JPS62211698A - 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 (Field of Industrial Application) The present invention relates to a speech signal processing method, and particularly to a speech segment detection method in a speech recognition device.

(Prior art) As a previously proposed speech interval detection method, Japanese Patent Application Laid-Open No. 1986-1
There is a method disclosed in No. 14900. This method makes use of the difference in power spectra between the noise section and the speech section, and detects the power spectrum based on the magnitude of the remaining power spectrum after subtracting the pre-registered noise power spectrum from the input Coverworth vector. .

This conventional method will be briefly explained below with reference to FIG.

FIG. 2 is a block diagram for explaining a conventional voice section detection method. The input audio signal received by the microphone 101 is passed through each of the low, middle and high band filters 102 and 1.
03, 104 and rectifying and smoothing section 105 for each band
is decomposed into the input power of each band by These three input powers pass through the multiplexer 10B and enter the environmental noise learning section 107 and the environmental noise removing section 108. The environmental noise learning unit 107 stores the noise power learned in advance, and the environmental noise removing unit 108 reads this noise power and subtracts the noise power from the input power. The difference power spectrum remaining after this subtraction enters the energy determination section 109, where the difference power spectrum is compared with a threshold value preset in the energy threshold memory 110, and the first sound/non-sound determination is performed. . If it is not possible to distinguish between voice and silence by this determination, in order to perform voice recognition with a high recognition rate for voice, a determination unit 111 using a statistical distance measure in the next stage uses the difference power spectrum and standard pattern memory 112. The distance between the pre-stored voiced consonant φ and the unvoiced consonant is calculated to perform the second voiced/unvoiced determination.

(Problems to be Solved by the Invention) However, according to this conventional speech interval detection method,
If the nature of the noise, such as the noise power, changes between the time of noise measurement and the time of voice detection, false detections may occur or detection performance will significantly deteriorate. There are problems in that it is necessary to re-register the basic noise power, etc., and a speech recognition method with a high discrimination rate for speech is also required.

The present invention has been made in order to solve the above-mentioned problems such as the weakness against changes in noise properties and power fluctuations, the need for noise registration, and the need for a voice identification method.

SUMMARY OF THE INVENTION Therefore, an object of the present invention is to provide a method for detecting a speech section that can accurately detect a speech section even under high noise and can be used without being aware of noise changes or registration.

(Means for Solving the Problems) In order to achieve this objective, the voice section detection method of the present invention includes the following processing.

First, time intervals other than the intervals determined to be voice interval candidates are determined to be noise intervals using an input feature pattern (referring to a set of feature values representing the characteristics of the input signal) for each predetermined frame length. Has processing.

In this case, for example, the input feature pattern in this noise interval includes a process of learning a set of characteristic values representing the characteristics of the noise as a noise standard pattern. It is preferable to learn a set of the average shape and the average fluctuation range of this noise feature pattern as the above-mentioned noise standard pattern. Further, this noise standard pattern learning is preferably performed only within a noise interval while following changes in the nature of the noise, for example, slow changes (this learning is referred to as interval limited tracking learning).

Furthermore, in this case, it includes a process of detecting the above-mentioned speech section candidate as a speech section based on the magnitude relationship between the inter-pattern distance between the input feature pattern and the noise standard pattern and the threshold value.
2. A set of the average value and average fluctuation width of the inter-pattern distance (hereinafter sometimes simply referred to as distance) within the noise period. 2. Distance during noise. A set of characteristic values representative of the characteristics (distance during noise standard. It is preferable to perform section-limited tracking learning as a pattern (referred to as a pattern). Further, this threshold value is preferably a value calculated based on the obtained noise distance standard pattern.

More specifically, this threshold value is calculated according to the following formula, where T is the threshold value, D is the estimated average value of the distance between patterns in the noise interval, E is the estimated average deviation, and C is an arbitrary constant. It is preferable to do so.

In addition, in the learning method of the noise standard pattern, specifically, the discrete time is l, the characteristic value to be learned for the input feature pattern is fi, the characteristic value in the noise standard pattern is FL, and the constant Kt. -K> 1, within the noise section ■i-((K -1) ■i-1/K) + fH/K
It is preferable to use a learning method in which calculations are made according to the following formulas within the speech interval and the following formulas.

In addition, the learning method for the distance standard pattern in noise is preferably as follows:
When the discrete time is i, the characteristic value to be learned for the inter-pattern distance is g4, the characteristic value in the noise time distance II standard pattern is Gi, and the constant is a constant with L>1, the noise 84 = ((L-1
) a, -1/L) + g, /L and within the speech interval, it is preferable to use a learning method in which calculations are made according to the following formulas.

