JPH0448400B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a speech recognition method that can perform highly accurate recognition even under high noise.

(Prior Art) Conventionally, this type of speech recognition method has been described in Journal of the Institute of Electronics and Communication Engineers, J68-1 (January 1985), pages 78-85. Figure 2 is a flowchart of a conventional speech recognition method using local peaks, in which the input speech is frequency-analyzed every 10 msec by a group of bandpass filters with 15 channels (second
(See 1 in Figure 2), as a normalization method for individual differences in vocal cord sound source characteristics, the voice spectrum is expressed logarithmically on both the amplitude and frequency axes, a least squares approximation straight line is found (see 2 in Figure 2), and the differences are corrected. . However, if the slope of the least squares approximation line is positive, the difference from the average value is taken. After that, as shown in Figure 3, for each part of each frame (10 msec) where the amplitude is 0 dB or more, the channel with the maximum value among those with an amplitude of 1/2 or more of each maximum value is regarded as having a local peak. Binarization is performed by setting the data to "1" and the others to "0" (see 3 in FIG. 2). The number of channels of the bandpass filter is 15, but when the slope of the least squares approximation straight line in the 16th channel is negative, it is regarded as a voiced sound and is set to 1.
When the slope is positive, it is regarded as an unvoiced sound, and "0" is set, and the sign of the slope is added (see 4 in Fig. 2).

The weighted average dictionary is obtained as a multivalued pattern by linearly extending and adding a plurality of binarized patterns to the longest one on the time axis (see 5 in FIG. 2).

To match a binary input pattern and a multi-value weighted average dictionary, the longer pattern in the time direction is linearly extended and matched, calculations are performed based on a certain degree of similarity, and the standard pattern that gives the maximum degree of similarity is selected. The category name is used as the recognition result (see 6 in Figure 2).

(Problems to be Solved by the Invention) The conventional speech recognition methods described above function effectively in an environment with a good SN ratio, such as when using a close-talking microphone, but in a high-noise environment, noise There was a problem that it was easy to pick up peaks due to fluctuations in the value, leading to an increase in erroneous recognition.

Even if there are peaks due to noise fluctuations as described above, the present invention uses a local peak vector calculation process that takes into consideration the difference in characteristics from the local peaks of speech, so that the peaks due to noise fluctuations are not picked up and the noise is reduced. The purpose of this invention is to provide a speech recognition method with strong resistance and high recognition accuracy.

(Means for Solving the Problems) The speech recognition method according to the present invention first performs frequency analysis on input speech into feature vectors of multiple channels for each speech frame.

On the other hand, the feature vector of the input voice is spectral normalized using a least squares approximation straight line in the voice frame to which the vector belongs. A binary window vector is calculated, in which each component of the feature vector after spectrum normalization is set to 1 if it is positive, and 0 if it is less than or equal to 0, the window vector is smoothed, and then each component of the window vector is The product of the component and each component of the feature vector after the spectrum normalization is calculated, and the component corresponding to the channel with the maximum value in the frequency direction is calculated from the product feature vector.
Calculate the local peak vector. The category of the input voice is determined by calculating the degree of similarity between the time series of local peak vectors of the input voice and a plurality of standard patterns prepared in advance.

(Operation) The present invention extracts a spectrum-normalized feature vector of input speech, smooths the window vector obtained from the spectrum-normalized feature vector before extracting the local peak vector, and multiplies this by the spectrum-normalized feature vector. Processing is in progress. Therefore, when extracting local peak vectors, it is possible to prevent peaks due to fluctuations in input noise from being mistakenly extracted as local peaks of input speech, and to stably extract local peak vectors of input speech.

(Embodiment) FIG. 1 is a block diagram showing an embodiment of the present invention. The configuration and operation of the speech recognition device shown in FIG. 1 will be explained below.

[Input processing]

Input audio is converted into an electrical signal through a microphone (not shown), passed through an amplifier (not shown), a low-pass filter (not shown), and an A/D converter (not shown) at a sampling frequency of 12 kHz, for example. The signal is sampled at the input terminal 101 and input to the input terminal 101.

[Frequency analysis processing]

A digital value inputted from an input terminal is subjected to frequency analysis in a frequency analysis section 102 and converted into a feature vector of an audio frame time series. This frequency analysis section 102 is composed of a band pass filter, an absolute value calculation section, and a low pass filter.

