JP4745916B2 - Noise suppression speech quality estimation apparatus, method and program - Google Patents

Description

  The present invention relates to a voice quality evaluation technique in a voice communication service using a noise suppression processing technique, and more particularly to a technique for evaluating voice quality in a communication environment where the influence of ambient noise is large.

  In voice communication in an environment with a high ambient noise, the noise is mixed in the transmitter, so that the listener listens to the voice with the superimposed noise. In hands-free communication, since the distance between the human mouth and the microphone is longer than when using a handset or headset, the sound collection range of the microphone is widened and is easily affected by ambient noise. Further, since voice communication using a mobile phone is often used in an outdoor environment where there is a large amount of ambient noise, the handset communication is also susceptible to ambient noise. Therefore, a noise suppression processing technique is important in such a communication form.

Conventionally, noise suppression processing techniques using various methods have been developed. In order to provide a high-quality voice communication service, it is important to accurately grasp the performance of the noise suppression processing technology, optimize the parameters of the method, and select the method. Therefore, a quality evaluation method for noise-suppressed speech is desired.
The basis of voice quality evaluation is subjective quality evaluation based on psychological evaluation by actually listening to voice or talking. Subjective quality evaluation can directly evaluate the quality perceived by the user, but it requires a sufficient number of subjects and dedicated equipment, and is not as simple as requiring significant costs and time.

  Therefore, a technique for efficiently estimating subjective quality based on a physical quantity of an audio signal is desired instead of human subjective evaluation. Such a technique is called objective quality evaluation. Currently, noise removal is one of the most widely used feature quantities as an objective index of noise suppression processing performance, but it is not always sufficient in terms of correspondence with subjective quality. This is because distortion occurs in the speech and noise during the noise suppression process, and this affects the subjective quality. Therefore, in order to estimate the subjective quality appropriately, it is necessary to consider such distortion. It is.

  Non-Patent Document 1 discloses PESQ (Perceptual evaluation of speech quality) as an objective quality evaluation technique capable of evaluating speech distortion. PESQ is a technique for measuring the quality of an evaluation target from the difference between the original speech signal and a degraded speech signal processed by an encoding method or apparatus to be evaluated. 8 and 9 are block diagrams showing a configuration example of a quality evaluation system using PESQ. FIG. 8 shows a configuration for evaluating the quality of an audio signal on which noise is not superimposed, and FIG. 9 shows a configuration for evaluating the quality of an audio signal on which noise is superimposed. 8 and 9, reference numeral 100 denotes an evaluation target apparatus, 101 denotes a PESQ apparatus, and 102 denotes an audio adder.

ITU-T Recommendation P.862, “Perceptual evaluation of speech quality (PESQ), an obective method for end-to-end speech quality assessment of narrow-band telephone networks and speech codecs”, Feb.2001

  As shown in FIG. 9, in PESQ, since the speech signal before noise is superimposed and the noise-suppressed speech signal are input, it is possible to evaluate the quality of the speech with superimposed noise in consideration of speech distortion. However, in PESQ, since there is no input related to noise before suppression processing, it is not possible to perform evaluation in consideration of noise distortion. Therefore, the PESQ disclosed in Non-Patent Document 1 has a problem that noise-suppressed speech cannot be accurately evaluated.

  The present invention has been made to solve the above-described problems, and an object of the present invention is to provide a noise-suppressed speech quality estimation apparatus, method, and program capable of accurately evaluating a noise-suppressed speech.

The present invention provides a noise-suppressed speech quality estimation device that objectively estimates the quality of noise-suppressed speech, and when the noise-superimposed speech signal is given as an input to a noise suppression processing device to be evaluated, the noise suppression Detection means for detecting a feature amount of a quality factor of a noise-suppressed speech signal output from the processing device, and estimation means for estimating the quality of the noise-suppressed speech signal based on the feature amount detected by the detection means The detecting means determines whether each section when the noise-suppressed speech signal is divided at a predetermined time is a speech section or a non-speech section; and a feature quantity of the quality factor of the noise-suppressed speech signal in the speech section Voice section feature quantity detecting means for detecting the noise section feature quantity detecting means for detecting the feature quantity of the quality factor of the noise-suppressed speech signal in the no voice section, and the voice section feature quantity The output means detects speech distortion as a feature quantity of the quality factor of the noise-suppressed speech signal by comparing the speech signal before noise is superimposed with the noise-suppressed speech signal in the speech section corresponding to the speech signal. A speech distortion measuring unit; and a volume measuring unit that detects a volume of the noise-suppressed speech signal in the speech segment as a feature amount of the quality factor of the noise-suppressed speech signal, By comparing the noise-suppressed speech signal in the non-speech interval with the corresponding noise-superimposed speech signal or the noise signal that is the basis of this noise-superposed speech signal, noise is used as a feature quantity of the quality factor of the noise-suppressed speech signal. A noise distortion measuring unit for detecting distortion; and a noise amount detecting unit for detecting a noise amount of the noise-suppressed speech signal in the non-speech interval as a feature factor of the quality factor of the noise-suppressed speech signal. The noise distortion is obtained by referring to the noise-suppressed speech signal in the no-speech interval as a degraded speech signal, the noise superimposed speech signal corresponding thereto, or the noise signal that is the basis of this noise superimposed speech signal. A PESQ value or Wideband-PESQ value when a signal is used, and the estimating means is configured to determine the noise based on the voice distortion, the volume of the noise-suppressed voice signal, the noise distortion, and the noise amount of the noise-suppressed voice signal. The quality of the suppressed speech signal is estimated .