According to the above-described learning of the noise standard pattern and the noise distance standard pattern, it is not necessary to newly learn the noise, and it is possible to accurately detect the speech section, and even if a new noise source occurs, it is possible to detect the speech section longer than the speech section. If it continues, it can learn from this and return to a state where it can detect voices again, allowing it to accurately detect voice sections.

Furthermore, in performing the above-mentioned judgment, in addition to comparing the magnitude relationship between the distance between patterns and the threshold value, the speech interval candidates obtained by the above-mentioned judgment that satisfy various characteristics regarding the utterance length of the speech are compared with the original one. It is preferable to detect it by determining that it is a voice section. In this way, even in the case of pulse-like noise, there will be no malfunction due to the use of the characteristics related to the speech generation length.
Voice sections can be detected.

In this case, the characteristics regarding the utterance length of the voice include, for example: (a) The time during which the distance between patterns becomes larger than the threshold value is about 40 ms or more.

(b) It can be determined that the time during which the distance between patterns becomes smaller than the threshold value is approximately 400 ms or less, and (c) that no word is continuously uttered for approximately 4 seconds or more. In this case, if the voice section candidate satisfies at least one of the above characteristics (a) and (b) and the feature (c), it is determined that the voice section candidate is a voice section.

Further, it is preferable to include a process of determining a speech segment candidate that is not detected as a speech segment as a noise segment based on the magnitude relationship with the above-mentioned threshold value.

(Function) With this configuration, a speech section candidate and a noise section are detected from the speech input signal section, learning is performed within the noise section,
Since this method determines a speech segment candidate as a speech segment or a noise segment based on the magnitude relationship between the distance between patterns and the threshold value, there is no need to set aside time for special noise learning, and there is no need for special recognition methods like in the past. This is not necessary, and the voice section can be detected accurately.

(Embodiments) Hereinafter, embodiments of the voice section detection method of the present invention will be described with reference to the drawings.

FIG. 1 serves to explain an embodiment of the voice section detection method of the present invention. FIG. 1 is a block diagram showing the configuration of the device. First, the configuration and basic processing of this device will be briefly explained.

In FIG. 1, 201 is an input terminal, and this input terminal 2
The audio input signal I is input to 01. This speech section candidate I passes through a low-pass filter 202 of, for example, 4 kHz, and then passes through a low-pass filter 202 of, for example, 10 kHz in the primary stage circuit 203.
The audio data is sampled and held at a sampling frequency of , and is A/D converted into digital audio data D1, which is output and sent to the next preprocessing signal processing processor 204.

This pre-processing signal processing processor 204 converts the audio data D1 from the center frequency 250)1z to 4.0k.
Q = 20 channels at 115 octave intervals up to Hz
6. The IIR type single resonant digital bandpass filter group is calculated, and the absolute value of the output is averaged by 128 points, for example, and the obtained signal is sent to the speech section detection speech recognition microprocessor 205 as the input feature pattern D2. .

This voice section detection voice recognition microprocessor 205
, various processes as will be described later are performed to detect a voice section, a voice recognition process is performed, and the recognition result R is outputted to the output terminal 20B.

Next, an outline of an embodiment of the voice section detection process of the present invention will be described with reference to the block diagram of FIG. 1 and the process flowchart of FIG. 3. In this flowchart, processing steps are indicated by S.

First, this device is started to start processing operations (S1).

Next, various variables are initialized (S2), and then the audio input signal I is received at the input terminal 201 (S3).

This audio input signal equipment is connected to the audio data D in the circuit 203.
1 and output, and feature extraction of the input signal is performed in the signal processing processor 204.

Here, the input feature pattern is obtained as, for example, 20 numerical values (the number of channels) in which one frame is about 13 ms, and the value of the feature amount of the j-th channel of the i-th frame is xj,i.

Next, the inter-pattern pressure MDi between this input feature pattern and the noise standard pattern is calculated (35).

In this calculation, it is assumed that the threshold value Ti for the i-th frame has already been determined in the voice section detection process for the immediately preceding (i-1)-th frame. In this embodiment, first, voice section candidates are determined using this input feature pattern, and the remaining sections are determined to be noise sections. It is assumed that the noise standard pattern, which is the input feature pattern of this noise section, is expressed by the average value and average deviation of the feature amounts in each channel. The noise of the j-th channel obtained by learning the noise standard pattern up to the i-1th frame within the noise interval while following the slow changes in the nature of the noise (limited interval tracking learning) The average value as a standard pattern is the characteristic value Aj,
Let i be the characteristic value Oj, and the average deviation be the characteristic value Oj,i. in this case,
The inter-pattern pressure fllDi between the noise standard pattern and the input feature pattern in the i-th frame is calculated by the following equation.

However, L (x) is a ramp function and is given by the following equation.

