JPS60129800A - 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 (Technical Field) The present invention relates to an improvement in a method for resampling voice data in order to improve the matching accuracy of a voice recognition device.

(Technical background) A conventional speech recognition device is configured as shown in Fig. 1, where 1 is an input terminal, 2 is a frequency analysis section, 3 is spectrum data, 4 is a speech interval determination section, and 5 is a reproduction section. Sample section, 6 is distance calculation section, 7 is standard pJ? Turn memory, 8 is the judgment section,
9 is a recognition result output terminal.

In a conventional speech recognition device, in the distance calculation between the spectrally converted input speech pattern and the standard pattern (K=1.K), the distance DK is input and the mth channel of the nth time sample of the turn is input. When the element is A (m, n) and the m-th channel element of the n-th time sample point in the standard pattern is SK (m, n), the standard D in the turn
The recognition result is the standard/o-turn catechism that minimizes K.

Here, there are many methods for calculating the weights W (m, n), but they are not the purpose of this invention and will therefore be omitted.

Conventional speech recognition devices frequency-analyze input speech, calculate a least squares approximation straight line of the speech spectrum, and use the slope of the least squares approximation straight line as a spectral slope value. and,
If the spectral slope value is negative, the input speech is determined to be voiced and a least square approximation line of the voice spectrum is drawn from the voice spectrum.If the spectral slope value is positive, the input voice is determined to be unvoiced and the voice is By subtracting the average of the voice spectrum from the spectrum, the slope of the vocal cord sound source characteristics of the input voice and the difference in phonation intensity are normalized.

The normalized audio data is resampled by the resampling section 5 to the audio section determined by the audio section determining section 4 to obtain a constant audio sample data length. Re-sampling is usually performed by dividing the utterance section of the voice into 16 or 32 points.

FIG. 2 shows the correspondence between input audio data and resampled data in the linear matching method.

That is, the resampling data (sample a constant value n) is associated with the start frame S of the input audio and the end frame EDFR of TFR1 in a linear relationship as shown in the figure.

In this case, depending on the number of frames of the input audio data and the value of the number of resampled data, the temporal linearity of the resampled data with respect to the input audio data may be lost.

FIG. 3 shows an example of a resampling state. As described above, in conventional linear matching, the resampling result of input audio data is non-linear, and temporal changes with respect to the correct input audio data spectrum are not included. As a result, for example, in word recognition, characteristic changes in the word ``sutakutor'' are not reflected in the resampling results.
There are problems such as easy recognition errors during matching.

(Objective of the Invention) The object of the present invention is to provide a speech recognition method capable of eliminating misrecognition and improving the recognition rate, and to accurately reflect changes in the spectrum of input speech data in resampled data, thereby achieving the above-mentioned method. It provides a means to solve the problem and will be explained in detail below.

(Structure of the Invention) FIG. 4 is a diagram showing the relationship between input data and resampled data of the present invention and their windows, and the input data is 6.
This is an example in which the number of resamples is set to 5 frames.

In such a relationship diagram, six frames of input audio data are generally replayed/pulled according to the following equation (1).

a-1 ・・・・・・・・・・・・・・・ (1)(
In formula 1), n is the number of input audio data frames, and m is the number of resamples. The resampling is resampled at the correct position within the input audio data to accurately reflect temporal changes in the spectrum of the input audio.

The window width wl1w2 is set to approximately one frame of each resampled data, but the window may be set wide and overlapped in each sample as shown in w2.

FIG. 5 is a diagram illustrating a method for calculating tract data in the resansol part.

The window length W has the relationship shown by the following equation (2).

” = , 1- r + 12 river...song (2
) Here, the length included in the first frame of the input audio data in the window W of the resampled frame (m) is 1.
1, and the length included in the (k+, 1) 7th frame is t
Set it to 2.

The resampled spectrum data v(m) is expressed by the following formula (3)
It is calculated by

In Equation (3), the input audio spectrum of the 1st frame is expressed as (k), the input audio spectrum data of the (k+1)th frame is expressed as u(k+1), and the input audio spectrum data of the (m1)th frame is expressed as v Represented by (m).

In FIG. 5, the calculation of v(m) requires multiplication and division, which complicates implementation in a circuit. Therefore, in order to simplify the calculation, one frame of input audio is divided into a fixed number of parts, and t□+L2 is replaced with an integer.

FIG. 6 is a diagram illustrating a method for re-calculating data using a frame internal division method. As an example, a diagram is shown in which one frame is divided into four and the window length of resampled data is four.

