JPS58211793A - Detection of voice section - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a voice segment detection method for automatically detecting voice segments and non-speech segments from an acoustic signal including speech uttered by a human.

Speech segment detection is essential in speech recognition systems, analysis systems for speech synthesis, speech information compression, and the like.

Figure 1 shows a block diagram of the automatic speech recognition system.

show. 1 is a sound processing section, 2 is a speech section detection section, and 3 is a recognition section. In the speech recognition system, the speech section detection section 3 is located at the front stage of the recognition section 3 as shown in Fig. 1, and even if the performance of the recognition section 3 at the rear stage is very good, it is difficult to correctly detect the speech section. If this is not possible, it will be difficult to obtain correct recognition results, so the quality of the performance of this speech section detection section is an important factor that greatly affects the entire system.

Conventional methods for detecting speech periods often focus on the difference in signal energy between a speech signal and a non-speech signal (Zunawaji environmental noise) and set an appropriate threshold for the signal energy value to detect speech periods. . In addition, in order to improve the detection accuracy of speech signals such as unvoiced consonants, which have low signal energy and are not much different from environmental noise, there is a method that also uses threshold processing of the number of zero crossings for each appropriate time length of the signal. be.

This 19th article is described in "Voice Recognition" written by Yasunaga Niimi.

In the conventional method using threshold processing, the threshold value to be set depends on the environment. This threshold setting is often determined experimentally, but it is a threshold for detecting voice sections5, and a threshold for separating the environmental noise signal of the environment used and the voice signal. If the usage environment changes, the threshold must be reset, and there is a drawback that it lacks adaptability to changes in the usage environment. In addition, if the environmental noise energy of the environment used is large and the noise is white noise, there will be almost no difference in energy and zero crossing number between the environmental noise and unvoiced speech, so the conventional method Voice intervals cannot be detected correctly. As described above, conventional voice segment detection has the drawbacks of being limited in the environment in which it can be used and not being able to adapt to changes in the environment.

The present invention aims to provide a voice section detection method that improves the drawbacks seen in the conventional example, and is a method that has an environment learning function.

Through environmental learning, the average energy level of environmental noise and
An average value vector of LPC cepstral coefficients representing average characteristics of the spectrum is determined in advance.

Next, the actually input audio signal containing noise is normalized to a signal level using the environmental noise average energy obtained through learning. Furthermore, the Euclidean distance between the LPC cepstral coefficients of the input signal and the LPC cepstral coefficients of the environmental noise is also determined.

According to the present invention, speech intervals are detected by threshold processing of the two parameters, the normalized 5 Beh2' energy and the Euclidean distance, which are obtained in this way.

By this method, compared to the conventional example, it is possible to obtain a remarkable improvement in response to changes in the environment in which the product is used. The voice section detection method according to the present invention will be explained in detail below.

FIG. 2 is a rough functional block diagram of the speech section detection section according to the present method. As shown in FIG. 2, the voice section detection using this method involves an acoustic analysis section 4 that calculates the feature parameters to be used, an environment learning section 5 that learns the characteristics of the environment to be used, and a system that actually detects the voice section. It is composed of a voice section detection section 6.

In the speech segment detection method of the present invention, first, a standard environment is learned in advance.

This process is almost the same as the work for setting thresholds in the conventional example, and includes the average energy Es of the standard environment and the threshold TJ! of the two signal energies that separate speech and non-speech. :II
``'j!:2'' is set. The constant obtained through this process is stored in the speech interval detection unit 6. In the conventional example, this constant is set every time the environment 6 ・-- changes drastically. However, this method does not need to be performed once the constant 'rE1.TE21 ES is found.〇 Set a standard environment with low environmental noise energy and good signal noise reduction, and convert the environmental noise signal to , for some suitable length of time (
Energy E and logarithmic energy EL are determined for each frame length (referred to as frame length) using equations (1) and (2).

E L = 10 X tog 1゜E...
(2) Obtain standard environmental average energy EB from the average value of E obtained within a certain period of time. Further, based on the average value and variance of EL, an energy threshold TE1 is set such that if EL is less than this value, it is a non-speech or voice frame. Furthermore, from the average value and variance of the logarithmic effort 7Behni-Gie EL of voiceless consonants in the speech signal uttered by many speakers under a standard environment, an energy threshold is determined that if EL exceeds this value, it is a speech frame. Set fE2.