Next, this distance Di is compared with the threshold value Ti of the i-th frame calculated from the noisy distance standard pattern learned up to the i-1th frame, and various values regarding the utterance length of the voice are calculated as follows. The voice section is determined based on the characteristics of (S6).

(2) If a section in which the distance D is greater than the threshold value Ti continues for three frames (approximately 40 ms) or more, that section is selected as a speech section candidate.

■30 frames (approximately 400ms) from the end of the voice section candidate
), if the distance Di is smaller than the threshold Ti, the voice section candidate is determined to be a voice section.

(2) If an interval in which the distance 1llll:D is greater than the threshold value Ti continues for more than 300 frames (approximately 4 seconds), or if it does not meet the criteria (2) and (2) described above, it is determined to be a noise interval.

Based on the results of the determination based on the criteria (1) to (2), a noise step-down standard pattern is learned in limited section tracking in order to obtain the first-order t+ as the threshold value TL of the th frame, which is 1. This learning is performed using a learning method that follows changes in the characteristics of the noise. This learning method will be explained below.

First, a case will be described in which it is determined that the i-th frame is part of a noise interval. In this case, learning is performed by updating the noise standard pattern and the noise distance standard pattern assuming that this judgment data is the latest noise data.This updating is performed according to the following equation.

Let the discrete time be i (equal to the frame number), and the characteristic value to be learned for the input feature pattern be fi
Let gi be the characteristic value to be learned for the inter-pattern distance, and let ■i be the characteristic value in the noise standard pattern,
When the characteristic value in the noise distance standard pattern is Gi, and the constants K and L are K>1 and I>1, respectively.

The characteristic value ■i within the noise section is ■i = ((K-1) Ft1/K) +f
The characteristic value Gi within the same interval as 4/ is G・=((L-1)Gi1/L)+
g4/L, therefore, the specific characteristic of the noise standard pattern Aj, 4 = ((Kt −L)
Aj, H-1/to 1) + xJ, i/K
qHowever, 1>1 0J, 1 = ((K2 -1) OJ, J-1/
Kl)"' xj-i-λj, i-11/Kz
However, 2>1, and the specific characteristic value Di of the distance standard pattern during noise
(Di is the average value of the distance between patterns in the noise interval up to the i-th frame) and Ei (Ei is the average deviation of the distance between patterns in the noise interval up to the i-th frame), provided that 3〉■ + l Di −D4−II/+ However, >I K1 ~N4 is an arbitrary constant that determines the followability of learning and is empirically determined, and the smaller it is, the more quickly fluctuations in noise can be coped with. However, if this constant is too small, speech that begins slowly will be determined to be noise.

Next, a case will be described in which it is determined that the first frame is a voice section candidate. In this case, the noise standard pattern and the noise standard distance pattern are not updated. Therefore, the following is the case for the noise standard pattern, and the same is for the noise standard distance pattern.

If the speech section candidate is not determined to be a speech section but is determined to be a noise section, learning is performed retroactively to previous frames.

Next, the next i+
Define the threshold value Ti and 1 for the first frame (Sa
), this threshold value Ti and 1 are calculated using the following equation using the characteristic values Di and Ei.

T・100・＋CEi 工◆11 Here, C is a constant that takes into account the safety factor, and as it increases, the detection sensitivity decreases, but the possibility of false detection due to noise decreases.

In this way, it is determined whether a voice section has been detected for the i-th frame (S9), and if it has not been detected, the process returns to step S3 and the same process is repeated.
Furthermore, if the detected voice section is detected, voice recognition is performed using this detected voice section (S10), and then it is determined whether the voice recognition for the input voice section has been completed (511), and the process is terminated. If so, this process ends (
S12), where the above-mentioned processing step 85
~Sll is performed by the voice section detection and voice recognition microprocessor 205.

FIG. 4 is a logarithmic diagram showing the transition of each variable for the voice "Tokkyo" uttered under noise in the embodiment of the present invention. The sum of the input feature patterns in this figure is the sum of the spectral intensities in multiple channels of the sonargrams obtained in the experiment. In the figure, the horizontal axis represents time, that is, frames, and the vertical axis represents the overall size as a logarithm.8 Curve I represents the distance Di of the input feature pattern, and curve ■ represents the threshold Ti. There is. As can be understood from this experimental data, the threshold value Ti
is the distance D in the noise section (1st to 20th frames)
It changes in accordance with the change in i, and remains constant within the voice section. Further, it can be seen that the change in distance #Di is larger than the total sum of input feature patterns, and that it is effective to use it for section detection. It should be noted that the start and end of the detected voice section are indicated by long scales, and it can be seen that the voice section can be detected accurately even under noise.

