JPH05249987A - Voice detecting method and device - Google Patents

Abstract

PURPOSE:To automatically detect voice with high precision with a comparatively simple structure in a voice detecting device for detecting only the voice, which is used as the pretreatment of a voice recognizing device. CONSTITUTION:A voice detecting device has a tolerance calculating part 13 for extracting a plurality of characteristic quantities from input signal every fixed time by a characteristic extracting part 11 and calculating the logarithmic tolerance with a vowel standard model formed by use of a number of learning data of vowel by a vowel standard model forming part 12: and a vowel judging part 14 for calculating a frame average logarithmic tolerance by collectively using the logarithmic tolerances for several frames and detecting vowels by comparison with a proper threshold. According to the number judged as vowels by the vowel judging part, it is judged by a final judging part 15 whether this section is a voice or not.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice detection method and a voice detection device for detecting only voice used in preprocessing of a voice recognition device in an actual environment where non-stationary noise exists.

[0002]

2. Description of the Related Art In a device for performing voice processing such as voice recognition,
If non-stationary noise other than voice is input and is mistakenly determined as voice, misrecognition occurs. Therefore, there is a need for a voice detection device that can accurately determine whether the input signal is voice.

In a conventional voice detection apparatus, a method is generally used in which a portion where the power value of an input signal is larger than a threshold value is determined to be voice for simplification of processing. However, considering that it is used in a real environment where voice recognition is performed, various sounds with large power other than voice, such as the sound of flipping over materials such as paper, noise caused by the vibration of a microphone such as a breath, or the sound of animals May be input, and speech cannot be detected only by power.

Therefore, some methods have been proposed for determining whether the input signal is voice or non-voice by using a plurality of voice feature amounts other than power. For example, there is a method according to "discrimination between voice / non-voice in an actual environment" (Akira Ishida and Hidefumi Obata, Journal of Acoustical Society of Japan, Volume 7, No. 12 (1991)). This is to distinguish voice / non-voice in a real environment by using acoustic features that are effective for distinguishing various non-stationary noises generated in everyday laboratories and offices from voice. There is. Specifically, in the high power portion of the voice, the voice / non-voice is discriminated according to how many parts can be regarded as vowels, and the acoustic feature quantity to be used is (a) periodicity (B) Pitch frequency (c) Optimal linear prediction order (d) Distance to five vowels (e) Formant sharpness Five types of feature quantities are obtained, and the upper limit value or the lower limit value is determined for each feature quantity, and the Voice and non-voice are discriminated according to the magnitude relation.

[0005]

However, in the above speech / non-speech discriminating apparatus, the feature amount based on the feature of each vowel in the voice is not used, and each vowel suitable for detection of each vowel in the voice is not used. There is a need for a highly accurate voice detection method that uses the standard model of. In addition, the upper limit value or the lower limit value is determined for each feature amount, and the voice /
In the method of discriminating non-speech, it is difficult to set a threshold value for each feature amount when the number of feature amounts increases, and at the same time,
It cannot be said to be robust against fluctuations in the value of the feature amount such as when stationary noise is added. Furthermore, it can be said that a more reliable determination method is to consider voices, especially vowels, by considering several frames as one block rather than determining vowelness for each analysis frame.

The present invention solves the above-mentioned problems.
An object of the present invention is to provide a high-performance voice detection device suitable for voice signal processing such as voice recognition. The present invention uses a standard model for each vowel suitable for detecting each vowel of the voice,
Provided is a voice detection device based on the detection of each vowel of a voice. Furthermore, as an evaluation value for comprehensively determining the feature amount indicating whether it is voice or not, it is calculated by considering several frames for one frame considered to be useful for determining vowelness as one block. By using the frame average log likelihood,
A relatively simple and high-performance speech detection device that is relatively easy to set the threshold and has a certain degree of robustness to fluctuations in the value of the feature amount such as when stationary noise is added. The purpose is to provide.

[0007]

In order to solve the above-mentioned problems, the present invention solves the above-mentioned problems by inputting a frame unit (every fixed time).
The presence probability of vowels is obtained by using at least one feature quantity of the autocorrelation coefficient of 1st order or more and the cepstrum coefficient of 1st order or more that characterize the extracted speech by collectively using several frames. The feature is that vowel detection is performed and only voice is detected by this.

Further, according to the present invention, at least one of the feature quantities of the autocorrelation coefficient of the first order or more and the cepstrum coefficient of the first order or more, which characterizes the voice extracted in advance on the basis of a large number of vowel learning data, is used. , A standard model is created for each vowel by collectively using several frames, and a voice is detected using the vowel standard model.