First, for frequency analysis, in this embodiment, a bandpass filter having a low Q characteristic as shown in FIG. 4 is used. Here, a low Q bandpass filter is used for the purpose of stable extraction of local peaks.

The output of each bandpass filter is subjected to absolute value calculation, input to a low-pass filter, and resampled every audio frame period (10 msec in this embodiment) to calculate a feature vector.

The output obtained by resampling the output of the low-pass filter of the k channel in the i-th audio frame is a _i
^k , the feature vector a _i in the i-th audio frame is expressed as a _i =a _i ¹ , a _i ² , . . . , a _i ^k . Here, K is the number of channels (K=22 in this embodiment), and a _i ¹ , a _i ² , . . . , a _i ^k are components of the feature vector a _i .

[Frame power calculation process]

The frame power calculation unit 103 receives the feature vector a _i output from the frequency analysis unit 102 for each audio frame, and calculates the frame power of the audio frame.
P _i is expressed by the following formula (1) Calculated by

[Voice section detection processing]

The voice section detection section 104 performs voice section detection using the frame power P _i output from the frame power calculation section 103.

Various algorithms have been proposed for speech interval detection, and the _algorithm itself is _not the purpose of the _present invention. When the frame power P _i continues to be below the threshold P _E for T ₂ frames after the start end of the audio, the frame in which the frame power P _i becomes below the threshold P _E for the first time is detected as the end I _E .

[Spectral normalization processing]

The spectrum normalization unit 105 is the frequency analysis unit 10
2, each component _{a i k} of the feature vector a _i is first logarithmically converted into an absolute value x _i ( ^k ) using the following equation (2).

X _i ^K =c log a _i ^k 0 a _i ^k ≧1 a _i ^k ≦0 (2) c is a constant determined from the number of bits of a _i ^k and the number of bits of x _i ^k .

Next, the least squares approximation straight line given by the following equation (3) y _i ^k = u _i · k + v _i However, The spectrum normalization process using the following equation (4) is performed.

z _i ^k =x _i ^k −y _i ^k (4) [Local peak vector calculation process] FIG. 5 shows the detailed configuration of the local peak vector calculation unit 106 according to the present invention.

In FIG. 5, 501 is an input terminal for spectrum normalized data z _i ^k , 502 is a binarization calculation unit, and 5
03 is a smoothing section, 504 is a multiplication operation section, 5
05 is a local maximum value extraction unit, and 506 is a local peak vector output terminal.

From the data z _i ^k whose spectrum has been normalized by the spectrum normalization unit 105, the binarization calculation unit 5
In 02, a binary window vector W _i =(W _i ¹ , W _i ² , . . . , W _i ^k , . . . , W _i ^K ) given by the following equation (5) is calculated.

(k represents the channel number.) W _i ^K = 1 0 z _i ^k > 0 z _i ^k 0 (5) Here, W _i ¹ , W _i ² , ..., W _i ^K are the components of the window vector W _i It is. Subsequently, the smoothing unit 503 smoothes the window vector W _i to obtain a smoothing window vector _i = (W _i ¹ , _i ^k , . . . , _i ^K ).

This smoothing is performed by setting the corresponding _i ^k to zero if the component W _i ^k of W _i does not become 1 continuously for two or more channels.

In other words, it is smoothed like...010110...000110...

Next, the product of each component _i ^k of the smoothed window vector _i and the spectrum-normalized data z _i ^k is calculated by the following equation (6) in the multiplication calculation unit 504.

L _i ^k =z _i ^k・_i ^k ……(6) (However, k=1,……
K) Next, using the L _i ^k obtained here, the maximum value extraction unit 505 calculates in the following equation (7) that L _i ^k >L _i ^k+1 and L _i ^k-1 <L _i ^k where k=1 , ..., K L _i ⁰ = −∞ L _i ^k+1 = −∞ (7) For k that satisfies the condition, r _i ^k = 1, and for k that does not satisfy the condition, r _i ^k = A local _peak vector r _i =r _i ¹ , r _i ² , . . ^. , _ri ^k , . Here, r _i ¹ , r _i ² , ..., r _i ^K are components of the local peak vector r _i .

Figure 6a shows the spectral normalized data z _i ^k
An example of the component w _i ^k of the window vector W _i is shown in Fig. 6b.
Figure 6c shows an example of the component w _i ^k of the smoothed window vector _i , and Figure 6d shows an example of the product L _i ^k of z _i ^k and _i ^k .
FIG. 6e shows an example of the component r _i ^k of the local peak vector r _i .