Next, we will discuss the acoustic analysis section.

After the input from the microphone etc. is completed, the sound processing section 1 shown in Fig.
The acoustic signal subjected to the /D conversion is sent to the acoustic analysis section 4 in FIG.
sent to. The acoustic analysis section 4 analyzes the input acoustic signal for each appropriate frame length, and then analyzes the input acoustic signal Ig for each appropriate frame length.
and calculates the parameters used in both 110 in the voice detection section 6. The parameters to be calculated are the signal energy E expressed by equation (1) and the LPC kegstral coefficient vector C, which is a parameter representing the spectral characteristics of the signal.

The explanation of the method for calculating the LPC cepstral coefficient C will be omitted, but details can be found in J, L), 1viarker and
A, f(.

Gray, Ir. :Linear Predict
ion of 5peech.

Springer-Ver Lag (1976).

The parameters calculated by the acoustic analysis unit 4 are used for environmental learning when the speech segment detection using this method starts operating or when the environment in which it is used changes significantly and many detection errors occur in speech segment detection. Sent to Department 5. The input acoustic signal at this time does not include a voice signal and is only an environmental noise signal.

The environment learning unit 5 examines the parameters for each frame sent from the acoustic analysis unit 4 and calculates the signal energy of the usage environment by comparing the signal energy of the standard environment. The normalization coefficient s for normalization to the W energy level, the average value vector C8 of the LPC cepstral coefficients representing the average spectral characteristics of the usage environment noise, and the distance threshold 'rD are calculated. Distance threshold T
D is a threshold value for determining whether the frame is a voice frame or a non-voice frame based on the Euclidean distance between the LPC cepstrum coefficient vector C and the average value vector C8.

The normalization coefficient NS is calculated as follows.

The average value ``rJ'' of the environmental noise A and Lugi sent for each frame is calculated, and the standard environment average energy E stored in advance in the speech section detection unit 6 through learning of the standard environment is calculated.
s is calculated using equation (3).

NS”'5-EN --(3)9 Be-
“・Ma15, average value vector of LPC cepstral coefficients C8
is obtained by calculating the average value for each element of the LPC cepstrum coefficient vector sent for each frame. Furthermore, the Euclidean distance between this C8 and the LPC cepstral coefficient vector used to calculate C8 is calculated for each frame, and from the average value and variance of D, if the Euclidean distance value is less than this value, it is a non-speech frame. A threshold value TD is calculated.

The speech interval detection unit 6° generates a normalized signal from the signal energy E1 sent from the acoustic analysis unit 4, the normalization coefficient Ns1 obtained by the LPC cepstrum coefficient vector human environment learning unit 5, and the LPC cepstrum coefficient average value vector C8. The Euclidean distance between the logarithmic energy ENL and C8,0 is determined to determine whether it is a voice signal or a non-voice signal.

The normalized signal logarithmic energy ENL is determined by equation (4). Further, the Euclidean distance #D is determined by equation (6).

0 ENL=10XtO(Jlo(E-Ns) ---
・From (4) -f(C-C5)T-(U-shis)l ---
・-(5n (represents Tco device matrix)) Whether it is an audio signal or a non-audio signal is determined as follows.

"rJL≦TE1 Non-voice TE, <E,, L<TE2AND D<'rD Non-voice TE1, TE2 ANDD>TD Voice" N
L≧TE2 Voice This determination method is a method of improving accuracy by using spectrum information of the signal when it is unclear whether the signal is a voice signal or non-voice based only on the energy of the signal. In fact, since it uses the characteristics of the entire spectrum called the LPC cepstrum, rather than information on a part of the signal spectrum like the number of zero crossings in the conventional example, there is little performance degradation due to changes in the noise spectrum.

The results determined for each frame in this way are subjected to smoothing processing, and the final voice section is determined and 11 bases are applied.

