JPH05173592A - Method and device for voice/no-voice discrimination making - Google Patents

Abstract

PURPOSE:To provide the voice/no-voice discrimination device of simple constitution which makes a voice/no-voice decision automatically with high precision. CONSTITUTION:A feature extraction part 11 extracts plural feature quantities from an input signal at constant intervals of time and a threshold value determination part 12 predetermines a threshold value by using learning data on many speeches and nonspeeches; and a rough decision part 13 and a detail decision part 14 decide the speeches and others by comparing the extracted feature quantities with the threshold value and a final decision part decides whether or not the section is a voice or not according to the presence rate of the number of frames decided by the rough decision part and detail decision part as voices.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice / non-voice discriminating method for discriminating whether an input signal for switching a television camera used in a video conference system or a microphone is a voice or another sound. The present invention relates to a discriminating device, a voice / non-voice discriminating method and a discriminating device which are used in preprocessing of a voice recognition device and which judge whether an input signal is a voice or another sound.

[0002]

2. Description of the Related Art In a video camera used in a video conference system or a voice / non-voice discriminating device for switching microphones, there is a possibility of switching to inputs other than voice such as various noises existing in a conference room. There is. Further, in a device that performs voice processing such as voice recognition, if a signal other than voice is input and is erroneously determined to be voice, erroneous recognition occurs. Therefore, there is a need for a voice / non-voice discrimination device capable of accurately determining whether or not the input signal is voice.

In the conventional speech / non-speech discriminating apparatus, a method is generally used in which a portion where the power value of the input signal is larger than a predetermined threshold value is determined as speech for simplification of processing. However, in consideration of using it in a real environment such as a conference room, it is possible that various sounds with large power other than the sound of flipping through materials such as paper and noise such as noise caused by the vibration of the microphone such as breathing may be input. However, it is not possible to distinguish between voice and non-voice only by the power.

Therefore, some methods have been proposed for determining the voicedness of an input signal by using a plurality of voice feature amounts other than power. For example, A Pattern Recognition Approa
ch to Voiced-unvoiced-silence classification with
application to speech recognition, IEEE Trans. Aco
ust., Speech, Signal Processing, ASSP-24-3 (1976)
There is a method by. This is a method of collectively determining silence / unvoiced / voiced by using a pattern recognition method.
If the silent / unvoiced sound is considered as noise having many high frequency components, it can be considered as a kind of speech / non-speech discrimination apparatus based on detection of voiced sound (vowel sound). Specifically, five types of feature quantities of the number of zero crossings of the input speech at constant time intervals, the logarithmic energy of the signal, the first-order autocorrelation coefficient, the first-order linear prediction coefficient, and the logarithmic energy of the linear prediction residual are calculated. Then, a normal distribution is assumed for each feature amount, and silent / unvoiced / voiced determination is performed based on their joint probability.

[0005]

However, the above speech / non-speech discriminating apparatus can remove only the noise with the energy predominantly in the high frequency range. Further, the feature amount based on the feature of each phoneme in the voice is not used, and highly accurate voice / non-voice discrimination is required by using the feature amount suitable for detecting each phoneme of the voice. .. The present invention solves the above problems, and an object thereof is to provide a high-performance voice / non-voice discrimination device for voice discrimination. INDUSTRIAL APPLICABILITY The present invention uses a feature amount that is excellent in removing dominant noise of energy in a low frequency range, thereby causing noise caused by vibration of a microphone due to breathing or the like and low frequency noise caused by resonance with a floor or a desk. It is possible to remove the speech, and by using a combination of feature quantities suitable for detecting each phoneme of the voice, it is possible to perform a high performance based on the detection of each phoneme of the voice and to distinguish the voice / non-voice with a simple configuration. An object is to provide a possible voice / non-voice discrimination device.