[Similarity calculation process]

The similarity calculation unit 107 receives the time series of the local peak vector r _i of the input voice output from the local peak vector calculation unit 106 and calculates the similarity with all standard patterns stored in the standard pattern memory 108 .

Here, the standard pattern is that before performing recognition on one or more training voices for each category, local peak vectors are calculated using the same process as during recognition, and the time axis is expanded/contracted and added. It has been created.

That is, the standard pattern is stored as a time series of weighted local peak vectors. In this embodiment, the number of standard patterns is M.

In the similarity calculation unit 107, the interframe similarity S(i,j) between the input voice and the standard pattern is
It is determined by the following equation (8).

Here, r _i represents the local peak vector of the input speech of the i-th frame, D _j represents the feature vector of the standard pattern of the j-th frame, r _i ^t represents the transposition of r _i , and D _j ^t represents the transposition of D _j .

Although there is a method of non-linearly matching i and j, in this embodiment linear matching is performed and the length of the m-th standard pattern is set to SL _n .

At this time, the degree of similarity S^ _n between the input voice and the m-th standard pattern is obtained by the following equation (9).

As described above, the degree of similarity S^ _n (where m=1 to M) with the input voice is calculated for all M standard patterns.

〔Determination process〕

The determining unit 109 determines the degree of similarity S^ _n (where m = 1 to
M), the one with the highest degree of similarity is extracted, and the category name of the standard pattern corresponding to the extracted degree of similarity is identified and output as a determination result.

That is, this judgment process determines m ₀ such that m ₀ = arg max S^ _n ...(10) by the process expressed by the following equation (10), and outputs the category name of the m _0th standard pattern. Output to terminal 110.

In the above explanation, each process is performed using hardware, but it is of course also possible to perform each process using software.

(Effects of the Invention) As described above in detail, according to the present invention,
A window vector is obtained from the spectral-normalized feature vector of the input audio, the window vector is smoothed, and the spectral-normalized feature vector is multiplied as a spectral window before the local peak vector is calculated. The local peaks obtained by the method are not mistaken for the local peaks of the voice, and highly accurate processing is performed in the similarity calculation processing and determination processing with each standard pattern, and as a result, a speech recognition device with high recognition accuracy can be realized.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention, FIG. 2 is a flowchart of a conventional speech recognition method, FIG. 3 is a diagram for explaining conventional binarization of an input signal, and FIG. Fig. 4 is a frequency characteristic diagram of a bandpass filter used for frequency analysis according to an embodiment of the present invention, Fig. 5 is a block diagram showing the configuration of a local peak vector calculation section of the present invention, and Figs. FIG. 3 is a diagram for explaining the process of extracting a local peak vector of input speech in FIG. 102... Frequency analysis section, 103... Frame power calculation section, 104... Voice section detection section, 105
... Spectrum normalization section, 106 ... Local peak vector calculation section, 107 ... Similarity calculation section,
108... Standard pattern memory, 109... Judgment section, 502... Binarization operation section, 503... Smoothing section, 504... Multiplication operation section, 505... Maximum value extraction section.

Claims (1)

[Claims] 1. A process of frequency-analyzing the input voice for each voice frame of a predetermined period and extracting a feature vector as a vector of frequency components of the input voice, and converting the feature vector of the input voice to the voice to which the feature vector belongs. A process of normalizing the spectrum using a least squares approximation straight line in the frame and extracting a spectrum normalized feature vector, and for each component of the spectrum normalized feature vector, if the component is positive, it is set as "1", and if it is less than or equal to 0, it is set as " A process of extracting a window vector consisting of each binary component converted as 0'', a process of smoothing the window vector to extract a smoothing window vector, and a process of extracting a smoothing window vector by smoothing the window vector, The process of calculating the product with each component and extracting it as a windowed feature vector, and determining the presence or absence of a local maximum value in the frequency direction for the windowed feature vector, and determining the component corresponding to the channel that has a local maximum value, that is, a local peak. A process of converting into a binary local peak vector with "1" for some and "0" for others, and calculation of the similarity between the time series of the local peak vector of the input audio and multiple standard patterns prepared in advance. 1. A speech recognition method, comprising: determining a category of input speech.