Further, the voice detecting apparatus of the present invention is designed to detect vowels, and to detect the vowels at intervals of a fixed time.
A feature amount extraction unit that extracts a plurality of voice feature amounts such as an autocorrelation coefficient of the second order or higher and a cepstrum coefficient of the first order or higher, and the feature values extracted by the feature extraction unit in advance for learning data of many vowels are used. A vowel standard model creation unit that creates an average value and a covariance matrix for each vowel, and creates a standard model for each vowel, and an autocorrelation coefficient of the first or higher order extracted from the input signal in frame units by the feature extraction unit. And at least one of the first-order cepstrum coefficients, a likelihood calculation unit that calculates a log likelihood with each vowel standard model created by the vowel standard model creation unit, and a likelihood calculation unit by the likelihood calculation unit. A frame average log likelihood is calculated for each vowel for the calculated several frames before and after, and a vowel judgment unit for judging whether or not the vowel is other than that by comparing with a certain appropriate threshold, and a certain level of power or more. Of the input signal, the final deciding section for deciding the lump as a voice when the ratio of the number of vowel samples judged to be any vowel by the vowel judging section is above a certain threshold value. Is.

[0010]

With the above-described structure, the present invention uses a feature quantity suitable for vowel detection based on the characteristics of each phoneme in a voice, and uses a large number of highly reliable vowel data to generate a vowel standard model for each vowel. It is possible to perform high-performance voice detection by creating the vowels for a few frames and detecting the vowels by a statistical method and detecting only the voice.

[0011]

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, 11 is a feature extraction unit that extracts a plurality of feature amounts for voice detection, and a power calculation unit 11a that calculates the power for each frame (constant time).
And an autocorrelation coefficient calculation unit 11b that calculates the 1st and 7th order autocorrelation coefficients for each frame, and 1 for each frame.
It is composed of a cepstrum coefficient calculation unit 11c for calculating the second and third-order cepstrum coefficients. These feature quantities are used to detect the vowel characteristics of the input signal.

Next, 12 is a vowel that creates a standard model for each vowel by calculating the average value and covariance matrix for each vowel using the feature quantities extracted in advance by the feature extraction unit 11 with respect to a large number of vowel learning data. It is a standard model creation unit. Reference numeral 13 is a vowel standard model created by the vowel standard model creation unit 12 for the first-order and seventh-order autocorrelation coefficients and the first-order and third-order cepstral coefficients of the input signal for each frame output from the feature extraction unit 11. A likelihood calculator that calculates a log-likelihood with a model, and 14 is a logarithmic likelihood calculated similarly by the likelihood calculator 13 in several frames before and after the frame used in the calculation by the likelihood calculator 13. Is a vowel determination unit that determines whether or not the input signal number frame is a vowel by calculating the frame average log likelihood for each vowel and comparing it with a certain appropriate threshold. Reference numeral 15 denotes a final determination unit that determines a vowel as a voice when the number of frames determined by the vowel determination unit 14 as a vowel for a lump of an input signal having a power level equal to or higher than a certain threshold is equal to or more than a certain threshold. ..

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. When the acoustic signal is input through the microphone, the feature extraction unit 11 first extracts a plurality of feature amounts. Power calculator 1
In 1a, the power value for every fixed time is calculated by (Equation 1), for example. The fixed time interval is 200 points (20 ms) here, for example, when the sampling frequency is 10 KHz.
This time unit is called a frame.

[0014]

[Equation 1]

Where P _i 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. In the autocorrelation coefficient calculation unit 11b, the first-order autocorrelation coefficient A _i (1) is calculated for each frame (Equation 2), and the 7th-order autocorrelation coefficient A _i (7)
Is calculated by (Equation 3).

[0016]

[Equation 2]

[0017]

[Equation 3]

Further, A _i (1) and A _i (7) are normalized by the zero-order autocorrelation coefficient. The cepstrum coefficient calculation unit 11c obtains the primary and tertiary cepstrum coefficients C _i (1) and C _i (3) in the frame i by the linear prediction analysis.

In the vowel standard model creating unit 12, the feature extracting unit 11 has previously created vowel parts of a large number of voice data.
The feature values obtained in step 1 are extracted, and the average value and covariance matrix for each vowel are calculated by the following method using these feature values to create a standard model for each vowel. That is, as the vowel data, the learning data y _N of the vowel k (the number of data N)
When y _N can be assumed to follow an m-dimensional multidimensional normal distribution, its mean value μ _k and covariance matrix Σ _k are given by (Equation 4),
It can be calculated as in (Equation 5).

