JPH0556512B2 - - Google Patents

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,
There is a method disclosed in No. 60-114900. This method makes use of the difference in power spectra between the noise section and the speech section, and detects based on the magnitude of the remaining power spectrum obtained by subtracting a pre-registered noise power spectrum from the input power spectrum.

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. An input audio signal received by the microphone 101 is decomposed into input power for each band by low, middle, and high band filters 102, 103, and 104, and a rectifying and smoothing unit 105 for each band. . These three input powers pass through a multiplexer 106 to the environmental noise learning unit 10.
7 and the environmental noise removal unit 108 . The noise power learned in advance is stored in the environmental noise learning section 107, and the environmental noise removing section 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 unit 109,
This difference power spectrum is compared with a threshold value preset in the energy threshold memory 110,
The first sound/non-sound determination is performed. If it is not possible to distinguish between voiced and unvoiced by this determination, the next step, the statistical distance scale-based determination unit 111, uses the difference power spectrum and the voiced consonants and unvoiced consonants stored in the standard pattern memory 112 in advance. In order to perform speech recognition with a good recognition rate for speech by calculating the distance between the consonant and the consonant, a second utterance/non-speech determination is performed.

(Problem to be Solved by the Invention) However, according to this conventional voice section detection method, if the nature of the noise, such as the noise power, changes between the time of noise or noise measurement and the time of voice detection, errors may occur. In order to avoid this, it is necessary to re-register the noise power, etc. that is the basis of the judgment, and a speech recognition method with a high discrimination rate for speech is also required. There was a problem.

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 Accordingly, 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 that 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.

It includes a process of learning the input feature pattern within this noise interval as a set of characteristic values representing the noise feature, that is, a noise standard pattern. In this case, for example, the input feature pattern in this noise interval is a set of noise feature amounts (hereinafter referred to as a noise feature pattern), and the set of the average shape and average variation width of this noise feature pattern is set as the above-mentioned noise standard. It is preferable to learn it as a pattern. In addition, learning this noise standard pattern is
Preferably, learning is performed only within the noise interval while following changes in the nature of the noise, such as gradual changes (this learning is referred to as interval-limited follow-up learning).

Furthermore, it includes a process of detecting the above-mentioned speech section candidate as a speech section based on the inter-pattern distance between the input feature pattern and the noise standard pattern and the magnitude relationship with a threshold value. In this case, the set of the average value and the average fluctuation range of the distance between patterns (hereinafter sometimes simply referred to as distance) within the noise interval is combined with the set of characteristic values representing the characteristic of distance during noise (noise It is preferable to perform section-limited follow-up learning as a time-distance standard pattern. Further, this threshold value is preferably a value calculated based on the obtained noise distance standard pattern.

To be more specific, this threshold value is defined as T = 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 calculate according to the formula D^+CE^.

In addition, in the learning method of the noise standard pattern, specifically, let the discrete time be i, the characteristic value to be learned for the input feature pattern be f _i , and the characteristic value in the noise standard pattern be F^ _i . and constant K is set to K>1, then in the noise section F^ _i = {(K-1)F^ _i-1 /K} + f _i /K and in the speech interval F^ _i = F^ _i- It is preferable to use a learning method in which each calculation is performed according to formula ₁ .

Further, in the learning method of the distance standard pattern in noise, preferably, a discrete time is i, a characteristic value to be learned for the inter-pattern distance is g _i , and a characteristic value in the distance standard pattern in noise is G. ^ _i and the constant L is a constant L > 1, then in the noise interval G^ _i = {(L-1)G^ _i-1 /L} + g _i /L and in the speech interval G^ _i It is best to use a learning method that calculates each according to the formula = G^ _i-1 .

According to the above-described learning of the noise standard pattern and the noise distance standard pattern, there is no need to newly learn noise, and speech sections can be detected accurately.

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, the voice section can be detected without malfunctioning because the characteristics related to the voice generation length are used, and even when a new noise source occurs, the voice section can be detected. If it continues for a longer time than that, it can learn this and return to the state where it can detect the voice again, and the voice section can be detected accurately.

In this case, the characteristics regarding the utterance length of the voice include (a) the time during which the distance between patterns became larger than the threshold value is approximately 40 ms or more; (b) the time during which the distance between patterns became smaller than the threshold value is approximately (c) No words may be uttered continuously for more than approximately 4 seconds. In this case, if the speech section candidate satisfies at least one of the above-mentioned features (a) and (b) and feature (c), it is determined that the speech section candidate is a speech 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, and the speech section candidate is divided into speech sections based on the magnitude relationship between the distance between patterns and the threshold value. Alternatively, since it is a method of determining a speech section as a noise section, there is no need to set aside special noise learning time, and there is no need for a special recognition method as in the past, and moreover, the speech 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 is a block diagram showing the configuration of an apparatus for explaining an embodiment of the voice section detection method of the present invention. First, the configuration and basic processing of this apparatus will be briefly explained.

