JP6435133B2 - Phoneme segmentation apparatus, speech processing system, phoneme segmentation method, and phoneme segmentation program - Google Patents

Description

  The present invention relates to a phoneme dividing device for dividing and extracting phonemes from a speech signal, a speech processing system using the phoneme dividing device, a phoneme dividing method, a phoneme dividing program, and a noise measuring device.

  In recent years, the need to protect conversation information in banks and offices has increased due to the enforcement of the “Personal Information Protection Law”, etc., but as a means of doing so, it has been opened apart from conventional sound and sound insulation that physically separates spaces. A speech privacy system (speech information secret device) that conceals (information masking) an audio signal with another noise or music in a plan office or the like has been proposed. As a speech privacy system, for example, a system that uses original speech as a masker (see, for example, Patent Document 1) is known.

  While a general SR (Sound Reinforcement) system and PA (Public Address) system using electroacoustics are used for the purpose of improving sound volume and clarity, the speech privacy system is a simplified structure of an audio signal by signal processing. By changing and processing in real time, without significantly changing statistical properties such as the spectrum and energy envelope of the audio signal, only the content of the audio is concealed / blocked and the listener cannot understand the contents of the conversation. It is intended to be.

  In the above Patent Document 1, “substantially a mountain” of a speech envelope is extracted as one phoneme and rearranged to change the structure of the speech to mask a masker (superimposed on the original speech to conceal its contents) Used as a separate voice).

  Conventionally, as a method for extracting substantially one peak of a speech envelope, a method in which one energy element (1 mora) from when the energy envelope of an input speech rises beyond a threshold and returns to the original level is generally used.

JP 2011-123141 A

  However, background noise is usually superimposed on input sound collected by a microphone or the like. The background noise in a general room or space is almost constant in a short time, but tends to fluctuate considerably in a long time. Therefore, even if the threshold is set based on the background noise level for a certain time, there is a possibility that appropriate phoneme division cannot be performed due to the background noise level fluctuation. Also, it is a difficult task to manually adjust the threshold according to the background noise level fluctuation.

  The present invention has been made in view of these problems, and an object of the present invention is to provide a technique that makes it possible to automatically detect the background noise level.

  In order to solve the above problems, a phoneme division device according to an aspect of the present invention is divided into a first branching unit that branches a sound signal in which a background noise signal is superimposed on a sound signal into two, and a first branching unit. A second branching unit for further branching the other sound signal into two, and a sound time constant unit for smoothing one sound signal branched by the second branching unit with a sound time constant of several tens to several hundreds of ms The rise of the other sound signal branched by the second branching unit is smoothed with a rise time constant that is at least 10 times larger than the sound time constant, and the rise of the other sound signal Is a comparison unit that compares the signal from the time constant part for background noise and the signal from the time constant part for background noise with the time constant part for background noise that is smoothed by the time constant for falling that is almost the same as the time constant for voice And other branches branched at the first branch according to the comparison result of the comparator And a gate portion for controlling transmission / non-transmission of the sound signal.

  Another aspect of the present invention is a speech processing system. This system collects the original sound and outputs a sound signal in which a background noise signal is superimposed on the sound signal, and receives the sound signal from the sound collector and divides the sound signal into phonemes The above-mentioned phoneme dividing device, a phoneme processing device that performs a predetermined process on a phoneme signal obtained from the phoneme dividing device, and an output device that outputs the phoneme signal processed by the phoneme processing device as a sound to space.

  Yet another embodiment of the present invention is a phoneme division method. This method includes a first branching step for branching a sound signal in which a background noise signal is superimposed on an audio signal into two, and a second branching step for branching one of the sound signals branched in the first branching step into two. A first smoothing step for smoothing one sound signal branched in the second branching step with a time constant for sound of several tens to several hundreds of ms, and a rise of the other sound signal branched in the second branching step Is smoothed with a rising time constant that is at least 10 times larger than the audio time constant, and the other sound signal is smoothed with a falling time constant that is substantially the same as the audio time constant. The first smoothing step, the comparison step comparing the signal calculated in the first smoothing step and the signal calculated in the second smoothing step, and the first branch depending on the comparison result of the comparison step Ste And a transmission controlling step for controlling passage / non-passage of the other sound signals branched by up.

  Yet another embodiment of the present invention is a phoneme division program. This program causes a computer to branch a sound signal obtained by superimposing a background noise signal on a sound signal into two, and to branch one sound signal branched in the first branch step into two. A second smoothing step, a first smoothing step for smoothing one sound signal branched in the second branching step with an audio time constant of several tens to several hundreds of ms, and the other sound branched in the second branching step The signal rise is smoothed by a rise time constant that is at least 10 times larger than the audio time constant, and the fall time constant of the other sound signal is substantially the same as the audio time constant. The second smoothing step smoothed in step 1, the comparison step comparing the signal calculated in the first smoothing step with the signal calculated in the second smoothing step, and the comparison result of the comparison step In response, a phoneme splitting program for executing a transmission control step of controlling transmission / non-transmission of the other sound signal branched by the first branching step.

  Yet another embodiment of the present invention is a noise measurement device. This device is a noise measurement device that measures the level of background noise included in ambient sound, and is a tens of times for a sound collection unit that collects ambient sound and a falling edge of a sound signal from the sound collection unit. Smoothing with a falling time constant of ˜100 ms, and for dark noise smoothing with a rising time constant that is at least 10 times greater than the falling time constant for the rise of the sound signal from the sound collection unit And a time constant part.

  Yet another embodiment of the present invention is also a noise measurement device. This device is a noise measuring device that measures the level of noise included in ambient sound, and a branching unit that branches the ambient sound into two, and one sound signal branched at the branching unit is several 10 to several 100 ms. The time constant for speech smoothing with the time constant for speech and the rise of the other sound signal branched at the branching portion are smoothed with a time constant for rise of at least 10 times greater than the time constant for speech. In addition, for the falling edge of the other sound signal, the background time constant part for smoothing with the time constant for falling substantially the same as the time constant for sound, the signal from the time constant part for sound, and for the background noise And a display unit for displaying a signal from the time constant unit.

  It should be noted that any combination of the above-described constituent elements, or those obtained by replacing the constituent elements and expressions of the present invention with each other between apparatuses, methods, systems, computer programs, recording media storing computer programs, and the like are also included in the present invention. It is effective as an embodiment of

  According to the present invention, the background noise level can be automatically detected.