FIG. 3 is a functional block diagram of a speech section detection circuit according to this method. The acoustic signal inputted from a microphone or the like is subjected to H/D conversion, and is sent to the signal energy calculation unit 7 and the LPC cepstral coefficient calculation unit 8 for each appropriate frame length. The signal energy calculation section 7 calculates the signal energy E, and the LPC cefnutrum coefficient calculation section 8 calculates the LPC cepnutrum coefficient vector C. The flow of the calculated parameters differs depending on whether environmental learning or voice section detection is performed, and this control is performed by the control unit 9. In the figure, the broken line indicates the control line.

In the case of environmental learning, the signal energy E is multiplexer 1
0 to the average value/variance calculation unit 11, where the average energy EN is calculated. This EN is further sent to the normalization coefficient determining section 12, where a normalization coefficient NB is determined. Also, the LPC cepstrum coefficient vector C is
1 through o to the mean value variance calculation unit 11.
．． It is stored in 4IFO buffer 13. Average value variance calculation unit 1982-211793 ('4') The calculation unit 11 calculates the average value vector CB and sends it to the LPC cepstrum coefficient average value vector memory 14. When the data is stored in the PC cepstrum coefficient average value vector m-v+)-14, the FIFO buffer 13 sends the LPC cepstrum coefficient C to the knee distance calculation unit 16 through the multiplexer 16, and the Euclidean distance is calculated. .

The calculated Euclidean distance is sent to the mean value variance calculating section 11 through the multiplexer 1Q, and the mean value and variance are calculated. This average value and variance value are sent to the threshold value TD determining section 17, and the threshold value °fD is determined.

On the other hand, when performing voice section detection, the signal energy Qi E is sent to the normalized logarithmic energy calculation unit 8, converted to normalized logarithmic energy ENL, and sent to the ternary comparison unit 19. Further, the LPC cepstrum coefficient vector C is sent through the multiplexer 16 to the Nino-Rinode distance calculation section 16 to calculate the Euclidean distance, and the value is sent to the binary comparison section 20. The binary comparison unit 19 uses the normalized logarithmic energy 13, -- ENL and the threshold TI8.1. “Compare with fE2.The result is “NL≦TE1 or”1’JL≧°1゛E
2 - The determination result of the ternary comparison section 190 is sent to the smoothing processing section 22 through the multiplexer 21. In cases other than those described above, the comparison result between the sharpness, distance, and threshold value TD by the binary comparison section 2o is sent to the smoothing processing section 22 via the multiplexer 21. The smoothing processing unit 22 smoothes the determination results sent for each frame, determines a voice section, and outputs the same.

FIG. 4 shows the effect of the present invention in normalizing signal energy through learning. FIG. 4A shows the environmental noise logarithmic energy distribution of a standard environment (upper row) and the logarithmic energy distribution of voiceless consonants of speech uttered under that environment (lower row). Now, when the environment changes and the average energy of the voice and the energy of the environmental noise become about 20 dB, the conventional signal energy distribution is shown in Figure 4B, and the signal energy distribution of the present invention is shown in Figure 4B. Figure 4C
Shown below. Figure 4B shows the environmental noise logarithmic energy distribution (upper row) when the environment changes and the signal noise ratio of voice energy and environmental noise energy is approximately 20 dB, and the Distribution of logarithmic energy of voiceless consonants in tailed speech (bottom row). Furthermore, C in FIG. 4 shows the distribution of each normalized logarithmic energy under the same environment as in FIG. 4B. In the figure, the broken line indicates the normal distribution assumption. These figures show that when the logarithmic energy is not normalized as in the past, the signal-to-noise degradation deteriorates and the logarithmic energy of the environmental noise approaches the energy of the unvoiced consonant, making it difficult to separate it using the threshold set under the standard environment. , and the threshold must be reset. In addition, even if the threshold is reset, the accuracy will decrease because the two distributions overlap greatly.In contrast, in the case of the present invention, the distribution of normalized logarithmic energy is the same as the distribution of logarithmic energy under the standard environment. Therefore, there is no need to change the threshold value at all, and the overlap of the two distributions is ! ! 4B, it is possible to reliably separate the environmental noise signal from the voiceless consonant of the voice.
A better result can be obtained than the logarithmic energy of 15 -, which is often used in the conventional example.