[0020]

[Equation 4]

[0021]

[Equation 5]

However, t indicates a transposed value. As a result, the model shape (average value μ _k and variance Σ _k ) of the standard model for each vowel is obtained. However, y _N and μ _k are m-dimensional vectors (m-dimensional feature parameters), and Σ _k is m × m
It is a matrix of dimensions. As the vowel data, for example, data of a vowel center frame ± 2 frames may be used by cutting out a vowel portion as data for learning a vowel k of a certain standard speaker. Further, by using the data of a plurality of speakers, it is possible to create a standard pattern that is strong against variations in the utterances of the speakers.

The likelihood calculation unit 13 has a logarithmic likelihood with respect to each vowel standard model created by the vowel standard model creation unit 12 for some feature parameters of the input signal for each frame output from the feature extraction unit 11. Is the part to calculate.
The distance measure used for vowel detection is a statistical distance measure assuming that the distribution of each feature parameter used is a multidimensional normal distribution. Logarithmic likelihood L _ik of the input vector (spectrum) x _i of the _i- th frame with respect to the standard model k for each vowel
Is calculated by (Equation 6).

[0024]

[Equation 6]

Where x _i is an m-dimensional vector (m-dimensional feature parameter), t is a transposed value, −1 is an inverse matrix,
C indicates a constant.

The vowel determination unit 14 performs vowel determination using N frames (which are referred to as segments) before and after the target frame to be detected in order to express the temporal continuity of the vowel. Using the log-likelihood _Lik calculated by the likelihood calculator 13 for each vowel, if the following conditional expression (Equation 7) is satisfied, the segment is regarded as a vowel.

[0027]

[Equation 7]

However, L _kTH is a discrimination threshold (a threshold of frame average log likelihood) for the vowel standard model k.

As described above, the evaluation value (logarithmic likelihood) with which the influence of each characteristic parameter can be effectively and comprehensively determined.
By using, it is possible to construct a system that is more robust against variations in the value of the feature amount such as when stationary noise is added, compared to the method of setting the threshold value for each feature parameter.
Further, there is an advantage that many thresholds do not need to be determined by a heuristic method. Further, by considering the number of frames of the acoustic signal as one block, the determination is more effective for continuous speech such as vowels.

The final decision unit 15 is a final decision unit for deciding a vowel sound as a voice when the number of vowel samples for the lump of the input signal having a power level above a certain level is above a certain threshold value. The final determination section 15 first detects, as a voice candidate section, a section that exceeds a predetermined power threshold value by a predetermined length or more from the power value series obtained by the power calculation section 11a. Within this voice candidate section,
When the number of segments determined to be vowels k by the vowel determination unit 14 is counted, and the number of vowel segments determined to be vowel k is C _k , and a threshold value M _{k of the} number of vowel segments in a predetermined section is set. If the conditions of (Equation 8) are satisfied, it is determined that the voice candidate section is voice. Do this for all vowels.

[0031]

[Equation 8]

The results of experiments actually carried out by this method are shown below. (Table 1) shows four feature parameters used in this method. As a result of preliminary experiments, these characteristic parameters are parameters that are relatively good in separating speech from other non-stationary noises, and are easily obtained in the process of calculating the LPC cepstrum coefficient. First, the first-order normalized autocorrelation coefficient and the first-order linear prediction coefficient are parameters suitable for voiced / unvoiced discrimination, and the seventh-order normalized autocorrelation coefficient distinguishes low-frequency noise. Is a parameter suitable for. The third-order LPC cepstrum coefficient is a parameter showing a characteristic characteristic of / i / among the five vowels.

[0033]

[Table 1]

The voice data is 200 Japanese words uttered by 10 men. The standard model was created by using the phonological center ± 2 frames of each vowel uttered by a standard speaker.
However, considering the calculation efficiency, it is assumed that there is no correlation between each parameter, and only the diagonal component of the covariance matrix is used for the calculation. As noise data, 5 noise groups (about 9
00 samples) non-stationary noise was used. The analysis conditions are shown in (Table 3).

[0035]

[Table 2]

[0036]

[Table 3]

A standard model was created from vowel data of 5 male speakers, and a voice detection experiment and a non-stationary noise removal experiment of (Table 2) were conducted for 10 speakers including the standard speaker. .. FIG. 2 shows the relationship between the voice detection rate and the noise erroneous detection rate when the vowel segment length is changed from 1 frame to 11 frames. The optimum value of the detection performance can be obtained by appropriately changing the discrimination threshold, but there is almost no difference in the discrimination performance in 5 frames or more. After all, the discrimination threshold = 7 frames with vowel segment length =
-When 1.2, voice detection rate 99.3% (noise false detection rate 9.0
%)was gotten.