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

This preprocessing signal processing processor 204 processes the audio data D1 from a center frequency of 250 Hz.
Calculate a group of 20 channels of Q = 6, IIR type single resonant digital bandpass filters up to 4.0kHz at 1/5 octave intervals, and calculate the absolute value of the output as follows:
The signal obtained by averaging each 128 points is sent to the speech section detection speech recognition microprocessor 205 as an input feature pattern D2.

The voice section detection voice recognition microprocessor 205 performs various processes as described below to detect a voice section, performs voice recognition processing, and outputs the recognition result R to an output terminal 206.

Next, an outline of an embodiment of the voice section detection process of the present invention will be explained with reference to the block diagram of FIG. 1 and the process flowchart of FIG. 3. In this flowchart, the 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 I is converted into audio data D1 in a circuit 203 and output, and a signal processing processor 204 extracts the characteristics of the input signal.

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

Next, the inter-pattern distance D _i between this input feature pattern and the noise standard pattern is calculated (S5).

In this calculation, it is assumed that the threshold value T _i of the i-th frame has already been determined in the voice section detection process of 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. Assume 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 j-th frame obtained by learning the noise standard pattern up to the i-1th frame within the noise interval while following slow changes in the nature of the noise (limited interval tracking learning). Let the average value of the channel noise standard pattern be the characteristic value A^ _j,i , and the average deviation be the characteristic value O^ _j,i . In this case, the inter-pattern distance D _i between the noise standard pattern and the input feature pattern in the i-th frame is calculated using the following equation.

D _i = ₂₀ 〓 ^j=1 {L(x _j,i −A^ _j,i-1 )}／O^ _j,i-1 However, L(x) is a ramp function and is given by the following formula .

L( _x )=0 (x<0) is compared with a threshold value T _i , and the speech section is determined based on various characteristics related to the length of speech utterance, such as the following (S6).

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

30 frames (approx.
If the distance D _i is smaller than the threshold T _i for a period of 400 ms or more, the speech section candidate is determined to be a speech section.

If an interval in which the distance D _i is greater than the threshold value T _i continues for 300 frames (approximately 4 seconds) or more, or if the interval does not meet the above-mentioned criteria, it is determined that the interval is a noise interval.

Based on the result of the determination based on such criteria ~, the threshold value T _i+1 of the next i+1th frame is determined.
In order to find the distance standard pattern during noise, limited tracking learning is performed. 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 determination data is the latest noise data. This update is performed according to the following formula.

Let the discrete time be i (equal to the frame number),
The characteristic value to be learned for the input feature pattern is f _i , the characteristic value to be learned for the inter-pattern distance is g _i , and the characteristic value in the noise standard pattern is
When F^ _i is the characteristic value within the standard distance pattern during noise, G^ _i is the characteristic value within the noise distance standard pattern, and the constants K and L are K>1 and L>1, respectively, the characteristic value F^ _i within the noise interval is F^ _i = {(K-1)F^ _i-1 /K} + f _i /K The characteristic value G^ _i in the same interval is G^ _i = {(L-1)G^ _i-1 /L} +g _i /L. Therefore, when expressing the specific characteristic values A^ _j,i and O^ _j,i of the noise standard pattern, A^ _j,i = {(K ₁ -1) A^ _j,i-1 /K ₁ }＋x _j,i ／K _1However , K ₁ ＞1 O^ _j,i = {(K ₂ −1)O^ _j,i-1 ／K ₂ }＋｜x _j,i −A^ _{j,i- 1}
|/K ₂ However, K ₂ > 1, and the specific characteristic value of the distance standard pattern during noise
D^ _i (D^ _i is the average value of the distance between patterns in the noise interval up to the i-th frame) and E^ _i (E^ _i is the average deviation of the distance between patterns in the noise interval up to the i-th frame). Expressed as D^ _i = {(K ₃ -1) D^ _i-1 /K ₃ } + D _i /K ₃ where K ₃ > 1 E^ _i = {(K ₄ -1) E^ _i-1 /K ₄ }+|D _i −D^ _i-1 |/K ₄ However, K ₄ > 1 K ₁ to _{K 4} are arbitrary constants that determine the followability of learning and are determined empirically, and the smaller this is, the faster the noise fluctuations. Be able to respond to 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 i-th frame is a voice section candidate. In this case, the noise standard pattern and the noise distance standard pattern are not updated. Therefore, in the case of the noise standard pattern, F^ _i = F^ _i-1 , and expressing this in concrete characteristic values, A^ _j,i = A^ _j,i-1 O^ _{j, i} = O^ _j,i-1 , and in the case of the noise distance standard pattern, G^ _i
= G^ _i-1 , and when expressed in concrete characteristic values, D^ _i = D^ _i-1 E^ _i = E^ _i-1 .