(Effect of the invention) As is clear from the above explanation, according to the present invention,
There is no need to set aside special noise learning time, and if the characteristics of the noise are changing slowly, the speech interval can be accurately determined without the need to newly learn the noise. can be detected.

In addition, since pulse-like noise is detected using characteristics related to the length of speech, only speech sections can be detected without malfunction.

Furthermore, even when a new noise source utters, if it continues longer than the word speech section to be accepted, this is learned and the state returns to a state where speech can be detected again.

Further, according to the method of the present invention, a special speech recognition method, which is necessary in the conventional method, is not used, so that it is simple and easy to implement the method of the present invention.

In order to achieve such effects, the present invention provides a speech section detection device in a speech recognition device, a transmission information compression device by transmitting only the speech section in a speech communication device, and other methods for detecting speech sections. It is suitable for application to devices in which

[Brief explanation of drawings]

FIG. 1 is a block diagram schematically showing an apparatus for explaining the speech interval detection method of the present invention. FIG. 2 is a block diagram showing an apparatus for explaining the conventional speech interval detection method. FIG. 4 is a flowchart of a process for explaining an embodiment of the speech interval detection method. FIG. 4 is a diagram showing experimental results for explaining an embodiment of the speech interval detection method of the present invention. 201...Input terminal, 202...Low pass filter 203...Sample hold, A/D converter 2
04... Preprocessing signal processing processor 205... Microprocessor for speech section detection and speech recognition. 206...Output terminal. Procedural amendment November 21, 1986

Claims (10)

[Claims] (1) A set of feature values representing the characteristics of an input audio signal for each preset fixed frame length is used as an input feature pattern, and when detecting a speech section using the input feature pattern, (a) Speech section candidates (b) processing to learn a set of characteristic values representative of noise characteristics within the noise interval as a noise standard pattern; (c) the above-mentioned A speech interval detection method comprising the step of detecting the speech interval candidate as a speech interval based on an inter-pattern distance between an input feature pattern and a noise standard pattern, and a magnitude relationship with a threshold value. (2) The threshold value is determined by learning a set of characteristic values representative of the characteristics of the inter-pattern distance at the time of noise within the noise interval as a distance standard pattern at the time of noise, and based on the obtained distance standard pattern at the time of noise. 2. The voice section detection method according to claim 1, wherein the value is calculated by using the calculated value. (3) The threshold value is T=■+C■, where the threshold value is T, the estimated average value of the distance between the patterns in the noise interval is ■, the estimated average deviation is ■, and an arbitrary constant is C. 3. The voice section detection method according to claim 2, wherein the voice section detection method is calculated according to a formula. (4) The speech interval detection method according to claim 1, wherein the learning of the noise standard pattern is performed by a learning method that follows changes in the characteristics of the noise. (5) In the learning method of the noise standard pattern, the discrete time is i, the characteristic value to be learned for the input feature pattern is f_i, and the characteristic value in the noise standard pattern is
When i and the constant K are K>1, within the noise interval ■_i={(K-1)■_i_-_1/K}+f_i/
5. The voice section detection method according to claim 4, wherein the learning method is used to calculate K and the inside of the voice section respectively according to the formula: ■_i=■_i_-_1. (6) The speech interval detection method according to claim 2, wherein the learning of the noise distance standard pattern is performed by a learning method that follows changes in the characteristics of the noise. (7) The learning method for the noise distance standard pattern is such that the discrete time is i, the characteristic value to be learned for the inter-pattern distance is g_i, and the characteristic value in the noise distance standard pattern is ■_i. When the constant L is a constant L>1, within the noise interval ■_i={(L-l)■_i_-_1/L}+g_i/
7. The voice section detection method according to claim 6, wherein the learning method is used to calculate L and the inside of the voice section according to the formula: ■_i=■_i_-_1. (8) Detection of a speech section based on the magnitude relationship between the distance between the patterns and the threshold value is performed by comparing the distance between the patterns with the threshold value, and in addition to the comparison between the distance between the patterns and the threshold value, the speech section candidate is 2. The method of detecting a voice section according to claim 1, wherein the voice section is detected as a section. (9) The characteristics regarding the utterance length of the voice are (a) the time during which the distance between the patterns becomes larger than the threshold value, and (b) the time during which the distance between the patterns becomes smaller than the threshold value. and (c) there are no words that are uttered continuously for more than approximately 4 seconds, and the speech segment candidate has at least the characteristics of (a) and (b) above and the characteristic of (c) above. 9. The voice section detection method according to claim 8, wherein the voice section is determined to be a voice section if one of the conditions is satisfied. (10) The speech segment detection method according to claim 1, further comprising a process of determining a speech segment candidate that is not detected as a speech segment to be a noise segment based on a magnitude relationship with the threshold value.