JPH0122638B2 - - Google Patents

Description

[Detailed Description of the Invention] As a speech analysis device used for speech analysis and synthesis or speech recognition, speech waveforms are autocorrelated using an analysis window such as a rectangular window or a Hamming window at every analysis period (frame period) of about 10 to 30 msec. 2. Description of the Related Art Speech analysis devices that extract voice characteristic parameters such as values and formant frequencies are known. However, the conventional speech analysis apparatus has a drawback that the feature parameter value extracted is affected by the position of the analysis window due to the fact that the speech waveform has a quasi-periodic waveform, and fluctuations occur due to analysis errors. In particular, when calculating characteristic parameter values within a short period of one pitch cycle or less, there is a risk that an analysis error may occur due to the large influence of the excitation point of the vocal cord sound source. In order to prevent this drawback, a method is known in which the excitation time point of the vocal cord sound source is extracted and analysis is performed in synchronization with the excitation time point, but accurate detection of the excitation time point is extremely difficult.

An object of the present invention is to provide a speech analysis device that is capable of reducing analysis errors caused by differences between the temporal position of an analysis window and the excitation point, etc., and extracting an accurate speech characteristic parameter value series with relatively little fluctuation. It's about doing.

The present invention provides a means for calculating feature parameter values using a plurality of analysis windows at different temporal positions for each analysis frame period, and a method for calculating a distance between feature parameter values extracted in temporally adjacent analysis sections. and selecting a series of feature parameter values that minimizes the sum of distances between the feature parameter values from among feature parameters for analysis windows at different time positions within a time interval including a plurality of said analysis intervals. It consists of means.

The feature of the present invention is that, within each analysis interval, first,
Characteristic parameter values such as formant frequencies are calculated for multiple analysis windows at different temporal positions, and the optimal one is selected from among these characteristic parameter values in temporally adjacent analysis intervals. The objective is to select such that the sum of the distances from the characteristic parameter values to be selected is the minimum.

Therefore, it is clear that according to the present invention, a feature parameter value series with relatively little fluctuation between analysis frames can be selected.

The optimal feature parameter value series is selected using the dynamic programming method (DP method) described below.
can be selected efficiently using

The feature parameter value for the i-th analysis window in the n-th analysis frame is represented by a vector _〓o , _i , and the distance between the feature vector _〓o , _i and 〓o _-1 , _j is expressed as R(n, i | n-1 , j), the optimal feature parameter value series in each analysis interval can be obtained by solving the following recurrence equations (1) and (2).

S(n, i)＝MinR(n, i｜n‐1, 1)+S(n‐1, 1) R(n, i｜n‐1, 2)+S(n‐1, 2) 〓 〓 R(n, i｜n-1, j)+S(n-1, j) 〓 〓 R(n, i｜n-1, I)+S(n-1, I) …(1) Here where i = 1, 2, ..., I (where I is the number of analysis windows) n = 1, 2, ..., N (N is the number of analysis sections), and S (n, i) is the number of analysis windows in the nth analysis frame. This is an evaluation value for the i-th analysis window, and is an integrated value of distance R.

Smin=MinS(N, 1) S(N, 2) 〓 S(N, I) ...(2) The optimal feature parameter value series is associated with the minimum evaluation value Smin expressed by equation (2) (n, i) Determined as a series.

Furthermore, for example, a formant frequency can be used as the characteristic parameter value. The formant frequency can be determined, for example, by converting an autocorrelation value into a linear prediction coefficient, and further converting the linear prediction coefficient into a polar frequency. This method is detailed in reference material (1) below, so the explanation will be omitted here.

Material (1): “Speech Analysis and Synthesis by
Linear Prediction of the Speech Wave”, B. S. Atal and SL Hanauer, The
Journal of the Acoustical Society of
America, Vol. 50, Num. 2, 1971 Next, the present invention will be explained in detail using the drawings. The figure is a block diagram showing one embodiment of the present invention.