[0006]

In order to achieve the above-mentioned object, the present invention aims to detect a voiced sound (vowel sound), and an autocorrelation coefficient of a first order or more for every constant time of an input signal,
It is extracted by a feature amount extraction unit that extracts a plurality of voice feature amounts such as first- and higher-order cepstrum coefficients, and the feature amount extraction unit that has extracted a large number of voice and non-voice learning data in advance within a predetermined time. A threshold value determining unit that determines a threshold value for determining whether it is voice or not using the feature amount, and a first-order or higher-order autocorrelation coefficient and a first-order cepstrum coefficient within a predetermined time. By comparing with the threshold value determined by the threshold value determination unit, it is determined whether it is voice or not,
A rough determination unit that roughly determines how vowel the voice has in the frequency domain, and a second or higher-order cepstrum coefficient that is a feature amount suitable for the detection of voiced sound based on the features of each phoneme in the voice. By comparing the threshold value determined by the threshold value determination unit to determine whether it is voice or otherwise, a detailed determination unit that determines in detail how much vowelness it has, the rough determination unit, and the The detailed determination unit includes a final determination unit that determines whether or not a section having a power level higher than a certain level is voice based on the existence ratio of frames determined to be voice.

[0007]

According to the present invention, with the above-described configuration, the feature amount that directly indicates in which frequency band the voiced sound such as a vowel has the most dominant energy, and the detection of the voiced sound based on the feature of each phoneme in the voice is detected. Using a feature amount suitable for the above, the voice is discriminated at regular intervals based on a preset threshold value based on a large number of highly reliable voice data and non-voice data including various noises. , High-performance voice /
Non-voice can be discriminated.

[0008]

EXAMPLE An example of the present invention will be described below.
FIG. 1 is a block diagram showing the overall configuration of an embodiment of the present invention. In FIG. 1, reference numeral 11 denotes a feature extraction unit that extracts a plurality of feature amounts for voice discrimination, and a power calculation unit 11a that calculates power for each frame and a primary autocorrelation coefficient for each frame. A first-order autocorrelation coefficient calculation unit 11b, a seventh-order autocorrelation coefficient calculation unit 11c that calculates a seventh-order autocorrelation coefficient for each frame, and a first-order cepstrum coefficient for each frame. And a third-order cepstrum coefficient calculation unit 11e that calculates a third-order cepstrum coefficient for each frame. These feature quantities are used to detect voiced sound (vowelness) of the input signal. The results of an investigation on the frequency distribution of these feature quantities are shown below.

The data used for the investigation are two types of acoustic data, voice data and noise data. As the voice data, one man recorded in an anechoic chamber was uttered 21
16 phonemes in 2 words (/ a /, / i /, / u /, / e /, / o /, / b /, / d /, / g
/, / m /, / n /, / N /, / s /, / h /, / r /, / w /, / y /) and the phoneme boundaries were determined by visual inspection for all phonemes. There is. Further, as the noise data, 22 kinds of noises as shown in (Table 1) that can be assumed in the conference room where the voice / non-voice discrimination apparatus of the present embodiment will be used are used. The analysis conditions for voice and noise data are shown in (Table 2).

[0010]

[Table 1]

[0011]

[Table 2]

As a result of the investigation, the frequency distributions for the first-order autocorrelation coefficient are 16 phonemes and 22 noises (FIG. 4) and FIG. The cases where the distributions are for 16 phonemes and 22 noises (FIG. 6) (FIG. 7) respectively, and the cases where the frequency distribution for the primary cepstrum coefficient are for 16 phonemes and 22 noises (FIG. 8) (FIG. 9) respectively The frequency distributions for the third-order cepstrum coefficient are shown in FIG. 10 and FIG. 11 for 16 phonemes and 22 noises, respectively. In each figure, the black circles represent the average value and the standard deviation in the vertical direction. As a result, the following tendencies were found.

First, the autocorrelation coefficient of the first or higher order is a feature quantity that reflects the difference in the concentrated frequency range of energy, and is used for noise with strong randomness such as unvoiced sound that exists predominantly in the frequency band where energy is high. The value of the first-order autocorrelation coefficient shows a small value close to 0 (FIG. 5), and the value of voiced sound shows close to 1 (FIG. 4). On the other hand, in the case of noise where energy is predominantly present in the low frequency band,
The value of the 7th-order autocorrelation coefficient is close to 1 (FIG. 7),
For voiced sounds, etc., its value is close to 0 (Fig. 6). The cepstrum coefficient is a feature quantity that represents the shape of the spectrum,
Even for the same voiced sound, the value varies greatly for each phoneme. The first-order cepstrum coefficient is an amount representing a rough difference in the shape of a spectrum such as voiced sound or unvoiced sound, and is a phonological unit / i /
The voiced sounds except for have a value of 1.0 or more, and the other sounds have a value of 1.0 or less (FIG. 8). The third-order cepstrum coefficient is a feature amount in which the characteristic of the phoneme / i / is particularly large, and the value is 0.5 or more at / i /, and 0.5 or less at other sounds (FIG. 10). ).