It is a figure for demonstrating an example of the conventional phoneme division | segmentation apparatus. 2A to 2E are diagrams for explaining phoneme division processing by the phoneme division apparatus shown in FIG. It is a figure for demonstrating the phoneme division | segmentation apparatus which concerns on embodiment of this invention. FIGS. 4A to 4F are diagrams for explaining phoneme division processing by the phoneme division apparatus shown in FIG. It is a figure for demonstrating the phoneme division | segmentation apparatus which concerns on a modification. It is a figure for demonstrating the phoneme division | segmentation apparatus which concerns on another embodiment of this invention. It is a figure for demonstrating the phoneme division | segmentation apparatus which concerns on another embodiment of this invention. 8A to 8C are diagrams for explaining phoneme division processing by the phoneme division apparatus shown in FIG. It is a figure for demonstrating the speech processing system which concerns on another embodiment of this invention. It is a figure for demonstrating the noise measuring device which concerns on another embodiment of this invention. It is a figure for demonstrating the noise measuring device which concerns on another embodiment of this invention. It is a figure which shows an example of the noise level display by a display part.

  The present invention will be described below based on preferred embodiments with reference to the drawings. The same or equivalent components, members, and processes shown in the drawings are denoted by the same reference numerals, and repeated descriptions are omitted as appropriate.

  First, an example of a conventional phoneme dividing apparatus will be described before describing a phoneme dividing apparatus according to an embodiment of the present invention.

  FIG. 1 is a diagram for explaining an example of a conventional phoneme dividing apparatus. 1 includes a microphone amplifier 104, an absolute value circuit 106, a time constant circuit 108, a comparator 110, a variable resistor 112, and a gate circuit 114.

  The microphone amplifier 104 is connected to the microphone 102. The microphone 102 collects the original voice (Muskey) and converts it into a sound signal. This sound signal is a sound in which background noise is superimposed. The microphone amplifier 104 amplifies the sound signal from the microphone 102. The sound signal X (t) output from the microphone amplifier 104 is branched into two by the branching unit 105, one sound signal X (t) is input to the absolute value circuit 106, and the other sound signal X (t) Is input to the gate circuit 114. The absolute value circuit 106 outputs the absolute value of the sound signal X (t). The absolute value | X (t) | of the sound signal output from the absolute value circuit 106 is input to the time constant circuit 108. The time constant circuit 108 is a primary low-pass filter composed of a resistor having a resistance value R and a capacitor having a capacitance value C, and the time constant τ = RC is set to about 100 ms. The time constant circuit 108 smoothes the absolute value | X (t) | of the sound signal. By the smoothing processing by the time constant circuit 108, a component faster than the time constant τ is removed from the absolute value | X (t) | of the sound signal, and an envelope signal A (t) is obtained. The comparator 110 compares the envelope signal A (t) with a predetermined threshold T, and opens the gate circuit 114 at a timing when A (t) ≧ T. Thereby, the section of the sound signal when the gate circuit 114 is in the open state is divided and extracted as a phoneme (mora).

  In the phoneme dividing apparatus 100, the threshold value T is manually set by adjusting the power supply voltage + Vc with the variable resistor 112. In order to divide speech into phonemes with high accuracy, the setting of this threshold value T is important.

  2A to 2E are diagrams for explaining phoneme division processing by the phoneme division apparatus 100 shown in FIG. 2A to 2E, the vertical axis represents the signal level in arbitrary units, and the horizontal axis represents time t. FIG. 2A shows the waveform of the sound signal X (t) output from the microphone amplifier 104. FIG. 2B shows the waveform of the absolute value | X (t) | of the sound signal output from the absolute value circuit 106 and the waveform of the envelope signal A (t) output from the time constant circuit 108. The “approximately one mountain” of the waveform of the envelope signal A (t) corresponds to one phoneme (1 mora).

  As shown in FIG. 2B, in this example, the envelope signal A (t) includes six “substantially one mountain”, that is, phonemes 1 to 6. FIG. 2B shows three-stage threshold values T (threshold values T1 to T3) used when the phoneme is divided from the envelope signal A (t) in the comparator 110. 2C to 2E show waveforms of output signals from the gate circuit 114, that is, phoneme division results by the phoneme division device 100. FIG. The phoneme segmentation result by the phoneme segmentation apparatus 100 depends on the threshold value T.

  FIG. 2C shows a phoneme division result when the threshold value T is set to a threshold value T1 sufficiently larger than the background noise level. In a method of dividing a phoneme by comparing an envelope of a sound signal and a threshold as in the phoneme dividing apparatus 100, the threshold T is set to a value sufficiently larger than the background noise level in order to divide and extract the phoneme as stably as possible. It will be set. In this case, as shown in FIG. 2C, there is a possibility that a small level of phonemes such as phonemes 4 and 6 may be lost.

  FIG. 2E shows a phoneme division result when the threshold value T is set to a small threshold value T3 comparable to the background noise level. In this case, as shown in FIG. 2E, the adjacent phonemes 1 and 2 and the adjacent phonemes 3 to 5 are not divided. Thus, when the threshold value T is set to a small threshold value T3, the boundary between phonemes becomes ambiguous, and a plurality of phonemes may be connected without being divided as shown in FIG.

  FIG. 2D shows a phoneme division result when the threshold T is set to an optimum threshold T2 obtained by adding a margin H to the background noise level. In this case, phonemes 1 to 6 are appropriately divided as shown in FIG. Thus, in order to perform appropriate phoneme division, it is important to accurately detect the background noise level and set the threshold T to a value slightly larger than the background noise level.

  The general background noise in a room or space is almost constant in a short time, but in the long time, it tends to fluctuate considerably in the morning and afternoon, lunch and office hours. Therefore, even if the threshold value T is set based on the background noise level for a certain period of time, there is a possibility that appropriate phoneme division cannot be performed due to background noise level fluctuations. In addition, it is a difficult task to manually adjust the threshold T in accordance with the background noise level fluctuation.

  After recognizing the problems of the conventional phoneme dividing device as described above, the present inventor has made it possible to automatically perform appropriate phoneme division even when the background noise level fluctuates. Invented a method and apparatus.

  FIG. 3 is a diagram for explaining the phoneme dividing apparatus 10 according to the embodiment of the present invention. As shown in FIG. 3, the phoneme dividing apparatus 10 includes a microphone amplifier 14, a bandpass filter 15, a square circuit 16, an audio time constant circuit 18, a background noise time constant circuit 20, and an audio square root circuit. 22, a background noise square root circuit 24, a buffer amplifier 26, an adder 28, a variable resistor 29, a comparator 30, and a gate circuit 32.

  The microphone amplifier 14 is connected to the microphone 12. The microphone 12 collects original voice (masky) such as conversation and converts it into a sound signal. The microphone amplifier 14 amplifies the sound signal from the microphone 12.

  The band-pass filter 15 passes a signal component in a predetermined pass band in the amplified sound signal from the microphone amplifier 14. The band pass filter 15 has a pass band (for example, 100 Hz to 7 kHz, more preferably 250 Hz to 4 kHz) corresponding to the average spectrum of adult speech, for example. By providing the band-pass filter 15, the phoneme segmentation / extraction accuracy can be improved.