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, a threshold value T _i+1 for the next i+1th frame is determined based on the noise distance standard pattern described above (S8). This threshold T _i+1 is calculated by the following formula using the characteristic values D^ _i and E^ _i .

T _i+1 = D^ _i + CE^ _i Here, C is a constant that takes into account the safety factor, and the larger it becomes, the lower the detection sensitivity becomes, 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. In addition, if a detection has been made, speech recognition is performed on the detected speech section (S10), and then it is determined whether the desired speech recognition has been completed (S11), and whether the desired speech recognition has been completed is determined. In this case, this process becomes the end (S12). Here, the above-mentioned processing steps S5 to S11 are 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 "Totsukiyo" uttered under noise in the embodiment of the present invention. The sum of input feature patterns in this figure is the sum of spectral intensities in multiple channels of sonargrams obtained in experiments. In the figure, the horizontal axis represents time, that is, frames, and the vertical axis represents relative size logarithmically. Curve I indicates the sum of input feature patterns, for example, ₂₀ 〓 ^j=1 x _j,i , and the curve indicates the distance between patterns D _i
and the curve shows the threshold T _i . As can be understood from this experimental data, the threshold value T _i changes according to the change in the distance D _i in the noise section (1st to 20th frames), and remains constant within the speech section. Furthermore, it can be seen that the change in distance D _i is larger than the total sum of input feature patterns, and that it is effective to use it for section detection. Note 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.

(Effects of the Invention) As is clear from the above description, according to the present invention, there is no need to provide a special time for noise learning, and if the nature of the noise changes slowly, it can be followed. Since the system performs learning based on the noise, it is possible to accurately detect speech intervals without the need to newly learn noise.

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 that is required in the conventional method is not used, and therefore, it is simple and easy to implement the method of the present invention.

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

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an outline of a device used to explain the speech period detection method of the present invention, FIG. 2 is a block diagram showing a device used to explain the conventional speech period detection method, and FIG. FIG. 4 is a flowchart of a process for explaining one embodiment of the speech interval detection method. FIG. 4 is a diagram showing experimental results for explaining one embodiment of the speech interval detection method of the present invention. 201...Input terminal, 202...Low pass filter, 203...Sample hold, A/D converter, 204...Preprocessing signal processing processor, 2
05... Microprocessor for speech section detection and speech recognition, 206... Output terminal.

Claims (1)

[Claims] 1. A set of feature values representing characteristics of an input audio signal for each preset constant frame length is used as an input feature pattern, and when detecting a speech section using the input feature pattern, (a ) a process of determining a time interval other than the interval determined to be a speech interval candidate as a noise interval; (b) a process of learning a set of characteristic values representative of noise characteristics within the noise interval as a noise standard pattern; (c) a process of detecting the speech segment candidate as a speech segment based on the inter-pattern distance between the input feature pattern and the noise standard pattern, and the magnitude relationship with a threshold; A speech interval detection method characterized in that it is performed using a learning method that follows changes in characteristics. 2. The threshold value is calculated based on the obtained distance standard pattern during noise by learning a set of characteristic values representative of the characteristics of the distance between patterns during noise within the noise interval as a standard distance pattern during noise. 2. The voice section detection method according to claim 1, wherein the voice interval detection method is set to a value of . 3 The threshold value is expressed by the formula T=D^+CE^, where the threshold value is T, the estimated average value of the distance between the patterns in the noise interval is D, the estimated average deviation is E, 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 the following. 4. The learning method for the noise standard pattern is such that the discrete time is i, the characteristic value to be learned for the input feature pattern is f _i , the characteristic value within the noise standard pattern is F^ _i , and the constant K is When K > 1, within the noise section, F^ _i = {(K-1)F^ _{i-1 /} K} + f _i /K, and within the voice section, the formula is F^ _i = F^ _i-1. The speech interval detection method according to claim 1, characterized in that a learning method is used in which each calculation is performed separately. 5. 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. 6. 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 G^ _i. and the constant L is L>1
According to the formula G^ _i = {(L-1)G^ _i-1 /L} + g _i /L in the noise interval and G^ _i = G^ _i-1 in the voice interval, 6. The speech interval detection method according to claim 5, characterized in that a learning method is used in which each calculation is performed. 7. Detection of a speech section based on the magnitude relationship between the distance between the patterns and the threshold value includes comparing the distance between the patterns with the threshold value, and also detecting the speech section candidate as a speech section based on the characteristics related to the utterance length of the speech. 2. A voice section detection method according to claim 1, wherein the voice section detection method comprises: detecting a voice section; 8. 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 is approximately 40 ms or more, and (b) the time during which the distance between the patterns becomes smaller than the threshold value is approximately 400 ms. and (c) there are no words that are uttered continuously for more than approximately 4 seconds, and the speech segment candidate has at least one of the features in (a) and (b) above and the feature in (c) above. 8. The voice section detection method according to claim 7, wherein the voice section is determined to be a voice section if the condition is satisfied. 9. 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.