In FIG. 1, reference numeral 12 is a threshold value deciding unit for deciding an appropriate threshold value for judging whether it is voice or not, using a large number of voice and non-voice learning data in advance. .. Reference numeral 13 indicates a comparison between the first-order and seventh-order autocorrelation coefficients and the first-order cepstrum coefficient output from the feature extraction unit 11 with a certain appropriate threshold value determined by the threshold value determination unit 12 on a frame-by-frame basis. Reference numeral 14 denotes a general determination unit that determines whether the sound is a voice or not, and the reference numeral 14 indicates a feature extraction unit 11 among those determined to be other than the voice by the general determination unit 13.
The third-order cepstrum coefficient output from the above and a certain suitable threshold value determined by the threshold value determination unit 12 are compared in a frame unit to determine whether the voiced sound is / i / other than that. is there. Reference numeral 15 is a certain threshold value determined by the threshold value determining unit 12 as a ratio of the number of frames determined to be voices by the rough determining unit 13 and the detailed determining unit 14 for a group of input signals having a power level higher than a certain level. It is the final determination unit that determines the block as a voice in the above case.

An embodiment of the present invention will be described below (FIG. 1).
This will be described in detail with reference to the block diagram of FIG. 3 and a flowchart for explaining the operation of the general determination unit 13 (FIG. 2) and a flowchart for explaining the operation of the detailed determination unit 14 (FIG. 3).

When the acoustic signal is input through the microphone, the feature extraction unit 11 first extracts five feature amounts. In the power calculation unit 11a, the power value for each fixed time is calculated by, for example, (Equation 1). The fixed time interval is
Here, for example, the sampling frequency is 10 KHz, 200 points (20 ms), and this time unit is called a frame.

[0017]

[Equation 1]

Where Pi is the power value at frame i,
S _k represents a sample value of the input signal in the frame. For this power value, the maximum value and the minimum value in the high power section are normalized to, for example, a value of 0 to 1 so that the difference in power due to the difference in utterance conditions can be handled in a unified manner. 1
In the next autocorrelation coefficient calculation unit 11b, the 1st-order autocorrelation coefficient Ai (1) is calculated for each frame and the 7th-order autocorrelation coefficient calculation unit 11b
In c, the 7th-order autocorrelation coefficient Ai (7) is calculated by (Equation 2) and (Equation 3) for each frame, and Ai (1), Ai (7)
(7) Both are normalized by the zero-order autocorrelation coefficient.

[0019]

[Equation 2]

[0020]

[Equation 3]

In the primary cepstrum coefficient calculation unit 11d, the primary cepstrum coefficient Ci (1) in frame i is
In the third-order cepstrum coefficient calculation unit 11e, the frame i
The third-order cepstrum coefficient Ci (3) at is calculated by linear prediction analysis. Note that instead of the first-order autocorrelation coefficient, it is 1
Even if the next partial autocorrelation coefficient or the first-order linear prediction coefficient is used instead of the first-order cepstrum coefficient, the absolute values of these values are equal to each other. Also, instead of the 7th-order autocorrelation coefficient, even if the 6th-order to 12th-order autocorrelation coefficients are used, it is used to remove the non-speech predominantly present in the frequency band where the energy is low. It does not matter in the sense. Further, in the present embodiment, by using the first and third-order cepstrum coefficients, the feature amount in which the feature of / i / is particularly large is focused on by focusing on the phoneme / i / in the voice. Other phonemes such as / a /, / u /, / e /, /
You may combine and use the 1st or more-order cepstrum coefficient which features that o / etc. show greatly. Further, as the cepstrum coefficient, an LPC cepstrum coefficient, an FFT cepstrum coefficient, or a mel cepstrum coefficient may be used in the sense that the phonological property is determined in detail based on the characteristics of each phoneme in the voice.