In addition, one configuration example of the noise-suppressed speech quality estimation apparatus of the present invention further includes a speech database in which speech signals are registered in advance, speech playback means for playing back the speech signals in the speech database in an actual call environment, Voice recording means for outputting a signal obtained when the reproduced voice is collected as the noise-superimposed voice signal.
In addition, one configuration example of the noise-suppressed speech quality estimation apparatus of the present invention further includes a speech database in which speech signals are registered in advance, speech recording means for collecting noise signals in an actual call environment, and the speech database. And a sound adding means for outputting a signal obtained by adding the sound signal and the noise signal collected by the sound recording means as the noise superimposed sound signal.
Further, one configuration example of the noise-suppressed speech quality estimation apparatus of the present invention further includes a speech database in which speech signals are registered in advance, a noise database in which noise signals are registered in advance, a speech signal in the speech database, and the noise And a voice adding means for outputting a signal obtained by adding the noise signal of the database as the noise superimposed voice signal.

The noise-suppressed speech quality estimation method of the present invention provides a quality of a noise-suppressed speech signal output from the noise suppression processing device when a noise superimposed speech signal is given as an input to the noise suppression processing device to be evaluated. A detection procedure for detecting a feature quantity of the factor, and an estimation procedure for estimating the quality of the noise-suppressed speech signal based on the feature quantity detected by the detection procedure, wherein the detection procedure includes the noise-suppressed speech signal. A discriminating procedure for discriminating whether each segment is a speech segment or a non-speech segment when divided by a certain time, a speech segment feature detection procedure for detecting a feature factor of the quality factor of the noise-suppressed speech signal in a speech segment, A non-voice section feature quantity detection procedure for detecting a feature quantity of a quality factor of the noise-suppressed voice signal in a voice section, and the voice section feature quantity detection procedure is performed before speech is superimposed with noise. A speech distortion measurement procedure for detecting speech distortion as a feature quantity of the quality factor of the noise-suppressed speech signal by comparing the signal and a noise-suppressed speech signal in the speech section corresponding thereto, and the noise-suppressed speech signal A volume measurement procedure for detecting a volume of a noise-suppressed speech signal in the speech segment as a feature factor of the quality factor, and the speechless feature feature detection procedure corresponds to a noise-suppressed speech signal in the speechless segment A noise distortion measurement procedure for detecting noise distortion as a feature quantity of a quality factor of the noise-suppressed voice signal by comparing the noise-superimposed voice signal or a noise signal that is a source of the noise-superimposed voice signal, and the noise suppression A noise amount measurement procedure for detecting a noise amount of a noise-suppressed speech signal in the silent period as a feature amount of a quality factor of the speech signal, The sound distortion is a PESQ value or Wideband when the noise-suppressed speech signal in the non-speech interval is a degraded speech signal, and the corresponding noise superimposed speech signal or the noise signal that is the basis of this noise superimposed speech signal is a reference signal. A PESQ value, and the estimation procedure estimates the quality of the noise-suppressed speech signal based on the speech distortion, the volume of the noise-suppressed speech signal, the noise distortion, and the amount of noise of the noise-suppressed speech signal. It is characterized by .

The noise-suppressed speech quality estimation program of the present invention provides a quality of a noise-suppressed speech signal output from the noise suppression processing device when a noise-superimposed speech signal is given as an input to the noise suppression processing device to be evaluated. A detection procedure for detecting a feature quantity of the factor and an estimation procedure for estimating the quality of the noise-suppressed speech signal based on the feature quantity detected by the detection procedure are executed by the computer, and the detection procedure includes the noise suppression. A determination procedure for determining whether each section is a speech section or a non-speech section when the speech signal is divided at a predetermined time, and a speech section feature amount detection procedure for detecting a feature amount of the quality factor of the noise-suppressed speech signal in the speech section And a voiceless section feature quantity detection procedure for detecting a feature quantity of a quality factor of the noise-suppressed voice signal in the voiceless section, and the voice section feature quantity detection procedure includes: A speech distortion measurement procedure for detecting speech distortion as a feature quantity of the quality factor of the noise-suppressed speech signal by comparing the speech signal before the sound is superimposed and the noise-suppressed speech signal in the speech section corresponding to the speech signal. And a volume measurement procedure for detecting the volume of the noise-suppressed speech signal in the speech section as a feature quantity of the quality factor of the noise-suppressed speech signal, and the silent section feature amount detection procedure includes noise in the silent section. Noise that detects noise distortion as a feature quantity of the quality factor of the noise-suppressed speech signal by comparing the suppressed speech signal with the noise-superimposed speech signal corresponding thereto or the noise signal that is the basis of the noise-superposed speech signal A noise amount for detecting a noise amount of the noise-suppressed speech signal in the silent period as a distortion measurement procedure and a feature amount of the quality factor of the noise-suppressed speech signal The noise distortion includes a noise-suppressed speech signal in the no-speech interval as a degraded speech signal, a corresponding noise superimposed speech signal or a noise signal that is a source of this noise superimposed speech signal as a reference signal. PESQ value or Wideband-PESQ value when the noise suppression speech is calculated based on the speech distortion, the volume of the noise-suppressed speech signal, the noise distortion, and the noise amount of the noise-suppressed speech signal. The signal quality is estimated .