First, a voice waveform is input to the voice waveform buffer 2 via the voice waveform input terminal 1 and temporarily stored. The audio waveform temporarily stored in the audio waveform buffer 2 is
The waveform output control data output from the control circuit 12 via the audio waveform buffer control data transmission line 6 is input to the window circuit 3. The window circuit 3 calculates a waveform obtained by multiplying the audio waveform by windows at different time positions according to the window position control data inputted from the control circuit 12 via the window circuit control data transmission line 7, and sequentially performs the filtering. It is output to the cloak parameter extraction circuit 4. The formant extraction circuit 4 calculates formant frequencies for each of the voice waveforms that are windowed at different time positions, and stores them in the formant parameter value temporary storage circuit 5. The formant parameter value temporary storage circuit 5 stores the analysis period and the previous analysis according to the formant parameter value output data given from the control circuit 12 via the formant parameter value temporary storage circuit control data transmission line 9. The two formant parameter values in the section are output to the distance calculation circuit 11. The distance calculation circuit 11 calculates the distance between the two formant parameter values (R in the above formula (1)) according to distance calculation circuit control data given from the control circuit 12 via the distance calculation circuit control data transmission line 10. corresponding) is calculated and output to the evaluation value calculation circuit 16.
The evaluation value calculation circuit 16 follows the evaluation value calculation circuit control data given from the control circuit 12 via the evaluation value calculation circuit control data transmission line 13, and calculates the distance data and the previous evaluation output from the evaluation value temporary storage circuit 15. value (corresponding to S(n-1, j) in equation (1)) to a new evaluation value (S in equation (1))
(corresponding to n, i)), and stores the optimal path data (corresponding to the value of j that gives the minimum value in equation (1)) in the optimal path data temporary storage circuit 17, and the new evaluation The value is stored in the evaluation value temporary storage circuit 15.

After repeating the above process a predetermined number of frames (corresponding to N in equations (1) and (2) above), the optimal data address generation circuit 18 is controlled by the control circuit 12 to control the optimal data address generation circuit. According to the control data given via the data transmission path 14, the minimum value in the Nth analysis interval output from the evaluation value temporary storage circuit 15 is detected (corresponding to S _nio in the above equation (2)). After that, the minimum value (S _nio )
By sequentially outputting the optimal path data associated with the optimal path data from the optimal path data temporary storage circuit 17, address data for the optimal formant parameter value in each analysis interval is generated, and the formant is transmitted via the address data transmission line 19. It is output to the parameter value temporary storage circuit 5. The formant parameter value temporary storage circuit 5 sequentially outputs the formant parameter values from the characteristic parameter value output terminal 20 according to the address data.

In the above explanation, the formant frequency was used as the feature parameter, but it is possible to realize a speech analysis device with similar effects using autocorrelation values, filter bank output values, linear prediction coefficients, etc. as the feature parameters. it is obvious.

[Brief explanation of drawings]

The figure is a block diagram for explaining an embodiment of the present invention. In the figure, 1 is an audio waveform input terminal;
2 is an audio waveform buffer, 3 is a window circuit, 4 is a formant parameter extraction circuit, 5 is a formant parameter value temporary storage circuit, 6 is an audio waveform buffer control data transmission line, 7 is a window circuit control data transmission line, 8 is a formant parameter extraction circuit control data transmission line, 9 is a formant parameter temporary storage circuit control data transmission line, 10 is a distance calculation circuit control data transmission line, 11 is a distance calculation circuit, 12 is a control circuit, and 13 is a control circuit. 14 is an evaluation value calculation circuit control data transmission line, 14 is an optimum data address generation circuit control data transmission line, 15 is an evaluation value temporary storage circuit, 16 is an evaluation value calculation circuit, 17 is an optimum path data temporary storage circuit, and 18 is an optimum data address. 19 is an address data transmission path, and 20 is a characteristic parameter value output terminal.

Claims (1)

[Claims] 1 Analyze the audio waveform for each analysis period and analyze each analysis section n
In a speech analysis device that extracts feature parameter values of speech at (=1, 2, ..., N), a plurality of analysis windows i (=
1, 2, ..., I) Feature parameter values 〓 _o,i
and means for calculating the analysis window j in the analysis interval n-1 for the analysis window i in the analysis interval n.
Distance R _{( n} , _i | _n
−1, j) for all i, j, n, and for each analysis interval n and analysis window i,
Feature parameter value by analysis window j (=1, 2, ..., I) of analysis section n-1 of analysis section n-1 held〓o _{-1 ,} cumulative distance value S(n-1, j)
The above R(n, i|n-1, j) is added to each of R(n, i|n-1, j)+S(n-1, j);
Analyze the process of obtaining (j = 1, 2, ..., I) and setting the minimum value among these I values as the cumulative distance value S (n, i) of this analysis interval n to the analysis window i. Section N
Distance cumulative value S(N, i); (i=1,
2,...,I), and this S(N,i); (i=1,
2, . . . , N) for outputting a combination of feature parameter values that gives the minimum value.