Figure 5 shows “lhutal” uttered by a male speaker.
This figure shows the Euclidean distance B between the normalized logarithmic energy A of the voice and the average value vector of the LPC cepstral coefficients. Normalized logarithmic energy only /h/
The starting point of /U/ does not matter, but by using the Euclidean distance of the LPC cepstral coefficients,
These awkward parts become clear and the degree of detection can be increased.

FIG. 6 is a comparison between the voice interval detection method according to the present invention and the conventional voice interval detection method in which signal energy is determined using a fixed threshold value.

This figure was obtained from an audio take of 200 words uttered by one male speaker, and shows the identification rate determined by equation (6) when the signal-to-noise ratio of the utterance environment changes. The one-dot chain line indicates the conventional example, and the solid line indicates the one according to the present invention.

JP-A-8-211793 (5) (6) In the conventional method, when the signal-to-noise ratio deteriorates below 3 odB, the identification rate decreases rapidly; Since the frame is determined to be a voice frame, the identification rate becomes 60%, making it impossible to detect a voice section. In contrast, with this method, the identification rate hardly changes up to a signal-to-noise ratio of about 20 dB, and even when the signal-to-noise ratio decreases to about 1 odB, a good value of 91% can be obtained, making it possible to adapt to environmental changes compared to the conventional method. It is possible to boast of a remarkable improvement in the shortcomings of gender.

As described above, the present invention is a voice section detection method characterized by learning the energy level and spectrum of noise in the environment in which the method is used, and also using spectrum information.

By performing learning, it is possible to normalize the energy level of the input signal containing noise to a constant level, reducing the influence of environmental changes, and it also makes use of spectral information. Therefore, voice sections can be detected with high accuracy.

[Brief Description of the Drawings] Fig. 1 is a block diagram showing a general configuration of an automatic speech recognition system, Fig. 2 is a block diagram showing a speech interval detection method according to the present invention, and Fig. 3 is a block diagram showing a general configuration of an automatic speech recognition system. A functional block diagram of the voice section detection circuit, shown in FIG. 4. C is a comparison diagram of the signal energy distribution of the present invention and the conventional example, and B is the normalized logarithmic energy of the actual speech signal of the present invention. and Euclidean distance, and FIG. 6 is a diagram showing changes in the present invention and the conventional example with respect to changes in the signal-to-noise ratio of the environment. 1...Acoustic processing unit, 2...Speech section detection unit, 3...Recognition unit, 4...K-sound analysis unit, 6... ...Environmental learning unit, 6...Speech section detection unit, 7...Signal energy calculation unit, 8.
...LPC cepstrum coefficient calculation section, 9...
...control section, io. 1621・・・Multibrefusa, 11・・・・・・
・Hei 18: Mean value variance calculation unit', 12...:...Normalization coefficient determination unit, 13...F, IFO buffer, 14...
・・LPC cepstral coefficient average value vector memory,
16...Euclidean distance calculation unit, 17...
...Threshold value determination unit, 18...Normalized logarithmic energy calculation unit, 19...Third value comparison unit, 20...
. . . binary comparison section, 22 . . . smoothing processing section. Name of agent: Patent attorney Toshio Nakao and 1 other person 1st
Bad Figure 2 l Figure 4 Positive KL-based Suke Elephant χ・ru 1000 Figure 5

Claims (2)

[Claims] (1) Set an energy threshold in advance based on the average energy level of environmental noise in the standard environment, use the environmental learning function to learn the average energy level of environmental noise in the environment to be used, and A normalization coefficient is determined from the average energy level of environmental noise in the environment, the energy level of the audio input signal is normalized using the normalization coefficient, and the energy level of the audio input signal is normalized using the energy level of the normalized audio input signal and the energy threshold. A speech interval detection method characterized by detecting an interval. (2) Using the environment learning function, learn the average value vector of LPC cepstral coefficients that expresses the spectral characteristics of the environmental noise of the environment in which it is used, and obtain the average value vector of the LPC cepstral coefficients and the audio input signal. L that can be done
Find the Euclidean distance with the PG cepstrum coefficient vector, set the distance and threshold based on the Euclidean distance, and calculate the energy level and energy threshold of the audio input signal normalized by the distance threshold and the normalization coefficient. 2. The voice interval detection method according to claim 1, wherein the voice interval is detected using the following.