The threshold value deciding unit 12 extracts feature values obtained by the feature extracting unit 11 in advance for vowel parts and non-voice data of a large number of voice data, and determines whether the voice is based on the distribution of these feature amounts. An appropriate threshold value other than the above is determined for each feature amount. Further, there is a certain appropriateness for judging voice / non-voice depending on the proportion of the feature amount obtained by the feature extraction unit 11 using the learning data of voice in a predetermined number of frames. Determine a threshold. As the non-voice data used when determining the voice / non-voice threshold value, for example, when the voice / non-voice discriminating apparatus of the present embodiment is used in a conference room or the like, a sound of tapping a desk or printing of paper can be performed. It is sufficient to use expected noise data such as sounds, cup sounds, and sounds of objects touching the microphone.

The feature amount obtained by the feature extraction unit 11 from the acoustic signal is input to the general determination unit 13 and the detailed determination unit 14, respectively. First, the first-order and seventh-order autocorrelation coefficients and the first-order cepstrum coefficients among the feature quantities obtained from the acoustic signal by the feature extraction unit 11 are input to the rough determination unit 13, respectively.

In step 21 shown in FIG. 2, 1
Depending on the value of the next autocorrelation coefficient, noise with strong randomness, such as unvoiced sound that exists predominantly in the frequency band with high energy, is removed. Let A1 be the threshold value of the first-order autocorrelation coefficient determined by the threshold value determination unit 12,
When i (1) ≧ A1, it is determined to be voice, and the others are non-voice. Next, in step 22, the noise that is predominantly present in the low energy frequency band is removed by the magnitude of the value of the seventh-order autocorrelation coefficient. That is, assuming that the threshold value of the 7th-order autocorrelation coefficient determined by the threshold value determination unit 12 is A7, it is determined that Ai (7) ≦ A7 is voice, and the others are nonvoice.

Further, in step 23, a voiced sound excluding the phoneme / i / is detected based on the magnitude of the value of the first-order cepstrum coefficient. 1 determined by the threshold determination unit 12
If the threshold value of the next cepstrum coefficient is C1, then Ci
(1) When ≧ C1, it is determined to be voice, and the others are determined to be non-voice. If it is voice, in step 24, Vi =
If the output value of 1 is non-voice, V is output in step 25.
The output value of i = 0 is sent to the detail determination unit 14.

Next, in step 31 of the detailed determination section 14, if the rough determination section 13 determines that the voice is present, that is, if Vi = 1, the output value is sent to the final determination section 15 as it is, and the rough determination section 13 is executed. In the case of non-speech, that is, only when Vi = 0
i / is detected. In step 32, assuming that the threshold value of the third-order cepstrum coefficient determined by the threshold value determination unit 12 is C3, comparison with the magnitude of the value of the third-order cepstrum coefficient extracted by the feature extraction unit 11 from the acoustic signal. Thus, it is determined that / i /, that is, voice only when Ci (3) ≧ C3, and non-voice other than that. If it is a voice, the output value of Vi = 1 is sent to the final determination unit 15 at step 33, and if it is a non-voice, the output value of Vi = 0 at step 34.

In the final determination section 15, first, the power calculation section 1
1a, the threshold value determining unit 12
The section which exceeds the predetermined power threshold value by a predetermined length or more is detected as a voice candidate section. Let C be the frame length of the voice candidate section at this time. For this voice candidate section, the number C1 of frames determined to be voice by the rough determination unit 13 and the detail determination unit 14 is counted, and the ratio of the number of frames of the section determined to be voice in the voice candidate section is calculated. When the threshold value determining unit 12 exceeds the predetermined threshold value M, that is, when the condition of (Equation 4) is satisfied, this voice candidate section is determined to be voice.

[0028]

[Equation 4]

As described above, according to the voice / non-voice discriminating apparatus of the present embodiment, the feature amount extracting section 11 for extracting a plurality of feature amounts of the voice at constant time intervals from the input signal, and a large number of voices in advance A threshold value determination unit 12 that determines a threshold value for determining whether the sound is voice or not using the feature quantity extracted for each frame of the learning data of the voice, and the feature quantity and the threshold value of a plurality of voices. By comparing with the threshold value determined by the determination unit, it is determined whether it is voice or not, and a rough determination unit 13 that roughly determines how vowel the voice has in the frequency domain; Based on the characteristics of each phoneme, a plurality of features suitable for voiced sound detection are compared with the threshold value determined by the threshold value determination unit to determine whether it is a voice or not, and how much vowelness it has. Detailed determination unit 14 for determining in detail A final determination unit that determines whether or not a section having a power level higher than a certain level by the rough determination unit and the detailed determination unit is a voice based on the existence ratio of frames determined to be voice. Due to
It is possible to provide a speech / non-speech discrimination device capable of accurately discriminating various acoustic signals with a simple configuration.