  The sound signal X (t) output from the bandpass filter 15 is obtained by superimposing the background noise signal n (t) of the room (space) on the sound signal x (t). That is, X (t) = x (t) + n (t).

  The sound signal X (t) output from the bandpass filter 15 is branched into two by the first branching unit 13. One sound signal X (t) branched by the first branching unit 13 is input to the square circuit 16, and the other sound signal X (t) is input to the gate circuit 32. As can be seen from FIG. 3, in the present embodiment, a bandpass filter 15 is provided in front of the first branching unit 13. Further, a square circuit 16 is provided between the first branch part 13 and the second branch part 17.

The square circuit 16 outputs a square signal X ² (t) of one sound signal X (t) branched by the first branching unit 13. The sound signal X (t) includes positive and negative values. Since the squaring circuit 16 squares the sound signal X (t), only a positive value needs to be processed, so that signal processing can be facilitated. The square signal X ² (t) output from the square circuit 16 is branched into two by the second branch unit 17. One square signal X ² (t) branched by the second branching unit 17 is input to the audio time constant circuit 18, and the other square signal X ² (t) is input to the background noise time constant circuit 20. .

The audio time constant circuit 18 is a primary low-pass filter including a first resistor 34 having a resistance value R and a second capacitor 36 having a capacitance value C. One terminal of the first resistor 34 is connected to the square circuit 16, and the other terminal is connected to the audio square root circuit 22. One terminal of the second capacitor 36 is connected to the other terminal of the first resistor 34, and the other terminal of the second capacitor 36 is grounded. The time constant (hereinafter referred to as “sound time constant”) τ _v = RC of the sound time constant circuit 18 is set to a relatively small value of several tens of ms to several hundreds of ms (for example, 125 ms). The audio time constant circuit 18 smoothes (averages) the square signal X ² (t) with the audio time constant τ _v . The smoothing by the audio time constant circuit 18 processes (averaging process), squared signal X ^{2 (t)} at the time for sound from the constant tau _v fast component than is removed, the envelope signal of the squared signal X ^{2 (t)} Is obtained.

The audio square root circuit 22 provided at the subsequent stage of the audio time constant circuit 18 calculates the square root of the signal input from the audio time constant circuit 18. The signal A (t) output from the sound square root circuit 22 can be regarded as an envelope of the sound signal x (t), that is, an effective value x _{rms of the} sound signal x (t) (see the following formula). . Hereinafter, A (t) is referred to as a “voice envelope signal”.

  The background noise time constant circuit 20 is a primary low-pass filter configured to have different time constants at the rise and fall of the input signal. The background noise time constant circuit 20 includes a second resistor 38 having a resistance value R ′, a second capacitor 40 having a capacitance C, a diode 42, and a third resistor 44 having a resistance value R. One terminal of the second resistor 38 is connected to the square circuit 16, and the other terminal is connected to the background noise square root circuit 24. The cathode terminal of the diode 42 is connected to the square circuit 16, and the anode terminal is connected to one terminal of the third resistor 44. The other terminal of the third resistor 44 is connected to the background noise square root circuit 24. One terminal of the second capacitor 40 is connected to the other terminal of the second resistor 38 and the third resistor 44, and the other terminal of the second capacitor 40 is grounded.

In the background noise time constant circuit 20 configured in this manner, the time constant (hereinafter referred to as “rising time constant”) composed of the second resistor 38 and the second capacitor 40 with respect to the rising of the input signal. Smoothing is performed with a low-pass filter of τ _u = R′C. On the other hand, the falling of the input signal is smoothed by a low-pass filter having a time constant (hereinafter referred to as “time constant for falling”) τ _d = RC composed of the third resistor 44 and the second capacitor 40. Is done.

In the background noise time constant circuit 20 according to the present embodiment, the rising time constant τ _u is set to a value that is much larger than the falling time constant τ _d . That is, the background noise time constant circuit 20 is composed of two asymmetric time constants. Specifically, the rising time constant τ _u is set to be at least 10 times, more preferably 100 times to 1000 times larger than the falling time constant τ _d . For example, τ _u = R′C ≧ 300τ _{d to} 3000τ _d may be set. On the other hand, the time constant tau _d for falling is set to the same value and the time constant tau _v audio voice for the time constant circuit 18. In the present embodiment, the falling time constant τ _d is equal to the audio time constant (ie, τ _d = τ _v = RC).

The background noise square root circuit 24 provided at the subsequent stage of the background noise time constant circuit 20 calculates the square root B (t) of the signal input from the background noise time constant circuit 20. Since the rising time constant τ _u of the background noise time constant circuit 20 is set to a value that is much larger than the sound time constant τ _v of the sound time constant circuit 18, B (t) is the sound signal x ( t) is almost insensitive to changes in level, and is generally considered to be almost constant (eg, background noise level in a bank lobby or hospital waiting room), that is, near the lowest level of the sound signal X (t) Maintained. That is, B (t) quickly decreases to the background noise level through a slight silence (discontinuity) between phonemes (mora), and always maintains a level equal to background noise.

However, in some cases, a loud speaker may speak continuously or a short start bell will ring, but the signal B (t) will rise slowly, but promptly when they stop. In order to return to the original background noise level, the falling time constant τ _d is different from the rising time constant τ _u , specifically, the same as the audio time constant τ _v of the audio time constant circuit 18. It is a time constant. As a result, the signal B (t) follows a slow background noise change in a symmetric space from early morning to morning and from lunch to afternoon, but almost follows a quick level change such as speech. It would be insensitive. The signal B (t) changing in this way is an envelope of the background noise signal n (t), that is, an effective value n _{rms of the} background noise signal n (t), with respect to the voice envelope signal A (t) changing every moment. (See formula below). Hereinafter, B (t) is referred to as a “background noise envelope signal”.

  The comparator 30 compares the speech envelope signal A (t) output from the speech square root circuit 22 with the background noise envelope signal B (t) output from the background noise square root circuit 24. Here, in this embodiment, before inputting to the comparator 30, the background noise envelope signal B (t) is amplified with a predetermined amplification factor m using the buffer amplifier 26 and further added with the adder 28. The offset value h is added. That is, in the present embodiment, the speech envelope signal A (t) and B ′ (t) = mB (t) + h are compared by the comparator 30. Hereinafter, B ′ (t) = mB (t) + h is referred to as “threshold signal”. The amplification factor m may be selected in the range of m = 1 to 3, for example. The offset value h may be selected from a range of h = 0 to [about 10 times the maximum value assumed for B (t)], for example. In general, the background noise level is sufficiently lower than the voice level. Thus, by comparing the threshold signal B ′ (t) slightly larger than B (t) with the voice envelope signal A (t), the phoneme division is performed. Can be performed safely and stably. In the modification, the voice envelope signal A (t) and the background noise envelope signal B (t) may be directly compared.