According to the present invention, when a noise-superimposed speech signal is given as an input to the noise suppression processing device to be evaluated and the feature quantity of the quality factor of the noise-suppressed speech signal output from the noise suppression processing device is detected, noise is detected. By comparing the noise signal that is the basis of the superimposed speech signal or the noise superimposed speech signal with the noise-suppressed speech signal, at least noise distortion is detected as a feature amount of the quality factor of the noise-suppressed speech signal, and based on this feature amount Estimating the quality of noise-suppressed speech signals makes it possible to estimate the quality of noise-suppressed speech in accordance with the user experience, making it possible to make appropriate settings for noise suppression processing technology and compare performance at low cost and easily. . For example, it is possible to know the performance of the noise suppression processing technique in each environment by providing a voice signal and a noise signal according to the environment where the voice call is performed as inputs. In addition, it is possible for a company developing noise suppression processing technology to know the performance of the developed technology. Further, when a voice communication terminal is designed, when the noise suppression processing technique is incorporated in the terminal, it is possible to select and set the best technique according to the assumed use environment.
In the present invention, when measuring the amount of noise of a noise-suppressed speech signal, sudden noise is captured by measuring the volume at regular intervals. As a result, it is possible to measure the noise amount of the noise-suppressed speech signal in consideration of the effect of sudden noise on the quality of experience.
In the present invention, when measuring the voice distortion of the noise-suppressed voice signal, the noise level is calculated based on the noise volume from the magnitude of the distortion detected by comparing the noise-suppressed voice signal and the voice signal that is the source of the noise-superimposed voice signal. To eliminate the effects of noise distortion. This makes it possible to measure pure speech distortion of the noise-suppressed speech signal.

  Further, the present invention reproduces a voice signal in a voice database under a real call environment, and outputs a signal obtained when the reproduced voice is collected as a noise superimposed voice signal, whereby noise in the real call environment is obtained. It is possible to accurately and easily perform quality estimation in accordance with the user experience of the suppressed speech.

  In the present invention, a noise signal is collected in a real call environment, and a signal obtained by adding the voice signal of the voice database and the collected noise signal is output as a noise superimposed voice signal. It is possible to accurately and easily perform quality estimation in accordance with the user experience of noise-suppressed speech.

  Further, in the present invention, a signal obtained by adding the voice signal of the voice database and the noise signal of the noise database is output as a noise-superimposed voice signal, so that the quality estimation according to the user experience of the noise-suppressed voice under various noise environments. Can be performed accurately and easily.

[First Embodiment]
Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a configuration example of a noise-suppressed speech quality estimation apparatus according to the first embodiment of the present invention.
As shown in FIG. 1, a noise-suppressed speech quality estimation apparatus 1 includes a speech database unit 2, a speech playback unit 3, a speaker 4, a speech recording unit 5, a microphone 6, a speech segment detection unit 7 (discriminating means), and a delay correction unit. 8-1, 8-2-1, 8-2-2, switch control unit 9, audio connection unit 10 (10-1 to 10-4), audio distortion measurement unit 11, noise distortion measurement unit 12, volume measurement unit 13, a noise amount measurement unit 14, a voice quality estimation unit 15 (estimation means), a voice quality output unit 16, and switches 20, 21 and 22.

  Voice section detection unit 7, delay correction unit 8-1, 8-2-1, 8-2-2, switch control unit 9, voice connection unit 10, voice distortion measurement unit 11, noise distortion measurement unit 12, and volume measurement unit 13, the noise amount measurement unit 14, and the switches 20 to 22 constitute detection means for detecting the feature amount of the quality factor of the noise-suppressed speech signal.

The audio reproducing unit 3 acquires the audio signal A from the audio database unit 2 and outputs the audio signal A from the speaker 4 in an environment where a voice call is performed.
The microphone 6 collects sound output from the speaker 4. At this time, since noise is superimposed on the sound collected by the microphone 6, the sound signal output from the microphone 6 is referred to as a noise superimposed sound signal B. The voice recording unit 5 acquires the noise superimposed voice signal B collected by the microphone 6 and inputs it to the evaluation target device 100 (noise suppression processing device) and the delay correction unit 8-2-2. By inputting the noise-superimposed speech signal B to the evaluation target device 100, let C be the speech signal that has been subjected to noise suppression processing and is output from the evaluation target device 100. An example of the evaluation target device 100 is a voice communication terminal, for example.

  The delay correction unit 8-1 receives the audio signal A and the noise-suppressed audio signal C, and measures the delay time of the noise-suppressed audio signal C with respect to the audio signal A. The delay correcting unit 8-1 measures the delay time of the noise-suppressed speech signal C by obtaining the time when the short-time cross-correlation coefficient between the speech signal A and the noise-suppressed speech signal C is maximized. The delay correction unit 8-2-1 delays the audio signal A by the delay time measured by the delay correction unit 8-1, thereby outputting the audio signal A ′ whose time is synchronized with the noise suppression audio signal C. To do. The delay correcting unit 8-2-2 delays the noise superimposed speech signal B by the delay time measured by the delay correcting unit 8-1, so that the noise superimposed speech signal in which the time is synchronized with the noise suppression speech signal C. B 'is output.

  The voice section detection unit 7 divides the voice signal A ′ into fixed short sections (20 ms), and uses VAD (Voice Activity Detection) to determine whether each section is a voice section where voice is present or a voiceless section. Determine for each. The voice section detector 7 transmits the type information (voice section or silent section) of each section of the voice signal A ′ obtained in this way to the switch controller 9.

The switch control unit 9 controls the switches 20 to 22 based on the type of each section of the audio signal A ′ notified from the audio section detection unit 7.
FIG. 2 is a diagram illustrating switch control when the short interval x1 of the audio signal A ′ is determined as the audio interval. When the type information indicating that the short interval x1 of the audio signal A ′ is an audio interval is input, the switch control unit 9 switches the switch 20 so that the short interval x1 of the audio signal A ′ is the audio connection unit 10−. At the same time, the switch 21 is switched so that the short section of the noise-suppressed speech signal C is input to the speech linking unit 10-3. Since the noise-suppressed audio signal C is synchronized with the audio signal A ′, the short interval of the noise-suppressed audio signal C corresponding to the short interval x1 of the audio signal A ′ is input to the audio connecting unit 10-3. . Further, the switch control unit 9 controls the switch 22 so that no signal is input to the voice connection unit 10-2.