1 in this case. The value of the ratio of and t2 is ・z, :z, ,=
3:x, and the resampled spectrum data is calculated using the following equation (4).

3u(n)-1-u(n+1) v(m)-□ ・・・・・・・・・・・・ (4) Seventh
The figure is a block diagram showing one embodiment of the present invention.

In FIG. 7, 100 is an input terminal, and 200 is a frequency analysis section. 30θ is the spectrum conversion section I), 4o
o is a voice section determining section. 1 in 500 resampling section, 50 accumulators, 50 spectral addition circuits
2. Division circuit 503, input frame number calculation circuit 504,
P1 arithmetic circuit 505, integer conversion circuit 506, addition circuit 50
7, division circuit 508, integer conversion circuit 509, resample Stezzo counter 5101 in-sample counter 511,
It consists of a resample control section 512. 600 is Ma.

Ching calculation section, 7O0 is standard/zotan memory section, 800
9 is a determination unit, and 900 is a recognition result output terminal.

In such a configuration, an input audio signal inputted from the input terminal 100 is inputted to the frequency analysis section 2θ0, and is frequency analyzed and sent to the vector transformation section 3θ0 as a quantized signal corresponding to the frequency band of the number of boards. It will be done.

The data sent to the spectrum conversion unit 3oo undergoes spectrum conversion to become spectrum information, audio power information, etc., and is sent to the audio interval determination unit 400 and the resampling unit 5θθ
sent to.

The voice section determination unit 400 determines the start and end points of the voice interval using the voice/voice information, and the resampling unit 50
0 and is sent to the matching calculation section 600. - Spectral data u(l) sent to the re-sansol unit 5θ0
, J) are resampled by the method described below to become resampled data v(i, j), which is sent to the matching calculation unit 600, and the distance from the standard pattern stored in the standard pattern 700 is calculated. The calculation is performed and sent to the determination section 800. The determination unit 800 compares the distance value with the total distance, and outputs the category name of the smallest total distance from the output terminal 9.00 as the recognition result.

Now, the operation of the resannier' /I/ calculation section 500 of the present invention will be explained.

The resampling methods described so far have described ideal resampling conditions.

An example in which this method is implemented as a circuit will be described.

In circuitization, some modifications were made to the method of calculating resampled data in order to simplify calculations. A case where the number of divisions of one frame is 4 and the window length is 4 will be explained.

Let the input audio data be u(1+jL!J) and the sample data be v(1+J), and calculate v(++j) according to the following equation.

P 1= [:4X(n-1)X(ji-i)/31
]-(6)P2=P1+3...
・・・・・・・・・・・・・・・・・・・・・ (7)
] means to convert the value into an integer. In addition, n indicates the total number of frames of input audio data that has been cut out, and P indicates the range of the window section of resample data in an example of a subroom number system in which each frame of input audio data is divided into four. Since the subframe number P1 is the starting subframe number of P, and P2 is the ending subframe number of window section 4, the relationship between PI and P2 is as shown in equation (7).

FIG. 8 shows the relationship between frame numbers and subframe numbers.

Equation (6) is an equation that gives the value P according to the resampling number j, and the calculation result is converted into an integer. The calculation result is as follows when the number n of input audio data frames is given. 〜 j=32 Px−Cφ−2 The first subframe is converted to the final 32nd subframe u (j=
32) indicates the first subframe of the last n@th frame of the input audio data.

By specifying each resampling interval as described above, the calculation of the resampled spectrum data example shown by equation (5) is performed.

That is, in equation (5), each resampling interval p l~P
In order to add the corresponding input audio spectrum data four times to which frame region of the input audio data 2 belongs and average the entire spectrum, by calculating the added value,
This results in an averaged resampled spectrum data sequence, which is the object of the present invention. The above calculation process is shown in FIG.

Spectral data 30' 0 data u(
i, j) are sent to the resampling unit 500 and stored in the achievator 501.

7*, the spectrum-exchanged audio data stored in the MU 501 is sent to the spectrum addition circuit 502 to add the input audio spectrum data, and then to the division circuit 503 for averaging the spectrum. is divided, and the re-sampled data v(i, j) is sent to the matching calculation unit 600.