  The comparator 30 outputs a high level to the gate circuit 32 in a section where the voice envelope signal A (t) is equal to or greater than the threshold signal B ′ (t) (that is, A (t) ≧ B ′ (t)), A low level is output to the gate circuit 32 in a section where the voice envelope signal A (t) is less than the threshold signal B ′ (t) (that is, A (t) <B ′ (t)).

  The gate circuit 32 controls passage / non-passage of the other sound signal X (t) branched by the first branching unit 13 according to the comparison result of the comparator 30. That is, the gate circuit 32 is opened when the high level is received from the comparator 30 and passes the sound signal X (t), and is closed when the low level is received from the comparator 30. Let X (t) not pass. With such an operation, a phoneme signal is output from the gate circuit 32.

  FIGS. 4A to 4F are diagrams for explaining phoneme division processing by the phoneme division apparatus 10 shown in FIG. 4A to 4F, the vertical axis represents the signal level v in the unit mV, and the horizontal axis represents the time t in the unit ms.

  FIG. 4A shows the waveform of the sound signal X (t) output from the bandpass filter 15. The sound signal X (t) is obtained by superimposing the background noise signal n (t) on the audio signal x (t). The sound signal X (t) is branched into two by the first branching unit 13. One of the branched sound signals X (t) is input to the square circuit 16, and the other sound signal X (t) is input to the gate circuit 32.

FIG. 4B shows the waveform of the square signal X ² (t) output from the square circuit 16. As shown in FIG. 4B, the square signal X ² (t) includes only a positive component. The square signal X ² (t) is branched into two by the second branching unit 17. One of the branched square signals X ² (t) is input to the audio time constant circuit 18, and the other square signal X ² (t) is input to the background noise time constant circuit 20.

The signal smoothed by the audio time constant circuit 18 is square-rooted by the audio square root circuit 22. This square root becomes the voice envelope signal A (t). FIG. 4C shows the waveform of the sound envelope signal A (t) output from the sound square root circuit 22. As shown in FIG. 4C, the speech envelope signal A (t) is a positive waveform that changes substantially following the effective value x _rms of the input original speech.

  On the other hand, the square root of the signal smoothed by the time constant circuit 20 for background noise is taken by the square root circuit 24 for background noise. This square root becomes the background noise envelope signal B (t). FIG. 4D shows the waveform of the background noise envelope signal B (t) output from the background noise square root circuit 24. As shown in FIG. 4 (d), the background noise envelope signal B (t) hardly follows the input original voice, and rapidly reaches the minimum value, that is, darkness only along the interrupted portion of the input original voice. Reduces to noise level. That is, B (t) is always maintained at the background noise level, and can be used as a phoneme division threshold.

  The background noise envelope signal B (t) output from the background noise square root circuit 24 is amplified m times by the buffer amplifier 26, and then the offset value h is added by the adder 28, whereby the threshold signal B '(t) ) = MB (t) + h. FIG. 4D shows a waveform of the threshold signal B ′ (t) in addition to the background noise envelope signal B (t).

  FIG. 4C illustrates the threshold signal B ′ (t) = mB (t) + h in addition to the voice envelope signal A (t). That is, in FIG. 4C, two signals to be compared by the comparator 30 are shown. As shown in FIG. 4C, the intersection of the voice envelope signal A (t) and the threshold signal B '(t) is obtained. FIG. 4E shows the output signal of the comparator 30. The comparator 30 sets the high level in the section where A (t) ≧ B ′ (t) among the sections effectively sandwiched between the intersections of the voice envelope signal A (t) and the threshold signal B ′ (t). And outputs a low level in a section where A (t) <B ′ (t).

  FIG. 4F shows the output signal of the gate circuit 32. The gate circuit 32 passes the note X (t) only when it receives a high level from the comparator 30, and does not pass the note X (t) when it receives a low level from the comparator 30. As a result, as shown in FIG. 4 (f), phonemes and background noise are clearly divided, and three phonemes are divided and extracted.

  The phoneme dividing device 10 according to the present embodiment has been described above. According to this phoneme dividing apparatus 10, since the background noise level is automatically detected, the threshold for dividing and extracting phonemes is always maintained at an optimum value even if the background noise changes with time. As a result, phoneme division can be performed with higher accuracy than in the past.

  According to the phoneme dividing apparatus 10 of the present embodiment, the work of manually adjusting the threshold value T in accordance with the fluctuation in the level of background noise is not required, so that significant rationalization and labor saving can be achieved.

  FIG. 5 is a diagram for explaining a phoneme dividing apparatus 50 according to a modification. The phoneme division apparatus 50 shown in FIG. 5 is different from the phoneme division apparatus 10 shown in FIG. 3 in that the bandpass filter 15 is provided between the first branching unit 13 and the square circuit 16.

  In the phoneme division device 50 according to the present modification, the bandpass filter 15 is provided between the first branch unit 13 and the square circuit 16, so that a sound signal that does not pass through the bandpass filter is received in the gate circuit 32. Entered. Therefore, since the phoneme division device 50 can obtain a phoneme signal closer to the original speech signal, the sound quality can be improved as compared with the phoneme division device 10 shown in FIG. Note that in the phoneme dividing device 50, the sound signal entering the square circuit 16 passes through the bandpass filter 15, and therefore the phoneme dividing / extracting accuracy is equivalent to that of the phoneme dividing device shown in FIG.

  FIG. 6 is a diagram for explaining a phoneme division device 60 according to another embodiment of the present invention. The phoneme dividing device 60 shown in FIG. 6 is different from the phoneme dividing device 10 shown in FIG. 3 in that an absolute value circuit 62 is provided instead of the square circuit.

  As shown in FIG. 6, the phoneme division device 60 includes a microphone amplifier 14, an absolute value circuit 62, a sound time constant circuit 18, a background noise time constant circuit 20, a buffer amplifier 26, an adder 28, and the like. The variable resistor 29, the comparator 30, and the gate circuit 32 are provided.

  The microphone amplifier 14 is connected to the microphone 12. The microphone 12 collects original voice (masky) such as conversation and converts it into a sound signal. The microphone amplifier 14 amplifies the sound signal from the microphone 12. The sound signal X (t) amplified by the microphone 12 is obtained by superimposing the background noise signal n (t) on the audio signal x (t).

  The sound signal X (t) output from the microphone amplifier 14 is branched into two by the first branching unit 13. One sound signal X (t) branched by the first branching unit 13 is input to the absolute value circuit 62, and the other sound signal X (t) is input to the gate circuit 32. Similarly to the phoneme dividing apparatus 10 shown in FIG. 3, a bandpass filter may be provided in the preceding stage of the first branching unit 13. Alternatively, a band pass filter may be provided between the first branching unit 13 and the absolute value circuit 62 as in the phoneme dividing device 50 shown in FIG. An absolute value circuit 62 is provided between the first branch unit 13 and the second branch unit 17.