  FIG. 3 is a diagram illustrating switch control when the short section x2 of the audio signal A ′ is determined to be a non-voice section. When the type information indicating that the short interval x2 of the audio signal A ′ is a non-audio interval is input, the switch control unit 9 switches the switch 21 so that the short interval of the noise-suppressed audio signal C is the audio connection unit 10. At the same time, the switch 22 is switched so that the short section of the noise superimposed audio signal B ′ is input to the audio connecting unit 10-2. Since the noise superimposed audio signal B ′ and the noise-suppressed audio signal C are synchronized with the audio signal A ′, the short interval of the noise superimposed audio signal B ′ corresponding to the short interval x2 of the audio signal A ′ 2 and the short section of the noise-suppressed speech signal C corresponding to the short section x2 is input to the speech linking unit 10-4. Further, the switch control unit 9 controls the switch 20 so that no signal is input to the voice connection unit 10-1.

  The voice linking unit 10-1 stores the first input signal of the short section, and thereafter, every time the short section signal is input, the latest signal that currently stores the input short section signal. The connected signal is newly stored after the connection. The voice connection unit 10-1 stores and stores the short sections of all input voice signals A 'as described above. The voice coupling units 10-2, 10-3, and 10-4 concatenately store short-period signals that are similarly input.

  As a result, the voice connection unit 10-1 stores the voice signal a obtained by connecting all the short voice sections of the voice signal A ′, and the voice connection section 10-2 stores all the voiceless short sections of the voice signal A ′. The noise signal b obtained by concatenating the short sections of the noise-superimposed speech signal B ′ corresponding to is stored, and the speech concatenation unit 10-3 stores the short of the noise-suppressed speech signal C corresponding to all the speech short sections of the speech signal A ′. The voice signal c1 in which the sections are connected is stored, and the noise signal c2 in which the short sections of the noise-suppressed speech signal C corresponding to all the voiceless short sections of the voice signal A ′ are stored in the voice connecting unit 10-4. The FIG. 4 shows the relationship among the audio signal A ′, the noise superimposed audio signal B ′, and the noise-suppressed audio signal C, and the audio signals a and c1 and the noise signals b and c2. In FIG. 4, the vertical axis represents signal level and the horizontal axis represents time.

Audio signals a and c1 are input to the audio distortion measurement unit 11. The audio distortion measurement unit 11 measures the audio distortion of the noise-suppressed audio signal C by comparing the audio signals a and c1. In this embodiment, a known PESQ is used for voice comparison, and voice distortion is measured using the voice signal a as a reference signal and the voice signal c1 as a degraded voice signal. PESQ outputs a distortion feature value as a PESQ value. The audio distortion measurement unit 11 transmits the measured PESQ value as the audio distortion amount x ₁ to the audio quality estimation unit 15.

Noise signals b and c2 are input to the noise distortion measurement unit 12. The noise distortion measurement unit 12 measures the noise distortion of the noise-suppressed speech signal C by comparing the noise signals b and c2. In the present embodiment, as in the case of measuring the audio distortion, a known PESQ is used for audio comparison, and the noise signal b is used as a reference signal and the noise signal c2 is used as a deteriorated audio signal. The noise distortion measurement unit 12 transmits the measured PESQ value as the noise distortion amount x ₂ to the voice quality estimation unit 15.

A sound signal c <b> 1 is input to the volume measuring unit 13. The volume measuring unit 13 measures the volume of the noise-suppressed audio signal C by measuring the volume of the audio signal c1. In this embodiment, a method standardized by ISO 532 is used for measuring the volume. The volume measuring unit 13 transmits the value of the measured volume x ₃ to the voice quality estimating unit 15.

A noise signal c <b> 2 is input to the noise amount measurement unit 14. The noise amount measurement unit 14 measures the noise amount of the noise-suppressed speech signal C by measuring the volume of the noise signal c2. In the present embodiment, a method standardized by ISO 532 is used for measurement, as in the case of measuring the volume. The noise amount measurement unit 14 transmits the value of the measured noise amount x ₄ to the voice quality estimation unit 15.

  The voice quality estimation unit 15 includes a voice distortion amount, a noise distortion amount, a volume, and a noise of the noise suppression voice signal C input from the voice distortion measurement unit 11, the noise distortion measurement unit 12, the volume measurement unit 13, and the noise amount measurement unit 14. Based on the amount, the subjective quality of the noise-suppressed speech signal C is estimated, and the estimated value of the subjective quality is transmitted to the speech quality output unit 16. The voice quality estimation unit 15 can estimate the subjective quality using, for example, an estimation formula obtained by the following method.

  First, in order to obtain an estimation formula, voice samples in which various feature quantities are given to voice distortion, noise distortion, volume, and noise amount are prepared in advance. Subjective quality assessment based on category scale. The average of the evaluation values obtained by this subjective quality evaluation is called a MOS (Mean Opinion Score) value. In the MOS value, 5 points are very good and 1 point is very bad.