Further, the data in which the start and end of the voice section are determined by the voice section determining section 400 using the voice/voice information is sent to the input frame number calculation circuit 504 . The number n of input frames calculated by the input frame number calculation circuit 504
is sent to the P1 calculation circuit 505. Pl calculation circuit 50
After the resampled spectrum data P1 is calculated using the above-mentioned equation (3), the data sent to step 5 is converted into an integer by the integer conversion circuit 506 using equation (6), and then the data is converted into an integer using equation (6).
After the P value is sequentially obtained by the addition circuit 507 according to the formula, the division circuit 5o8' calculates the frame number of the input audio data 501, and the spectrum addition circuit 50. ).

Next, according to the count of the in-sample counter 511, the addition circuit repeatedly adds 1 to the P value, and sequentially from P1 to P2.
The four values are converted into integers, and the values are sent to the achievator 50 and the spectrum addition circuit 502.

The spectral addition circuit 502 reads out the spectral data corresponding to the four frames obtained above from the achievator 501, adds them, and averages them in the division circuit 503 to obtain one resampled data.

In addition, when the re-sampling start signal from the voice interval determining section is sent to the re-sampling control section 512, the re-sampling 500 counts from the start frame 1 frame to the end frame 32 frames in the re-sampling stelaph 0 counter 510. 32 re-sampled data are sequentially created.

When the re-sample step counter 51o reaches the required count, the re-sample control unit 512 sends a matching start signal to the matching calculation unit 6oθ.

(Effects of the Invention) As explained above, by employing the configuration of the present invention, loss of temporal linearity of resampled data with respect to input speech data can be avoided, and at the same time, in word recognition. , characteristic spectral changes in words are reflected in the re-sansol results, reducing misrecognition during matching and improving recognition performance. In addition, as a result of actually conducting experiments using this method, we found that when compared with the conventional resampling method, the recognition rate improved, the distance value between the 1st and 2nd place expanded, and the stability during recognition improved. Two were seen.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of a conventional speech recognition device, Fig. 2 is a diagram showing the correspondence between input speech data and resampled data in the linear matching method, Fig. 3 is an example of the state of resampling, and Fig. 4 The figure shows the relationship between the input data and resampled data of the present invention and their windows.
Diagram showing the calculation method of resampled spectrum data, No. 6
7 is a diagram illustrating a re-sannor data calculation method using a frame internal division method, and FIG. 7 is a block diagram illustrating an embodiment of the present invention. FIG. 8 is a diagram showing the relationship between frame numbers and subframe numbers. lθQ...input terminal, 2θθ...frequency analysis section, 3
θ0...spectrum conversion section, 4oo...speech interval determination section, 5oo...resampling section, 50...7 cumulator, 5o2...spectrum addition circuit, 5o3.
・・Division circuit, 5o4 ・・Input frame number calculation circuit,
505...P, arithmetic circuit, 505...integer conversion circuit,
507... Addition NO path, 50B... Division circuit, 50
9... Integer conversion circuit, 51o... Re-sample step counter, 51no... Sunnor internal counter, 612...
...Resampling control unit, 60Q...Matching calculation unit, 7θθ...Standard Noreen memory unit, 8θθ...Judgment unit, 9θ0...Recognition result output gl, i child. Patent applicant: Oki Electric Industry Co., Ltd. Figure 2 5TFREDFR -in f1@? -r'-ri Figure 3 Mefu Shimukyo 10 Re-on 7''Jl/& 6 around 1≧\Nozukuri Figure 4 Figure 5 Procedural amendment (voluntary) 59.416 Showa year Month Date Commissioner of the Japan Patent Office 1 Indication of the case 1982 Patent Application No. 236345 2 Name of the invention Voice recognition method 3 Person making the amendment Relationship with the case Patent Applicant

Claims (1)

[Claims] Frequency analysis of the input voice is performed, spectral characteristics are normalized, time normalization is performed by resampling, matching is performed with the standard/mother turn, and the category with the highest degree of similarity is determined to be the category of the input voice. In the speech recognition device to be judged, (a) each original frame data of input audio data is divided into a plurality of subframe data carrying the same information as the spectral characteristics of the original frame data, and (b) input audio is divided into a whole frame information series. Extract the corresponding subframe information series, (c) Divide the subframe information series at equal intervals by a constant number corresponding to the number of resamples, and (2) Divide the subframe information series into a set time length of the spectral data to be averaged as resampled data. The starting subframe and the ending subframe of each resample are determined according to the corresponding window length, and the resampling is performed by calculating the average of the spectral characteristics carried by all subframes between the starting end and the ending subframe. A voice recognition method that uses