  The absolute value circuit 62 outputs an absolute value | X (t) | of one sound signal X (t) branched by the first branching unit 13. As in the embodiment using the square circuit, the absolute value circuit 62 takes the absolute value of the sound signal X (t), and only a positive value needs to be processed, thereby facilitating signal processing. . The absolute value signal | X (t) | output from the absolute value circuit 62 is branched into two by the second branching unit 17. One absolute value signal | X (t) | branched by the second branching unit 17 is input to the audio time constant circuit 18, and the other absolute value signal | X (t) | Is input.

The audio time constant circuit 18 is a primary low-pass filter including a first resistor 34 having a resistance value R and a second capacitor 36 having a capacitance value C. The audio time constant circuit 18 smoothes (averages) the absolute value signal | X (t) | with an audio time constant τ _v of several tens to several hundreds of ms. The signal A (t) output from the audio time constant circuit 18 can be regarded as an envelope of the audio signal x (t), that is, an effective value x _{rms of the} audio signal x (t). Hereinafter, A (t) is referred to as a “voice envelope signal”.

The background noise time constant circuit 20 is a primary low-pass filter configured to have different time constants at the rise and fall of the input signal. In the background noise time constant circuit 20, the rising of the input signal is smoothed by a low-pass filter having a rising time constant τ _u = R′C. On the other hand, the falling of the input signal is smoothed by a low-pass filter having a falling time constant τ _d = RC.

In the background noise time constant circuit 20 according to the present embodiment, the rising time constant τ _u is set to a value that is much larger than the falling time constant τ _d . Specifically, the rising time constant τ _u is set to be at least 10 times, more preferably 100 times to 1000 times larger than the falling time constant τ _d . For example, τ _u = R′C ≧ 300τ _{d to} 3000τ _d may be set. On the other hand, the falling time constant τ _d is set to substantially the same value as the audio time constant τ _v of the audio time constant circuit 18. In the present embodiment, the falling time constant τ _d is equal to the audio time constant (ie, τ _d = τ _v = RC).

The signal B (t) smoothed by the background noise time constant circuit 20 is almost insensitive to the level change of the audio signal x (t) and is maintained near the background noise level. The signal B (t) can be regarded as an envelope of the background noise signal n (t), that is, an effective value n _{rms of the} background noise signal n (t). Hereinafter, B (t) is referred to as a “background noise envelope signal”.

  The comparator 30 compares the voice envelope signal A (t) output from the voice time constant circuit 18 with the background noise envelope signal B (t) output from the background noise time constant circuit 20. Here, in this embodiment, before inputting to the comparator 30, the background noise envelope signal B (t) is amplified with a predetermined amplification factor m using the buffer amplifier 26 and further added with the adder 28. The offset value h is added. That is, in the present embodiment, the speech envelope signal A (t) and B ′ (t) = mB (t) + h are compared by the comparator 30. Hereinafter, B ′ (t) = mB (t) + h is referred to as “threshold signal”. The amplification factor m may be selected in the range of m = 1 to 3, for example. The offset value h may be selected from a range of h = 0 to [about 10 times the maximum value assumed for B (t)], for example. In the modification, the voice envelope signal A (t) and the background noise envelope signal B (t) may be directly compared.

  The comparator 30 outputs a high level to the gate circuit 32 in a section where the voice envelope signal A (t) is equal to or greater than the threshold signal B ′ (t) (that is, A (t) ≧ B ′ (t)), A low level is output to the gate circuit 32 in a section where the voice envelope signal A (t) is less than the threshold signal B ′ (t) (that is, A (t) <B ′ (t)).

  The gate circuit 32 controls passage / non-passage of the other sound signal X (t) branched by the first branching unit 13 according to the comparison result of the comparator 30. That is, the gate circuit 32 is opened when the high level is received from the comparator 30 and passes the sound signal X (t), and is closed when the low level is received from the comparator 30. Let X (t) not pass. With such an operation, a phoneme signal is output from the gate circuit 32.

  The phoneme dividing device 60 according to the present embodiment has been described above. Also in this phoneme dividing device 60, since the level of background noise is automatically detected, the threshold for dividing and extracting phonemes is always maintained at an optimum value even if the background noise changes with time. As a result, phoneme division can be performed with higher accuracy than in the past.

  In the phoneme dividing apparatus 60 according to the present embodiment as well, it is not necessary to manually adjust the threshold value T in accordance with the fluctuation in the level of background noise, so that significant rationalization and labor saving can be achieved.

  FIG. 7 is a diagram for explaining a phoneme dividing device 70 according to still another embodiment of the present invention. In the embodiment shown in FIGS. 3, 5, and 6, the phoneme division process is realized by an analog circuit, but in the present embodiment shown in FIG. 7, the phoneme division process is realized by software.

  The phoneme dividing device 70 includes a microphone amplifier 14 and a DSP (Digital Signal Processor) board 71. The DSP board 71 includes an input amplifier 72, an A / D converter 73, a DSP 74, a D / A converter 75, an output amplifier 76, a ROM 77, an SD-RAM 78, an input port 79, and an output port. 80 is implemented.

  The microphone amplifier 14 is connected to the microphone 12. The microphone 12 collects original voice (masky) such as conversation and converts it into a sound signal. The microphone amplifier 14 amplifies the sound signal from the microphone 12. The sound signal X (t) amplified by the microphone 12 is input to the input port 79 of the DSP board 71. The sound signal X (t) is an analog signal, and a background noise signal is superimposed on a sound signal. The sound signal X (t) input from the input port 79 is amplified by the input amplifier 72 and then converted to a digital signal by the A / D converter 73. The digital signal of the sound signal X (t) output from the A / D converter 73 is input to the DSP 74.

  The DSP 74 is connected to a ROM 77 that stores a program for performing phoneme division processing and an SD-RAM 78 that stores data being processed by the DSP 74. The DSP 74 reads a phoneme division program from the ROM 77 and performs phoneme division processing.

The phoneme division program stored in the ROM 77 squares the DSP 74 with the first branch step for branching the sound signal X (t) into two and the one sound signal X (t) branched at the first branch step. and square step, a second branching step of branching squared signal ^X 2 (t) into two, when a voice of one squared signal ^X 2 (t) several tens to several hundreds of 100ms to branched by the second branching step a first smoothing step for smoothing constant tau _v, a first square root operation step of calculating the square root of the smoothed signal by the first smoothing step, the other squared signals X branched by the second branching step ² (t) rising to at least 10 times more than the time constant tau _v for voice, more preferably with smoothing in time constant tau _u for greater rise above 100 to 1000 times, the other squared signals ^{X 2} a second smoothing step for smoothing a time constant tau _d for falling substantially the same standing as the time constant tau _v audio for the fall of t), the square root of the smoothed signal by the second smoothing step The second square root calculation step to be calculated is compared with the voice envelope signal A (t) calculated in the first square root calculation step and the background noise envelope signal B (t) calculated in the second square root calculation step. It may be a program for executing a comparison step and a passage control step for controlling passage / non-passage of the other sound signal branched in the first branch step according to the comparison result of the comparison step. In the comparison step of this program, the voice envelope signal A (t) may be compared with the threshold signal B ′ (t) = mB (t) + h. The amplification factor m may be selected in the range of m = 1 to 3, for example. The offset value h may be selected from a range of h = 0 to [about 10 times the maximum value assumed for B (t)], for example.