Then, based on the MOS value for each audio sample, by using multiple regression analysis to obtain an expression for estimating subjective quality using the variables of the four quality factors of audio distortion, noise distortion, volume, and noise amount as variables, The following equation (1) is derived.
y = α ₁ · x ₁ + α ₂ · x ₂ + α ₃ · x ₃ + α ₄ · x ₄ + α ₅ (1)
Here, x ₁ represents the amount of voice distortion, x ₂ represents the amount of noise distortion, x ₃ represents the volume, x ₄ represents the amount of noise, and y represents the MOS value (subjective quality estimate). α ₁ , α ₂ , α ₃ , α ₄ and α ₅ are constants. The voice quality estimation unit 15 obtains an estimated value of subjective quality using Expression (1).

  The speech quality output unit 16 outputs the subjective quality estimation value of the noise-suppressed speech signal C input from the speech quality estimation unit 15 as an output value of the noise-suppressed speech quality estimation device 1.

  As described above, in this embodiment, in order to solve the conventional problems, the noise before noise suppression processing is compared with the noise after processing to detect noise distortion of noise-suppressed speech. For this reason, in this embodiment, a noise-superimposed speech signal before the suppression process is used as information regarding the noise signal before the noise suppression process. Furthermore, in order to obtain noise after noise suppression processing, the noise-suppressed voice signal is divided into a voice section and a non-voice section. The noise distortion of the noise-suppressed speech can be accurately detected from the difference between the noise before and after the noise suppression processing in the no-speech section. Further, in the present embodiment, the amount of noise is detected by measuring the volume of the noise-suppressed speech in the silent period. Furthermore, in the present embodiment, the sound distortion is detected from the difference between the sound and the noise-suppressed sound in the sound section of the noise-suppressed sound signal, and the sound volume is detected by measuring the sound volume of the noise-suppressed sound in the sound section.

  The user experience quality of the noise-suppressed speech is estimated from the speech distortion, noise distortion, volume, and noise amount that are the quality factors of the noise-suppressed speech detected in this way. In the present embodiment, an estimation equation obtained in advance is used to estimate the quality of noise-suppressed speech. This estimation formula prepares noise-reduced speech with various features for each quality factor, obtains each subjective quality assessment value through subjective quality assessment experiments, and obtains the subjective quality assessment value and the characteristics of the quality factor It is derived from the relationship of quantity.

  Thus, according to the present embodiment, it is possible to easily perform quality evaluation of noise-suppressed speech in consideration of noise distortion, which is impossible with conventional objective quality evaluation. Thereby, in this Embodiment, estimation close | similar to a user experience quality can be performed rather than a prior art. In the present embodiment, the voice signal of the voice database unit 2 is reproduced in an actual call environment, and a signal obtained when the reproduced voice is collected is used as a noise-superimposed voice signal. Using the noise superimposed speech signal generated below, it is possible to accurately and easily estimate the quality of the noise-suppressed speech in accordance with the user experience.

[Second Embodiment]
Hereinafter, a second embodiment of the present invention will be described with reference to the drawings. FIG. 5 is a block diagram showing a configuration example of a noise-suppressed speech quality estimation apparatus according to the second embodiment of the present invention, and the same components as those in FIG.
Also in the present embodiment, the configuration of the noise-suppressed speech quality estimation apparatus is almost the same as that of the first embodiment, so only the parts that are different from the first embodiment will be described.

First, in the first embodiment, when the noise amount measurement unit 14 measures the volume of the noise signal c2, the method standardized by ISO 532 is used. However, the method standardized by ISO 532 is implemented. In doing so, the following method can also be used. The noise amount measurement unit 14 substitutes the sound volumes B ₁ , B ₂ ,..., B _n measured every predetermined time t (ms) into the expression (2), so that sudden noise gives the quality of experience. The volume x ₅ [dB] of the noise c2 considering the influence is calculated. In this embodiment, t = 120 and p = 4. However, the present invention is not limited to these values. The noise amount measurement unit 14 transmits the measured volume x ₅ to the voice quality estimation unit 15 as the noise amount of the noise c2. In addition to the noise amount x ₅ , the noise amount measurement unit 14 calculates the volume x ₄ [dB] of the noise c 2 using a method standardized by ISO 532. The noise amount measurement unit 14 transmits the measured volume x ₄ to the audio distortion measurement unit 11.

Next, in the first embodiment, PESQ is used for audio comparison in the audio distortion measurement unit 11, but Wideband-PESQ may be used instead of PESQ. Wideband-PESQ is a technology in which the target range of known PESQ is expanded from a telephone band to a wide band, and it is possible to evaluate voice distortion considering up to a 7 kHz band. In the case of the present embodiment, the sound distortion measuring unit 11 receives the sound volume x ₃ from the sound volume measuring unit 13 and the sound volume x ₄ from the noise amount measuring unit 14 in addition to the sounds a and c1.

The voice distortion measuring unit 11 obtains the voice distortion of the noise-suppressed voice C by comparing the voice a and the voice c1. In the present embodiment, Wideband-PESQ is used, and W-PESQ value x ₆ ′ is calculated using speech a as a reference signal and speech c1 as degraded speech. However, since Wideband-PESQ also captures superimposed noise as distortion, the W-PESQ value x ₆ ′ is corrected based on the sound volume x ₃ in the voice section and the noise volume x ₄ in the non-voice section. . By substituting the W-PESQ value x ₆ ′ and the sound volumes x ₃ and x ₄ into Equation (3), a value x ₆ obtained by correcting the W-PESQ value x ₆ ′ is obtained.

Equation (3) means that the smaller one of α ₁ and x ₆ ′ / (1−x ₆ ′ / α ₂ ^{α3 (x3−x4)} ) is set as the correction value x ₆ . Here, α ₁ , α ₂ , and α ₃ are constants. In this embodiment, α ₁ = 4.644, α ₂ = 3, and α ₃ = 0.07, but the present invention is not limited to these values. The audio distortion measurement unit 11 transmits the measured value x ₆ to the audio quality estimation unit 15 as the audio distortion amount of the audio c1.