Next, in order to evaluate the detection performance of the present method under steady noise, the relationship between the S / N ratio and the detection rate when white noise was added was examined. FIG. 3 shows the relationship between the voice detection rate and the noise erroneous detection rate for each S / N ratio when the vowel segment length is fixed to 7 frames. As a result, the detection performance is hardly affected up to the S / N ratio of 12 dB.

As described above, according to the voice detecting apparatus of the present embodiment, the feature amount extracting section 11 for extracting a plurality of feature amounts of the voice at a constant time from the input signal, and the learning data regarding a large number of vowels are framed beforehand. Calculate the average value and covariance matrix for each vowel using the feature amount extracted in units,
Vowel standard model creation unit 1 that creates a standard model for each vowel
2, a likelihood calculation unit 13 that calculates a logarithmic likelihood of a plurality of voice feature values obtained from an input signal and each vowel standard model created by the vowel standard model creation unit, and a frame for performing vowel determination. Calculate the frame average log-likelihood for each vowel by using the log-likelihood of several frames before and after, and compare it with a certain appropriate threshold to determine whether the input signal several frames are vowels. And a vowel determining section 14 for a block of input signals having a power level higher than a certain level.
And a final determination unit 15 for determining the vowel sample as a voice when the number of vowel samples determined to be any vowel is equal to or larger than an appropriate threshold value,
It is possible to provide a voice detection device that can accurately determine voices in various acoustic signals with a relatively simple configuration.

In the above embodiment, an example in which the autocorrelation coefficient and the cepstrum coefficient are used as the feature quantity for detecting the vowel characteristic of the input signal in the feature extraction section has been described, but the present invention is not limited to this. Alternatively, a partial autocorrelation function or a mel cepstrum coefficient may be used.

[0041]

As is apparent from the above embodiments, according to the present invention, a plurality of feature quantities that characterize speech are extracted, a vowel standard model is created using a large number of learning vowel data, and input. It is configured to calculate the log-likelihood obtained from the vowel standard model and a plurality of feature quantities obtained from the signal, and to detect the vowels by collectively detecting the vowels for several frames. It is possible to provide a voice detection device capable of accurately determining whether an input signal is voice or not with a simple configuration.

[Brief description of drawings]

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

FIG. 2 is a diagram showing a relationship between a voice detection rate and a noise false detection rate due to the influence of a vowel segment length.

FIG. 3 is a diagram showing a relationship between a voice detection rate and a noise erroneous detection rate due to the influence of the S / N ratio.

[Explanation of symbols]

 11 feature extraction unit 11a power calculation unit 11b autocorrelation coefficient calculation unit 11c cepstrum coefficient calculation unit 12 vowel standard model creation unit 13 likelihood calculation unit 14 vowel judgment unit 15 final judgment unit

Claims (3)

[Claims] 1. At least one of a first or higher order autocorrelation coefficient or partial autocorrelation function and a first or higher order cepstrum coefficient or mel-cepstrum coefficient which characterizes a voice extracted from an input signal in frame units (every constant time). A voice detection method characterized by detecting the presence of a vowel by detecting the vowel presence by collectively using a few frames for the feature quantity and detecting only the voice. 2. At least one of a first-order or higher autocorrelation coefficient or partial autocorrelation function and a first-order or higher-order cepstrum coefficient or mel-cepstrum coefficient that characterizes speech extracted in advance from a large number of vowel learning data in frame units. A voice detection method characterized in that a standard model is created for each vowel by collectively using several frames for a feature amount and a voice is detected using the vowel standard model. 3. A feature extraction unit for extracting a first-order or higher-order autocorrelation coefficient or partial autocorrelation function and a first-order or higher-order cepstrum coefficient or mel-cepstral coefficient, which characterizes a sound at regular intervals from an input signal, and a large number of pre-existing features. A vowel standard model creation unit that creates a standard model for each vowel by calculating an average value and a covariance matrix for each vowel using the feature amount extracted by the feature extraction unit for vowel learning data, and a frame unit from an input signal In the vowel standard model creation unit, at least one or more feature amounts of the first or higher-order autocorrelation coefficient or partial autocorrelation function extracted by the feature extraction unit and the first-order cepstrum coefficient or mel-cepstral coefficient Likelihood calculator that calculates log-likelihood with each created vowel standard model, frame to detect vowels and several frames before and after it In the vowel determination section for calculating the frame average log likelihood for each vowel using the log likelihood calculated by the likelihood calculation section and comparing it with a certain appropriate threshold And a final determination unit that determines a vowel as a voice when the number of vowel samples determined to be one of the vowels by the vowel determination unit with respect to a lump of the input signal having a power level higher than or equal to a certain threshold And a voice detection device.