[0030]

As is apparent from the above description, according to the present invention, a plurality of feature quantities that characterize a voice are extracted, and an appropriate threshold value is set in advance from the feature quantities of many vowels and non-voice data. Every time, it is configured to judge whether it is voice or non-voice for each frame, and to judge whether it is voice or non-voice based on the existence ratio of the frames determined as the voice section candidates for the part with a large power. It is possible to provide a speech / non-speech discriminating apparatus capable of accurately discriminating whether an input signal is voice or not with such a configuration.

[Brief description of drawings]

FIG. 1 is a block diagram showing the overall configuration of a voice / non-voice discrimination device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing the operation of an embodiment of a general determination section of the present invention.

FIG. 3 is a flowchart showing the operation of an embodiment of a detail determination unit of the present invention.

FIG. 4 is a frequency distribution diagram of first-order autocorrelation coefficients in 16 phonemes.

FIG. 5 is a frequency distribution diagram of the first-order autocorrelation coefficient in 22 noise.

FIG. 6 is a frequency distribution diagram of the 7th-order autocorrelation coefficient in 16 phonemes.

FIG. 7 is a frequency distribution diagram of the 7th-order autocorrelation coefficient in 22 noise.

FIG. 8 is a frequency distribution diagram of first-order cepstrum coefficients in 16 phonemes.

FIG. 9 is a frequency distribution diagram of the first-order cepstrum coefficient in 22 noise.

FIG. 10 is a frequency distribution chart of the third-order cepstrum coefficient in 16 phonemes.

FIG. 11 is a frequency distribution diagram of the third-order cepstrum coefficient in 22 noise.

[Description of Reference Signs] 11 feature extraction unit 11a power calculation unit 11b first-order autocorrelation coefficient calculation unit 11c seventh-order autocorrelation coefficient calculation unit 11d first-order cepstrum coefficient calculation unit 11e third-order cepstrum coefficient calculation unit 12 Threshold value determination unit 13 General determination unit 14 Detailed determination unit 15 Final determination unit

Claims (3)

[Claims] 1. A feature amount of a plurality of voices is extracted from an input signal at regular intervals using at least one of a first-order autocorrelation coefficient and a second-order or higher autocorrelation coefficient that characterize the voice, A voice / non-voice discrimination method for discriminating between voice and non-voice. 2. A feature quantity of a plurality of voices is extracted from an input signal at regular time intervals by using at least one of a first-order cepstrum coefficient and a second-order or higher-order cepstrum coefficient that characterize the voice, and a voice or non-voice is extracted. A voice / non-voice discrimination method for determining whether or not 3. A feature extraction unit for extracting a first-order autocorrelation coefficient, a second-order or higher autocorrelation coefficient, a first-order cepstrum coefficient, and a second-order or higher-order cepstrum coefficient, which characterize speech at regular intervals, from an input signal. And threshold value determination for determining a threshold value for determining whether it is voice or not using the feature amount extracted by the feature extraction unit within a predetermined time with respect to a large number of voice and non-voice learning data in advance. And at least one of the first-order autocorrelation coefficient and the first-order cepstrum coefficient extracted by the feature extraction unit within a predetermined time from the input signal, the threshold determination unit determines. A rough judgment unit that judges whether it is voice or not by comparing it with a threshold value and a rough judgment unit that is judged to be other than voice are extracted by the feature extraction unit within a predetermined time from the input signal. At least one feature amount of the second or higher-order cepstrum coefficients that has been output is compared with the threshold value determined by the threshold value determination unit to determine whether it is voice or not, and a constant power. When the ratio of the number of frames determined to be speech by the rough determination unit and the detailed determination unit with respect to a block of the input signal having a level or higher is equal to or higher than the threshold value determined by the threshold value determination unit, the block is voiced. A voice / non-voice discriminating apparatus comprising: