JPS63262695A - Voice recognition system - 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 recognition method, and in particular aims to improve the recognition probability by removing the influence of noise from an input signal in which noise is added to the speech to be recognized. Speech Recognition Method [Prior Art] Conventionally, this Ume speech recognition method has been implemented using three basic means shown in FIG. The voice standard pattern storage means 5 is composed of a number of types. In the feature parameter analysis, the input speech S1 containing noise inputted to the extraction means 1 is extracted and output by analyzing the feature parameter time series Dv+Ni necessary for speech identification, and this feature parameter time series includes the feature parameters Dv and noise. Contains a noise component N extracted from the input. The pattern matching identification means 4 compares this feature parameter time series with the speech standard pattern time series Dpi read from the speech standard pattern storage means 5, and performs speech recognition on i that designates the most similar speech standard pattern time series. Input speech is recognized in a format in which the standard speech that corresponds to the selected result is determined to be the input speech.

[Problem that the invention seeks to solve]

In the conventional speech recognition method described above, the characteristic parameter time series Dv+N input to the pattern matching identification means 4 contains a noise component N, and the speech standard pattern time series i does not contain a noise component. Therefore, the presence or absence of this noise component causes a difference in the values of both time series, which may lead to incorrect determination of which of the speech standard pattern time series is most similar. This is done by determining whether the amount of components that characterize speech from a time series with high sensitivity is greater or less than a certain threshold. The fact that the feature parameter time series contains noise components other than the speech components that should be recognized is
There is a high possibility that this speech segmentation judgment will be inaccurate, and therefore, the comparison between the feature parameter time series and the speech standard pattern time series will not be performed well, and the recognition result will likely contain errors.

One of the traditional solutions to this problem is
Before inputting the voice to be recognized to the feature parameter analysis and extraction means 1, only noise is input, the time series of the characteristic parameters are averaged and temporarily stored as a noise component, and next time when a voice containing noise is input, This method removes temporarily stored noise components from the feature parameter time series to reduce the shadow caused by noise contained in the feature parameter time series input to the slam matching identification means 4.

Another method is to include noise components included in the input signal to be recognized when generating the speech standard pattern time series to be stored in advance. None of these methods can respond to temporal changes in noise, and in particular cannot remove the effects of noise whose feature parameters change during a short period of time when speech is being input. First, there is a drawback that the recognition rate inevitably decreases.

The purpose of the present invention is to eliminate the above-mentioned drawbacks, and to provide an analysis and extraction method for receiving input speech containing speech to be recognized and noise in a speech recognition method using a bang matching recognition method. The feature parameter time series is analyzed and extracted, a second analysis and extraction means is newly provided, and only signals that have a correlation with the noise or that include all the components of the noise are inputted to the first analysis and extraction means. (by calculating and removing the noise component included in the feature parameter time series output by the second analysis extraction means from the noise feature parameters output by the second analysis extraction means and then performing the slam matching identification process), the recognition probability is significantly improved. [Means for solving all the problems] The method of the present invention includes a feature parameter analysis and extraction means for analyzing input speech and extracting all acoustic feature parameters in time series; A voice standard pattern storage means which stores in advance a standard voice characteristic parameter time series as a voice standard pattern time series, and reads out the characteristic parameter time series extracted by the characteristic parameter analysis and extraction means from the voice standard pattern storage means. In a speech recognition device, the speech recognition device includes a pattern matching identification means that outputs a classification result of the input speech by comparing it with a speech standard pattern time series that has been recognized. a first feature parameter analysis and extraction means that takes the entire signal input; and a second feature that takes as input an acoustic signal that has a correlation with the noise input to the first feature parameter analysis and extraction means or contains a component of the noise. a parameter analysis and extraction means, and a noise component calculated from the noise feature parameter time series outputted by the second feature parameter analysis and extraction means is removed from the feature parameter time series outputted by the first feature parameter analysis and extraction means; and a noise component removing means for supplying the pattern matching discrimination means to the pattern matching identification means.

〔Example〕

Next, the present invention will be explained with reference to the drawings.

FIG. 1 is a block diagram showing a first embodiment of the present invention. Feature parameter analysis and extraction means 1, noise feature parameter analysis and extraction means 2, noise component removal means 3, pattern matching and identification means 4, speech standard pattern The feature parameter analysis and extraction means 1 receives the input signal S1f, and the noise feature parameter analysis and extraction means 2 receives the noise source signal input to the feature parameter analysis and extraction means 1 or its Noise signal 82 (processed by 5 people) that has a correlation with noise and includes components that characterize the noise.

The feature parameter analysis and extraction means 1 outputs a feature parameter time series Dv+N, and the noise feature parameter analysis and extraction means 2 outputs a noise feature parameter time series DN.

The noise component removing means 3 temporally synchronizes the two input time series, calculates the noise component included in the feature parameter time series Dv+N from the noise feature parameter time series DN, and calculates the noise component included in the feature parameter time series Dv+ from the feature parameter time series Dv+. By removing the noise component N described above, the feature parameter time series Dv from which the noise component has been removed is output. The slam matching identification means 4 converts the standard pattern time series read from the voice standard pattern storage means 5 into the input feature parameter time series,
i (1==l p 2 t...N), determine the standard pattern time series that is most similar to the feature parameter time series D, and then select i that specifies this Dpi. Speech recognition result 1 (
i=1.2. ......N).

FIG. 2 is a block diagram showing a second embodiment of the present invention.

The second embodiment of the present invention shown in FIG.
sFilter]116. BPF(2)7. Noise component remover DP (D
FIG. 2 also shows a noise transfer system 101 having an impulse response h(tle), and the speech generated in this noise transfer system. A virtual adder 102 representing spatial addition of noise is also shown.

The second embodiment shown in FIG. The spectral envelope of each frequency band output from each of these band-pass filters is supplied to the noise component remover 9 as a characteristic parameter. Therefore, in this case, the feature parameter used for voice identification is the center frequency fj (j=1, 2, etc.) of the group of bandpass filters.
・・・・It is the power for each frequency band expressed as 1, and in the case of voice recognition, it is usually 8 to 20 CH (channels).
The feature parameters are analyzed and extracted using a filter bank consisting of several band-pass filters. The two bandpass filters shown in FIG. 2 each consist of a group of bandpass filters having the same center frequency.

Now, the input background of the signal force s (t) and the noise force n (tl that are input to the BPF (1) 6 and BPF (2) 71C, respectively) will be explained.

Figure 3 shows the vocalist Pe〇- in the embodiment of hayashi 2 in Figure 2. FIG. 2 is an explanatory diagram showing the generation situation of force and noise input.

The voice v (t) uttered by the speaker 11 is the microphone 12
The noise n generated from the noise source 13 is converted into an electrical signal by s (t) and supplied to the BPF (116) as a voice signal s (t). Therefore, the audio input s(t) output from the microphone 12 is composed of the audio v(t) and the input noise n(t)*h(
t) can be expressed as the sum v(t)+n(tl*h(t). Here, the symbol sign represents convolution multiplication. Also, the noise n is converted into an electrical signal by the microphone 14 and becomes a noise It is supplied to BPF (2) 7 as input n(t).

Here, the noise source refers to any sound other than the voice V to be recognized; for example, the voices of humans other than the speaker 11 also constitute noise.

FIG. 4 is an explanatory diagram showing another situation in which voice input and noise input occur in the second embodiment of FIG. 2.

In this FIG. 4, the noise transmission system 101 is used for an aircraft pilot's acid mask, soundproof wall, etc. In many cases, the audio output 5 (
t) is the same as in Figure 3 (v (t) + n (t
) * h (t), and the microphone 14 also inputs surrounding uniform noise n and calculates the noise human power n(t
lffi output O Returning again to the embodiment of FIG. 2, the explanation will be in line t-α. In FIG. 2, v(t) is the total voice to be recognized, and noise human power n(t) is transmitted through the noise transmission system 101 to h(t)*n
(t) is added to the audio signal v(t) to generate audio input 5(1). Adder 102 is noise transmission system 10
This represents the spatial addition in 1.

BPF (1) 6 consists of a group of l-channel BPFs, and the output of each channel is the feature parameter S (1, S (fz
)t...8(fJ). The analysis and extraction of this characteristic parameter is carried out at a constant period of T seconds,
In the case of audio, this is usually repeated every so-called analysis cycle of about 10 to 20 m5. Therefore, the four feature parameters S (fx), S (fz), ......
The set of 5 (fl) becomes a time series every T seconds. In exactly the same way, the noise human power n(t) input to B P F (2) 8 is also analyzed, resulting in N(, ft).

Noise characteristic parameters N(fz), . . . NCf1) are extracted every T seconds and output as a time series.

Here, the time domain function 5(t)=v(t)+h(g
*n (If Ll is Fourier transformed and expressed as a function in the frequency domain, it can be expressed as 5(1)=V)+l(+N). Therefore, in the noise component remover 8, 5
(fj) and N(fj), V(fj)=S(fj
)-H(fj)・N(fj) is performed once for j=t, z,...l, then the voice V(1) which does not contain any noise components can be calculated as % characteristic Haraj- Ta V (
j't), ■(fz), =-VCfl). H(fj) used in this calculation is based on the frequency characteristic of the noise transfer system 101, which is 8(fD=HCfo
) ・NCfj ) (j=x, z, ------, (
Since )i(fj)=S(fj)/N(
, fi) (j = i, z, ...2), and this value can be temporarily memorized and used as the voice input since it does not change suddenly like the noise force n (tj). When v(t) is input, it can be used for noise component removal calculations.

Denoised feature parameters V Ul) , V(f
z).

The time series of VCf, ) is supplied to the DP processor 9 as a pattern matching identification means, and the audio standard pattern Vi(ft
), Vi(fz), . . . are compared and verified with Vi(fl). This slam matching process usually uses a dynamic programming method to
“The degree of similarity between feature parameter time series can be determined by matching while performing non-linear matching that takes into account the temporal expansion and contraction of the input.

Generally, a memory such as RAM or ROM is used as the audio standard pattern memory No., and here the audio standard patterns ViCfx), vi(fz), vi(fz),
...viCfl) (i= 1.2 m...
... M) feature parameter time series of M standard voices are registered in advance. In dynamic programming, for the standard pattern M time series, the input time series V (fl) - V (fz), ...... ■
(, fJ3) and which is the most similar to all judgments, Vi(
ft)-Vi(fz) and "-""''Vi(fz), the standard pattern with the most similar i is output as a recognition result.

The above is a description of the embodiment shown in FIG. 2, but many other variations are possible, and all of these can be easily implemented without departing from the spirit of the invention, as will be described below.

For example, in the embodiments of FIGS. 1 and 2, two BPFfts are provided as feature parameter analysis and extraction means or feature parameter analysis and extraction devices, respectively. On the other hand, an input voice having a correlation with the above-mentioned constriction or containing a noise component is supplied, but the above-mentioned feature parameter analysis and extraction means or BPF' are each realized by one common unit, and all of the input It is clear that this could be easily implemented by time-sharing.

Furthermore, the amount of calculations required for the processing performed to remove noise components is relatively small, and it is possible to perform it before or during pattern matching and identification processing, so physically these two processes have the same configuration. It can also be easily implemented.

〔Effect of the invention〕

As explained above, the present invention uses the second feature parameter analysis and extraction means to extract the feature parameters of the noise that is added to the speech to be recognized and causes a decrease in the recognition rate, and performs the matching identification process. By removing noise components from the feature parameter time series to be used,
As the noise components are removed, the similarity of the speech feature parameter time series increases9, recognition of the input speech,,! Improving the ratio has a corresponding effect. In addition, detection of the start and end of speech is an important factor that affects the recognition rate in comparison with the speech standard pattern time series, that is, in matching identification processing, but noise components are detected at the start and end where the power level of the speech signal is low If it is removed, it will be easier to detect it and the recognition rate will be greatly improved.

Furthermore, in conventional speech recognition methods, it is necessary to remove noise at the acoustic waveform level of the input signal, especially when used under high noise conditions, but removal at the acoustic waveform level requires at least 8 KF.
A product of about 200 points in Iz.

While the sum operation (approximately 1,600,000 times/second was required), in the noise component removal in the feature parameter domain as in the present invention, when using BPF = 2 for 16 channels and on the gold side, the period eloms is also 1,600 times/second. This has the effect that the amount of calculations required for noise removal can be greatly reduced by requiring only one product-sum operation.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing a first embodiment of the present invention;
Fig. 3 is a block diagram showing the second embodiment of the present invention, Fig. 3 is an explanatory diagram showing the occurrence of voice input and noise input in Figs. 1 and 2, Fig. 4 is 1... Feature parameter analysis and extraction means, 2...Noise feature parameter analysis and extraction means, 3. Mountain...Noise analysis and removal means, 4...
...Pattern matching identification means, 5...Audio standard pattern storage means, 6...BPF (1),
7...BPF (2),...Noise component remover, 9...DP processor, 10...Audio standard pattern memo! J, 11...Speaker, xz
...Microphone, 13...Noise source,
14...Microphone.

Claims (2)

[Claims] (1) Feature parameter analysis and extraction means for analyzing input speech and extracting acoustic feature parameter time series; speech standard pattern storage means for storing feature parameter time series of standard speech in advance as speech standard pattern time series; and pattern matching identification means for comparing the feature parameter time series extracted by the feature parameter analysis and extraction means with the speech standard pattern time series read from the speech standard pattern storage means and outputting an identification result of the input speech. a first feature parameter analysis and extraction means that receives as input an acoustic signal containing the speech to be recognized and other sounds as noise;
a second feature parameter analysis and extraction means that receives as input an acoustic signal that has a correlation with the noise input to the first feature parameter analysis and extraction means or contains a component of the noise;
The noise component calculated from the noise feature parameter time series outputted by the second feature parameter analysis and extraction means is
a noise component removing means for removing a noise component from the feature parameter time series outputted by the feature parameter analysis and extraction means and supplying the noise component to the pattern matching identification means. (2) The first feature parameter analysis and extraction means and the second feature parameter analysis and extraction means are realized by one common feature parameter analysis and extraction means, and the sound to be input to the first feature parameter analysis and extraction means is The speech recognition method according to claim 1, wherein the signal and the acoustic signal to be input to the second feature parameter analysis and extraction means are inputted in a time-sharing manner.