Next, in the first embodiment, PESQ is used for voice comparison in the noise distortion measurement unit 12, but Wideband-PESQ may be used instead of PESQ. Noises b and c2 are input to the noise distortion measurement unit 12. The noise distortion measurement unit 12 obtains the noise distortion of the noise-suppressed speech C by comparing the noise b and the noise c2. In the present embodiment, Wideband-PESQ is used, and W-PESQ value x ₇ is calculated using noise b as a reference signal and noise c2 as degraded speech. The noise distortion measurement unit 12 transmits the calculated W-PESQ value x ₇ to the voice quality estimation unit 15 as the noise distortion amount of the voice c1.

When Wideband-PESQ is used instead of PESQ for voice comparison, the voice quality estimation unit 15 uses the input volume x ₃ , noise amount x ₅ , voice distortion amount x ₆ , and noise distortion amount x ₇ as follows. The subjective quality of the noise-suppressed speech signal C is estimated. The voice quality estimation unit 15 calculates the value Q by substituting the volume x ₃ , the noise amount x ₅ , the voice distortion amount x ₆ , and the noise distortion amount x ₇ into Equation (4).

However, in Expression (4), x ₃ −x ₅ ≧ 10 is a constraint condition. Here, β _{1 to} β ₁₁ are constants. In this embodiment, β ₁ = 3.7, β ₂ = 0.215, β ₃ = 0.4, β ₄ = 3, β ₅ = 1.1, β ₆ = 0.9, β ₇ = 0. .05, β ₈ = 0.2, β ₉ = 4, β ₁₀ = 0.0002, and β ₁₁ = 24, but are not limited to these values. The voice quality estimation unit 15 transmits the calculated value Q to the voice quality output function 16 as an estimated value of the subjective quality of the noise-suppressed voice signal C.
Other configurations are the same as those of the first embodiment. Thus, also in this embodiment, the same effect as that of the first embodiment can be obtained.

[Third Embodiment]
The third embodiment of the present invention will be described below with reference to the drawings. FIG. 6 is a block diagram showing a configuration example of a noise-suppressed speech quality estimation apparatus according to the third embodiment of the present invention, and the same components as those in FIG.
As shown in FIG. 6, the noise-suppressed speech quality estimation device 17 includes a speech database unit 2, a speech recording unit 5, a microphone 6, a speech segment detection unit 7, and delay correction units 8-1, 8-2-1, 8-. 2-2, switch control unit 9, voice linking unit 10 (10-1 to 10-4), voice distortion measurement unit 11, noise distortion measurement unit 12, volume measurement unit 13, noise amount measurement unit 14, voice quality estimation unit 15, a voice quality output unit 16, a voice addition unit 18, and switches 20 to 22 are provided.

In the present embodiment, the real environment noise acquired by the microphone 6 and acquired by the voice recording unit 5 in the call environment is referred to as a noise signal B. The voice recording unit 5 inputs the noise signal B to the voice addition unit 18 and the delay correction unit 8-2-2.
The audio adder 18 receives the audio signal A and the noise signal B acquired from the audio database unit 2 and adds the audio signal A and the noise signal B to generate a noise superimposed audio signal. By inputting the noise-superimposed speech signal to the evaluation target device 100, the speech signal subjected to noise suppression processing output from the evaluation target device 100 is set as C.

Similarly to the first embodiment, the delay correction unit 8-1 receives the audio signal A and the noise-suppressed audio signal C, and measures the delay time of the noise-suppressed audio signal C with respect to the audio signal A. The delay correction unit 8-2-1 delays the audio signal A by the delay time measured by the delay correction unit 8-1, thereby outputting the audio signal A ′ whose time is synchronized with the noise suppression audio signal C. To do. The delay correcting unit 8-2-2 gives the noise signal B a delay corresponding to the delay time measured by the delay correcting unit 8-1, thereby outputting a noise signal B ′ whose time is synchronized with the noise-suppressed voice signal C. To do.
As in the first embodiment, the voice section detection unit 7 divides the voice signal A ′ into short sections and transmits type information of each section to the switch control unit 9.

  The switch control unit 9 and the voice connection unit 10 perform the same processing as in the first embodiment. However, the signal input to the audio connecting unit 10-2 via the switch 22 is not the noise superimposed audio signal but the noise signal B '. As a result, the noise signal b obtained by connecting the short sections of the noise signal B ′ corresponding to all the non-voice short sections of the voice signal A ′ is stored in the voice connecting unit 10-2. Other than that, as in the first embodiment, the audio connection unit 10-1 stores the audio signal a obtained by connecting all the audio short sections of the audio signal A ′, and the audio connection unit 10-3 stores the audio signal A ′. The speech signal c1 obtained by concatenating the short sections of the noise-suppressed speech signal C corresponding to all the speech short sections is stored, and the speech concatenation unit 10-4 stores noise corresponding to all the speechless sections of the speech signal A ′. A noise signal c2 obtained by connecting short sections of the suppressed speech signal C is stored.

The audio distortion measurement unit 11, the noise distortion measurement unit 12, the volume measurement unit 13, the noise amount measurement unit 14, the audio quality estimation unit 15, and the audio quality output unit 16 perform the same processing as in the first embodiment.
With the above configuration, the present embodiment can obtain the same effects as those of the first embodiment.