Alternatively, the phoneme division program stored in the ROM 77 can be obtained by feeding the DSP 74 the first branch step for branching the sound signal X (t) into two and one of the sound signals X (t) branched in the first branch step. An absolute value calculation step for calculating the absolute value | X (t) |, a second branch step for branching the absolute value signal | X (t) | from the absolute value calculation step into two, and a branch at the second branch step First absolute value signal | X (t) | smoothed with a time constant τ _v for speech of several tens to several hundreds of ms, and the other absolute value signal branched in the second branch step The rise of | X (t) | is smoothed with a rise time constant τ _u that is at least 10 times, more preferably 100 to 1000 times greater than the audio time constant τ _v , and the other absolute value Signal | X (t) | A second smoothing step for smoothing a time constant tau _d for falling substantially the same standing as the time constant tau _v audio for falling, first smoothing step smoothing speech envelope signal A (t) And the background noise signal B (t) smoothed in the second smoothing step, and the other sound signal branched in the first branching step according to the comparison result of the comparison step And a passage control step for controlling passage / non-passage of X (t). Also in the comparison step of this program, the voice envelope signal A (t) may be compared with the threshold signal B ′ (t) = mB (t) + h. The amplification factor m may be selected in the range of m = 1 to 3, for example. The offset value h may be selected from a range of h = 0 to [about 10 times the maximum value assumed for B (t)], for example.

  The phoneme digital signal output from the DSP 74 is converted into an analog signal by the D / A converter 75, amplified by the output amplifier 76, and output from the output port 80.

  8A to 8C are diagrams for explaining phoneme division processing by the phoneme division device 70 shown in FIG. 8A to 8C, the vertical axis represents signal level in arbitrary units, and the horizontal axis represents time in arbitrary units.

  FIG. 8A shows the waveform of the voice envelope signal A (t). FIG. 8B shows the waveform of the background noise envelope signal B (t). FIG. 8C shows a waveform (ie, A (t) −B ′ (t)) comparing the voice envelope signal A (t) and the threshold signal B ′ (t) = mB (t) + h. Here, the amplification factor m = 1 and the offset value h = 200 are set. As shown in FIGS. 8A and 8B, when the phoneme division process is performed by software, the sound envelope signal A (t) and the background noise envelope signal B (t) are appropriately obtained from the sound signal. Can be separated. Since the background noise level is automatically detected, as shown in FIG. 8C, the phoneme segmentation / extraction is performed with high accuracy by comparing the voice envelope signal A (t) and the threshold signal B ′ (t). It can be carried out. Further, according to the phoneme dividing apparatus 70 of the present embodiment, the work of manually adjusting the threshold value T in accordance with the background noise level fluctuation is not required, so that significant rationalization and labor saving can be achieved.

  FIG. 9 is a diagram for explaining an audio processing system 90 according to still another embodiment of the present invention. The speech processing system 90 performs predetermined processing on speech input using the above-described phoneme splitting device and outputs it to a space.

  As shown in FIG. 9, the sound processing system 90 includes a microphone 12 as a sound collecting device, a microphone amplifier 14, a phoneme dividing device 92, a phoneme processing device 94, an amplifier 95, and a speaker 96 as an output device. Is provided. The microphone 12 collects the original sound and outputs a sound signal in which a background noise signal is superimposed on the sound signal. The microphone amplifier 14 amplifies the sound signal from the microphone 12. The phoneme dividing device 92 receives the amplified sound signal from the microphone amplifier 14 and divides the audio signal into phonemes. As the phoneme dividing device 92, the above-described phoneme dividing devices 10, 50, 60, and 70 can be preferably used. The phoneme processing device 94 performs predetermined processing on the phoneme signal obtained from the phoneme processing device 94. An example of this predetermined process will be described later. The amplifier 95 amplifies the phoneme signal processed by the phoneme processing device 94. The speaker 96 outputs the amplified phoneme signal as a sound to the space.

  The voice processing system 90 may be, for example, a speech privacy system (voice information secret talk device). The speech privacy system changes and processes the structure of the audio signal in real time through signal processing, so that only the content of the audio is obtained without greatly changing the statistical properties such as the spectrum of the audio signal and the energy envelope. It hides / blocks and makes the contents of the conversation unintelligible to the listener. Unlike conventional voice masking systems, this speech privacy system produces no sound except when the original voice is generated (during utterance), thus increasing the noise level of the room and increasing the listener's discomfort. Only the contents can be effectively concealed. For details of the speech privacy system, see, for example, Patent Document 1 described above.

  In the speech privacy system, the phoneme processing device 94 rearranges the phonemes (mora) divided and extracted by the phoneme dividing device 92, for example, rearranges the order of the phonemes. The rearranged phoneme signals are output from the speaker 96 as sound to the space. Since the original voice is masked by the sound from the speaker 96, the contents of the original voice can be made unintelligible to the listener.

Alternatively, the voice processing system 90 may be a communication system such as a mobile phone, a radio, or a transceiver. For example, when a mobile phone is used at a construction site, under a guard, or at a railway platform, background noise is superimposed on the received voice on the receiving side, which hinders understanding of the conversation content. That is, listening and sentence comprehension are reduced. Therefore, the phoneme processing device 94 sets the output between the phonemes divided and extracted by the phoneme dividing device 92 (that is, the discontinuous part of the conversation) to zero (silence). By outputting the phoneme signal thus processed from the speaker 96, it is possible to reduce noise and improve call quality. In such a communication system, it is preferable that the rising time constant τ _u = R′C in the background noise time constant circuit of the phoneme dividing device 92 is set smaller than that used in the speech privacy system. Note that background noise is still superimposed on the audio part, but the background noise is hardly recognized due to auditory complementation, and perceived as if the noise was almost eliminated. Therefore, listening is greatly improved.

Or the phoneme division | segmentation apparatus which concerns on the above-mentioned embodiment may be used for the vehicle-mounted navigation system which included the speech recognition function. By using the phoneme signal output from the above phoneme splitting device for speech recognition, the speech recognition rate can be improved without being affected by the constantly changing traveling noise. In this case, the rise time constant τ _u = R′C in the background noise time constant circuit of the phoneme division apparatus is preferably set smaller than that used in the speech privacy system.

