to be voice as a voice section candidate based on a voiced section of the time waveform of the sound signal;
A feature amount calculation unit that calculates a feature amount indicating the degree of speech likeness of frequency change in the voice section candidate;
A speech speed calculation unit that calculates a speech speed in the speech segment candidate;
A correction unit that corrects the feature amount according to the speech speed;
A voice detection unit that detects the presence or absence of voice in the voice segment candidate based on the corrected feature amount;
A voice detection apparatus characterized by having:   The speech speed calculation unit calculates the autocorrelation of the time waveform in the speech segment candidate, and estimates the number of syllables included in the speech segment candidate from the number of consecutive segments whose correlation coefficient is greater than or equal to a predetermined threshold, The speech detection apparatus according to claim 1, wherein the speech speed per unit time is calculated from the number of syllables. The correction unit corrects a threshold compared with the feature amount according to the speech speed.
The voice detection apparatus according to claim 1 or 2, wherein the voice detection unit detects the presence or absence of voice in the voice section candidate by comparing the feature amount after correction and the threshold value after correction. .   The speech detection apparatus according to any one of claims 1 to 3, further comprising: a section setting unit that performs expansion or contraction of a speech section candidate that calculates the feature quantity according to the speech speed.   The speech detection according to claim 4, wherein the section setting unit performs expansion or contraction of a speech section candidate for calculating the feature quantity according to the reliability of data in the speech section candidate and the speech speed. apparatus. Get a sound signal,
The candidate of the section which seems to be voice is extracted as a voice section candidate based on the sound section of the time waveform of the sound signal,
Calculating a feature amount indicating a degree of speech likeness of frequency change in the speech segment candidate;
Calculate the speech speed in the speech segment candidate;
Correcting the feature amount according to the speech speed;
Detecting the presence or absence of speech in the speech segment candidate based on the corrected feature amount;
A voice detection program that causes a computer to execute a process.