[Fourth Embodiment]
The fourth embodiment of the present invention will be described below with reference to the drawings. FIG. 7 is a block diagram showing a configuration example of a noise-suppressed speech quality estimation apparatus according to the fourth embodiment of the present invention. The same reference numerals are given to the same configurations as those in FIG.
The noise-suppressed speech quality estimation device 19 of the present embodiment uses a noise database unit 23 instead of the speech recording unit 5 and the microphone 6 in the noise-suppressed speech quality estimation device 17 of the third embodiment. .

In the present embodiment, it is assumed that a noise signal registered in advance in the noise database unit 23 is B. This noise signal B is input to the voice adder 18 and the delay corrector 8-2-2. Subsequent operations are the same as those in the third embodiment.
With the above configuration, the present embodiment can obtain the same effects as those of the first embodiment. Furthermore, in the present embodiment, since a signal registered in advance in the noise database unit 23 is used as the noise signal B, the quality estimation according to the user experience of the noise-suppressed speech under various noise environments is performed accurately and easily. It becomes possible.

Needless to say, the third and fourth embodiments may be applied to the second embodiment.
The noise-suppressed speech quality estimation apparatus according to the first to fourth embodiments can be realized by a computer having a CPU, a storage device, and an external interface, and a program for controlling these hardware resources. In such a computer, the noise-suppressed speech quality estimation program for realizing the noise-suppressed speech quality estimation method of the present invention is recorded on a recording medium such as a flexible disk, a CD-ROM, a DVD-ROM, or a memory card. Provided in. The CPU writes the program read from the recording medium into the storage device, and executes the processes described in the first to fourth embodiments according to the program.

  The present invention can be applied to a voice quality evaluation technique.

It is a block diagram which shows the structural example of the noise suppression audio | voice quality estimation apparatus which concerns on the 1st Embodiment of this invention. It is a figure which shows switch control when the short area of an audio | voice signal is discriminate | determined as an audio | voice area in the 1st Embodiment of this invention. It is a figure which shows switch control when the short area of an audio | voice signal is discriminate | determined as a non-voice area in the 1st Embodiment of this invention. FIG. 4 is a waveform diagram showing the relationship between the audio signal, the noise superimposed audio signal, and the noise-suppressed audio signal in the first embodiment of the present invention, and the audio signal and the noise signal after connecting these signals in a short section. is there. It is a block diagram which shows the structural example of the noise suppression audio | voice quality estimation apparatus which concerns on the 2nd Embodiment of this invention. It is a block diagram which shows the structural example of the noise suppression speech quality estimation apparatus which concerns on the 3rd Embodiment of this invention. It is a block diagram which shows the structural example of the noise suppression speech quality estimation apparatus which concerns on the 4th Embodiment of this invention. It is a block diagram which shows the structural example of the quality evaluation system of a noiseless audio | voice signal using PESQ. It is a block diagram which shows the structural example of the quality evaluation system of a noise superimposed audio | voice signal using PESQ.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1,17,19 ... Noise suppression voice quality estimation apparatus, 2 ... Voice database part, 3 ... Voice reproduction part, 4 ... Speaker, 5 ... Voice recording part, 6 ... Microphone, 7 ... Voice section detection part, 8-1, 8-2-1, 8-2-2 ... delay correction unit, 9 ... switch control unit, 10 ... audio connection unit, 11 ... audio distortion measurement unit, 12 ... noise distortion measurement unit, 13 ... volume measurement unit, 14 ... Noise amount measurement unit, 15 ... voice quality estimation unit, 16 ... voice quality output unit, 18 ... voice addition unit, 20, 21, 22 ... switch, 23 ... noise database unit.

Claims (6)