  Alternatively, the phoneme division device according to the above-described embodiment may be used for a VOX (Voice Operating tX; tx = Transmitter) function of half-duplex communication. By accurately grasping the occurrence of an utterance based on the phoneme signal output from the phoneme dividing device, it is possible to switch between transmission and reception with certainty.

  FIG. 10 is a diagram for explaining a noise measurement device 120 according to still another embodiment of the present invention. The noise measurement device 120 shown in FIG. 10 can measure the level of background noise included in the ambient sound. The noise measuring device 120 includes a microphone 12, a microphone amplifier 14, a bandpass filter 15, a square circuit 16, a background constant circuit 20 for background noise, and a square root circuit 24 for background noise.

  The microphone 12 collects ambient sounds and converts them into sound signals. The microphone amplifier 14 amplifies the sound signal from the microphone 12. The band-pass filter 15 passes a signal component in a predetermined pass band in the amplified sound signal from the microphone amplifier 14. The sound signal X (t) output from the bandpass filter 15 is obtained by superimposing the background noise signal n (t) of the room (space) on the sound signal x (t).

The square circuit 16 outputs a square signal X ² (t) of the sound signal X (t). The sound signal X (t) includes positive and negative values. Since the squaring circuit 16 squares the sound signal X (t), only a positive value needs to be processed, so that signal processing can be facilitated. The square circuit 16 may be replaced with an absolute value circuit. In this case, the background noise square root circuit 24 is not necessary.

The background noise time constant circuit 20 is a primary low-pass filter configured to have different time constants at the rise and fall of the input signal. In the background noise time constant circuit 20, the rising of the input signal is smoothed by a low-pass filter having a rising time constant τ _u = R′C. On the other hand, the falling of the input signal is smoothed by a low-pass filter having a falling time constant τ _d = RC.

In the background noise time constant circuit 20 according to the present embodiment, the rising time constant τ _u is set to a value that is much larger than the falling time constant τ _d . Specifically, the falling time constant τ _d is set to a relatively small value of several tens of ms to several hundreds of ms (for example, 125 ms). On the other hand, the rising time constant τ _u is set at least 10 times, more preferably 100 times to 1000 times larger than the falling time constant τ _d . For example, τ _u = R′C ≧ 300τ _{d to} 3000τ _d may be set.

The background noise square root circuit 24 provided at the subsequent stage of the background noise time constant circuit 20 calculates the square root B (t) of the signal input from the background noise time constant circuit 20. As described above, the signal B (t) can be regarded as an envelope of the background noise signal n (t), that is, an effective value n _rms (that is, background noise level) of the background noise signal n (t).

  Thus, according to the noise measurement apparatus 120 according to the present embodiment, the level of background noise included in the ambient sound can be measured. The noise measuring device 120 according to the present embodiment is particularly effective for measuring background noise in a space where there is conversation or specific variable noise (significant noise). The noise measurement device 120 may include a display unit that displays a signal output from the background noise square root circuit 24. In this case, background noise can be visually recognized.

  The noise measuring device 120 can be used for an acoustic system such as a television system, an in-vehicle television system, a car stereo system, and the like. The level of background noise measured by the noise measuring device 120 is not affected by conversational voices between the occupants and intermittent noise for a short time. Therefore, referring to this background noise level, for example, the speaker volume is increased when the background noise level is high, and the speaker volume is decreased when the background noise level is low. It can be controlled optimally.

  FIG. 11 is a diagram for explaining a noise measurement device 130 according to still another embodiment of the present invention. The noise measuring device 130 shown in FIG. 11 can measure the level of noise included in the ambient sound. The noise measurement device 130 includes a microphone 12, a microphone amplifier 14, a bandpass filter 15, a square circuit 16, a sound time constant circuit 18, a background noise time constant circuit 20, a sound square root circuit 22, The background noise square root circuit 24 and a display unit 134 are provided.

  The microphone 12 collects ambient sounds and converts them into sound signals. The microphone amplifier 14 amplifies the sound signal from the microphone 12. The band-pass filter 15 passes a signal component in a predetermined pass band in the amplified sound signal from the microphone amplifier 14. The sound signal X (t) output from the bandpass filter 15 is obtained by superimposing the background noise signal n (t) of the room (space) on the sound signal x (t).

The square circuit 16 outputs a square signal X ² (t) of the sound signal X (t). The sound signal X (t) includes positive and negative values. Since the squaring circuit 16 squares the sound signal X (t), only a positive value needs to be processed, so that signal processing can be facilitated. The square circuit 16 may be replaced with an absolute value circuit. In this case, the voice square root circuit 22 and the background noise square root circuit 24 are not required.

The square signal X ² (t) output from the square circuit 16 is branched into two by the branch unit 132. One square signal X ² (t) branched by the branch unit 132 is input to the audio time constant circuit 18, and the other square signal X ² (t) is input to the background noise time constant circuit 20.

The audio time constant circuit 18 is a first-order low-pass filter having an audio time constant τ _v of several tens to several hundreds of ms. The audio time constant circuit 18 smoothes (averages) the input signal with the audio time constant τ _v . The signal A (t) output from the audio time constant circuit 18 may be regarded as an envelope of the audio signal x (t), that is, an effective value x _rms (that is, an audio signal level) of the audio signal x (t). it can.

The background noise time constant circuit 20 is a primary low-pass filter configured to have different time constants at the rise and fall of the input signal. In the background noise time constant circuit 20, the rising of the input signal is smoothed by a low-pass filter having a rising time constant τ _u = R′C. On the other hand, the falling of the input signal is smoothed by a low-pass filter having a falling time constant τ _d = RC. The smoothing by the audio time constant circuit 18 processes (averaging process), squared signal X ^{2 (t)} at the time for sound from the constant tau _v fast component than is removed, the envelope signal of the squared signal X ^{2 (t)} Is obtained.

The audio square root circuit 22 provided at the subsequent stage of the audio time constant circuit 18 calculates the square root of the signal input from the audio time constant circuit 18. The signal A (t) output from the audio square root circuit 22 can be regarded as an envelope of the audio signal x (t), that is, an effective value x _rms (that is, an audio signal level) of the audio signal x (t). it can.

In the background noise time constant circuit 20 according to the present embodiment, the rising time constant τ _u is set to a value that is much larger than the falling time constant τ _d . Specifically, the rising time constant τ _u is set to be at least 10 times, more preferably 100 times to 1000 times larger than the falling time constant τ _d . For example, τ _u = R′C ≧ 300τ _{d to} 3000τ _d may be set. On the other hand, the falling time constant τ _d is set to substantially the same value as the audio time constant τ _v of the audio time constant circuit 18. In the present embodiment, the falling time constant τ _d is equal to the audio time constant (ie, τ _d = τ _v = RC).

The background noise square root circuit 24 provided at the subsequent stage of the background noise time constant circuit 20 calculates the square root B (t) of the signal input from the background noise time constant circuit 20. As described above, the signal B (t) can be regarded as an envelope of the background noise signal n (t), that is, an effective value n _rms (that is, background noise level) of the background noise signal n (t).