A noise-suppressed speech quality estimation device that objectively estimates the quality of noise-suppressed speech,
Detecting means for detecting a feature quantity of a quality factor of a noise-suppressed speech signal output from the noise suppression processing device when a noise-superimposed speech signal is given as an input to the noise suppression processing device to be evaluated;
An estimation unit that estimates the quality of the noise-suppressed speech signal based on the feature amount detected by the detection unit;
The detection means includes
Discriminating means for discriminating whether each section when the noise-suppressed speech signal is divided at a certain time is a speech section or a silent section;
Speech section feature amount detecting means for detecting a feature amount of a quality factor of the noise-suppressed speech signal in a speech section;
A voiceless section feature quantity detecting means for detecting a feature quantity of a quality factor of the noise-suppressed voice signal in a voiceless section,
The voice section feature amount detection means includes:
A speech distortion measuring unit that detects speech distortion as a feature quantity of the quality factor of the noise-suppressed speech signal by comparing the speech signal before noise is superimposed with the noise-suppressed speech signal in the speech section corresponding to the speech signal When,
A volume measuring unit that detects the volume of the noise-suppressed speech signal in the speech section as a feature quantity of the quality factor of the noise-suppressed speech signal;
The silent section feature quantity detecting means is
By comparing the noise-suppressed speech signal in the no-speech section with the corresponding noise-superimposed speech signal or the noise signal that is the basis of this noise-superposed speech signal, the characteristic amount of the quality factor of the noise-suppressed speech signal A noise distortion measurement unit for detecting noise distortion;
A noise amount measuring unit that detects a noise amount of the noise-suppressed speech signal in the no-speech interval as a feature amount of the quality factor of the noise-suppressed speech signal;
The noise distortion is a PESQ value obtained when the noise-suppressed speech signal in the silent period is a degraded speech signal, the noise superimposed speech signal corresponding to the noise suppressed speech signal or the noise signal that is the basis of the noise superimposed speech signal is a reference signal, or Wideband-PESQ value,
The estimation means estimates the quality of the noise-suppressed speech signal based on the speech distortion, the volume of the noise-suppressed speech signal, the noise distortion, and the amount of noise of the noise-suppressed speech signal. Voice quality estimation device. The noise-suppressed speech quality estimation apparatus according to claim 1,
Furthermore, an audio database in which audio signals are registered in advance,
Voice playback means for playing back the voice signal of the voice database in an actual call environment;
An apparatus for estimating a noise-suppressed voice quality , comprising: voice recording means for outputting a signal obtained when the reproduced voice is collected as the noise-superimposed voice signal . The noise-suppressed speech quality estimation apparatus according to claim 1 ,
Furthermore, an audio database in which audio signals are registered in advance,
A voice recording means for collecting noise signals in an actual call environment;
A noise-suppressed speech quality estimation apparatus comprising speech adding means for outputting a signal obtained by adding the speech signal of the speech database and the noise signal collected by the speech recording means as the noise superimposed speech signal . The noise-suppressed speech quality estimation apparatus according to claim 1 ,
Furthermore, an audio database in which audio signals are registered in advance,
A noise database in which noise signals are registered in advance;
A noise-suppressed speech quality estimation apparatus comprising speech adding means for outputting a signal obtained by adding a speech signal of the speech database and a noise signal of the noise database as the noise superimposed speech signal . A noise suppression speech quality estimation method for objectively estimating the quality of noise suppression speech,
A detection procedure for detecting a feature quantity of a quality factor of a noise-suppressed speech signal output from the noise suppression processing device when a noise-superimposed speech signal is given as an input to the noise suppression processing device to be evaluated;
An estimation procedure for estimating the quality of the noise-suppressed speech signal based on the feature amount detected by the detection procedure,
The detection procedure includes:
A determination procedure for determining whether each section when the noise-suppressed speech signal is divided at a certain time is a speech section or a silent section;
A speech section feature amount detection procedure for detecting a feature amount of a quality factor of the noise-suppressed speech signal in a speech section;
A silent section feature amount detection procedure for detecting a feature amount of a quality factor of the noise-suppressed speech signal in a silent section,
The speech segment feature amount detection procedure includes:
A speech distortion measurement procedure for detecting speech distortion as a feature quantity of the quality factor of the noise-suppressed speech signal by comparing the speech signal before the noise is superimposed with the noise-suppressed speech signal corresponding to the speech section. When,
The volume measurement procedure for detecting the volume of the noise-suppressed speech signal in the speech section as a feature quantity of the quality factor of the noise-suppressed speech signal,
The silent section feature amount detection procedure includes:
By comparing the noise-suppressed speech signal in the no-speech section with the corresponding noise-superimposed speech signal or the noise signal that is the basis of this noise-superposed speech signal, the characteristic amount of the quality factor of the noise-suppressed speech signal Noise distortion measurement procedure for detecting noise distortion,
A noise amount measurement procedure for detecting a noise amount of the noise-suppressed speech signal in the no-speech interval as a feature amount of the quality factor of the noise-suppressed speech signal,
The noise distortion is a PESQ value obtained when the noise-suppressed speech signal in the silent period is a degraded speech signal, the noise superimposed speech signal corresponding to the noise suppressed speech signal or the noise signal that is the basis of the noise superimposed speech signal is a reference signal, or Wideband-PESQ value,
The estimation procedure estimates the quality of the noise-suppressed speech signal based on the speech distortion, the volume of the noise-suppressed speech signal, the noise distortion, and the noise amount of the noise-suppressed speech signal. Speech quality estimation method . A noise-suppressed speech quality estimation program that operates a computer as a noise-suppressed speech quality estimation device that objectively estimates the quality of noise-suppressed speech,
A detection procedure for detecting a feature quantity of a quality factor of a noise-suppressed speech signal output from the noise suppression processing device when a noise-superimposed speech signal is given as an input to the noise suppression processing device to be evaluated;
Causing the computer to execute an estimation procedure for estimating the quality of the noise-suppressed speech signal based on the feature amount detected by the detection procedure,
The detection procedure includes:
A determination procedure for determining whether each section when the noise-suppressed speech signal is divided at a certain time is a speech section or a silent section;
A speech section feature amount detection procedure for detecting a feature amount of a quality factor of the noise-suppressed speech signal in a speech section;
A silent section feature amount detection procedure for detecting a feature amount of a quality factor of the noise-suppressed speech signal in a silent section,
The speech segment feature amount detection procedure includes:
A speech distortion measurement procedure for detecting speech distortion as a feature quantity of the quality factor of the noise-suppressed speech signal by comparing the speech signal before the noise is superimposed with the noise-suppressed speech signal corresponding to the speech section. When,
The volume measurement procedure for detecting the volume of the noise-suppressed speech signal in the speech section as a feature quantity of the quality factor of the noise-suppressed speech signal,
The silent section feature amount detection procedure includes:
By comparing the noise-suppressed speech signal in the no-speech section with the corresponding noise-superimposed speech signal or the noise signal that is the basis of this noise-superposed speech signal, the characteristic amount of the quality factor of the noise-suppressed speech signal Noise distortion measurement procedure for detecting noise distortion,
A noise amount measurement procedure for detecting a noise amount of the noise-suppressed speech signal in the no-speech interval as a feature amount of the quality factor of the noise-suppressed speech signal,
The noise distortion is a PESQ value obtained when the noise-suppressed speech signal in the silent period is a degraded speech signal, the noise superimposed speech signal corresponding to the noise suppressed speech signal or the noise signal that is the basis of the noise superimposed speech signal is a reference signal, or Wideband-PESQ value,
The estimation procedure estimates the quality of the noise-suppressed speech signal based on the speech distortion, the volume of the noise-suppressed speech signal, the noise distortion, and the noise amount of the noise-suppressed speech signal. Voice quality estimation program.

Images