The display unit 134 displays the signal A (t) from the audio square root circuit 22 and the signal B (t) from the background noise square root circuit 24. For example, the display unit 134 may distinguish and display the signal A (t) and the signal B (t) in a binary manner. The signal A (t) represents a conversational voice that changes in a short time or an intermittent noise level of the construction site, and a signal B (t) represents a background noise level. For the signal A (t), A (t) includes background noise. In view of this, the display unit 134 can display “the background noise corrected signal A ′ (t)” obtained according to the following formula as the “net noise level” instead of or in addition to the signal A (t). .

  FIG. 12 shows an example of noise level display by the display unit 134. In the noise level display example shown in FIG. 12, the signal A (t) from the voice square root circuit 22 is displayed as “total noise”, and the signal B (t) from the background noise square root circuit 24 is displayed as “background noise”. The signal A ′ (t) displayed and corrected for background noise is displayed as “net noise”. As shown in FIG. 12, by visualizing the levels of total noise, background noise, and net noise, the user can grasp each noise level instantaneously.

  Although the present invention has been described based on the embodiments, the embodiments merely show the principle and application of the present invention, and the embodiments are defined in the claims. Needless to say, many modifications and arrangements can be made without departing from the spirit of the present invention.

  Envelope acquisition for the input signal, that is, envelope detection is not only the method of taking the square root of the mean square value described in the above embodiment and the method of smoothing the absolute value, but also the wavelet transform, the Hilbert transform, or simply May be performed by other similar methods such as a method of smoothing the result of half-wave rectification by a diode or the like.

  10, 50, 60, 70, 92 Phoneme splitting device, 12 microphone, 13 first branching unit, 14 microphone amplifier, 15 bandpass filter, 16 square circuit, 17 second branching unit, 18 time constant circuit for voice, 20 dark Time constant circuit for noise, 22 Square root circuit for sound, 24 Square root circuit for background noise, 26 Buffer amplifier, 28 Adder, 30 Comparator, 32 Gate circuit, 62 Absolute value circuit, 71 DSP board, 74 DSP, 76 Output amplifier 77 ROM, 90 voice processing system, 94 phoneme processing device, 96 speaker, 120, 130 noise measuring device, 134 display unit.

Claims (10)

A phoneme dividing device that divides a sound signal in which a background noise signal is superimposed on a sound signal into phonemes that are substantially a single envelope signal based on the sound signal,
A first branching section for branching the sound signal into two;
A second branching section for branching the one sound signal branched at the first branching section into two further;
A sound time constant part for smoothing one sound signal branched by the second branch part with a sound time constant of several tens to several hundreds of milliseconds;
The rising edge of the other sound signal branched by the second branching unit is smoothed with a rising time constant that is at least 10 times larger than the sound time constant, and the rising edge of the other sound signal is A background time constant portion for smoothing with a time constant for falling substantially the same as the time constant for sound,
A comparison unit that compares the signal from the time constant unit for sound and the signal from the time constant unit for background noise;
A gate unit for controlling passage / non-passage of the other sound signal branched by the first branch unit according to a comparison result of the comparison unit;
A phoneme segmentation device comprising:   2. The phoneme division apparatus according to claim 1, wherein the rising time constant is 100 to 1000 times larger than the voice time constant. An amplifier that amplifies the signal from the time constant portion for background noise at a predetermined amplification rate;
An adder for adding a predetermined offset value to the output from the amplifier, and
The phoneme division apparatus according to claim 1, wherein the comparison unit compares a signal from the time constant unit for speech with a signal from the adder. The first is provided in front of the branching portion, the phoneme splitting apparatus according to any one of claims 1 to 3, further comprising a band-path filter having a pass band corresponding to the average spectrum of speech. Either it provided from claim 1, further comprising a band-path filter having a pass band corresponding to the average spectrum of the speech 3 between the second branch portion and the first branch portion A phoneme division device according to claim 1. A square section that squares one of the sound signals branched by the first branch section, provided between the second branch section and the second branch section after the first branch section;
A voice square root calculation unit for calculating a square root of a signal from the voice time constant unit, provided at a subsequent stage of the voice time constant unit;
A background noise square root calculation unit for calculating a square root of the signal from the background noise time constant unit, which is provided at a subsequent stage of the background noise time constant unit,
The phoneme dividing apparatus according to claim 1, further comprising:   An absolute value calculation unit that is provided between the first branching unit and the second branching unit and outputs an absolute value of one of the sound signals branched by the first branching unit is further provided. The phoneme division device according to any one of claims 1 to 5. A sound collector that collects the original sound and outputs a sound signal in which a background noise signal is superimposed on the sound signal;
Wherein receiving the sound signal from the sound collecting device, the phoneme splitting device according to any one of 7 to claim 1 for dividing the sound signal into phonemes,
A phoneme processing device that performs predetermined processing on a phoneme signal obtained from the phoneme splitting device;
An output device that outputs the phoneme signal processed by the phoneme processing device as a sound to a space;
A speech processing system comprising: A phoneme division method that divides a sound signal in which a background noise signal is superimposed on an audio signal into phonemes that are substantially a mountain of envelope signals based on the sound signal,
A first branching step for branching the sound signal into two;
A second branching step for branching one of the sound signals branched in the first branching step into two;
A first smoothing step of smoothing one of the sound signals branched in the second branching step with an audio time constant of several tens to several hundreds of ms;
The rising edge of the other sound signal branched in the second branching step is smoothed with a rising time constant that is at least 10 times larger than the sound time constant, and the rising edge of the other sound signal is A second smoothing step of smoothing with a falling time constant substantially the same as the sound time constant;
A comparison step for comparing the signal calculated in the first smoothing step with the signal calculated in the second smoothing step;
A passage control step for controlling passage / non-passage of the other sound signal branched in the first branch step according to a comparison result of the comparison step;
A phoneme segmentation method comprising: A phoneme division program that divides a sound signal in which a background noise signal is superimposed on an audio signal into phonemes that are substantially a single envelope signal based on the sound signal,
On the computer,
A first branching step for branching the sound signal into two;
A second branching step for branching one of the sound signals branched in the first branching step into two;
A first smoothing step of smoothing one of the sound signals branched in the second branching step with an audio time constant of several tens to several hundreds of ms;
The rising edge of the other sound signal branched in the second branching step is smoothed with a rising time constant that is at least 10 times larger than the sound time constant, and the rising edge of the other sound signal is A second smoothing step of smoothing with a falling time constant substantially the same as the sound time constant;
A comparison step for comparing the signal calculated in the first smoothing step with the signal calculated in the second smoothing step;
A passage control step for controlling passage / non-passage of the other sound signal branched in the first branch step according to a comparison result of the comparison step;
Phoneme segmentation program to execute.

Images